Cross Reference: /external/clang/test/CodeGen/arm_neon

Lines Matching full:bitcast
18 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
19 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %c to <8 x i8>
20 // CHECK:   [[VABD_V_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
21 // CHECK:   [[VABD_V1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
23 // CHECK:   [[VABD_V3_I_I:%.*]] = bitcast <4 x i16> [[VABD_V2_I_I]] to <8 x i8>
24 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I_I]] to <4 x i16>
32 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
33 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %c to <8 x i8>
34 // CHECK:   [[VABD_V_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
35 // CHECK:   [[VABD_V1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
37 // CHECK:   [[VABD_V3_I_I:%.*]] = bitcast <2 x i32> [[VABD_V2_I_I]] to <8 x i8>
38 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I_I]] to <2 x i32>
54 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
55 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %c to <8 x i8>
56 // CHECK:   [[VABD_V_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
57 // CHECK:   [[VABD_V1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
59 // CHECK:   [[VABD_V3_I_I:%.*]] = bitcast <4 x i16> [[VABD_V2_I_I]] to <8 x i8>
60 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I_I]] to <4 x i16>
68 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
69 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %c to <8 x i8>
70 // CHECK:   [[VABD_V_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
71 // CHECK:   [[VABD_V1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
73 // CHECK:   [[VABD_V3_I_I:%.*]] = bitcast <2 x i32> [[VABD_V2_I_I]] to <8 x i8>
74 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I_I]] to <2 x i32>
90 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %b to <16 x i8>
91 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %c to <16 x i8>
92 // CHECK:   [[VABDQ_V_I_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
93 // CHECK:   [[VABDQ_V1_I_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
95 // CHECK:   [[VABDQ_V3_I_I:%.*]] = bitcast <8 x i16> [[VABDQ_V2_I_I]] to <16 x i8>
96 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VABDQ_V3_I_I]] to <8 x i16>
104 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %b to <16 x i8>
105 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %c to <16 x i8>
106 // CHECK:   [[VABDQ_V_I_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
107 // CHECK:   [[VABDQ_V1_I_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
109 // CHECK:   [[VABDQ_V3_I_I:%.*]] = bitcast <4 x i32> [[VABDQ_V2_I_I]] to <16 x i8>
110 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VABDQ_V3_I_I]] to <4 x i32>
126 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %b to <16 x i8>
127 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %c to <16 x i8>
128 // CHECK:   [[VABDQ_V_I_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
129 // CHECK:   [[VABDQ_V1_I_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
131 // CHECK:   [[VABDQ_V3_I_I:%.*]] = bitcast <8 x i16> [[VABDQ_V2_I_I]] to <16 x i8>
132 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VABDQ_V3_I_I]] to <8 x i16>
140 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %b to <16 x i8>
141 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %c to <16 x i8>
142 // CHECK:   [[VABDQ_V_I_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
143 // CHECK:   [[VABDQ_V1_I_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
145 // CHECK:   [[VABDQ_V3_I_I:%.*]] = bitcast <4 x i32> [[VABDQ_V2_I_I]] to <16 x i8>
146 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VABDQ_V3_I_I]] to <4 x i32>
164 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
165 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %c to <8 x i8>
166 // CHECK:   [[VABD_V_I_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
167 // CHECK:   [[VABD_V1_I_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
169 // CHECK:   [[VABD_V3_I_I_I:%.*]] = bitcast <4 x i16> [[VABD_V2_I_I_I]] to <8 x i8>
170 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I_I_I]] to <4 x i16>
171 // CHECK:   [[TMP3:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>
172 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x i16>
181 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
182 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %c to <8 x i8>
183 // CHECK:   [[VABD_V_I_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
184 // CHECK:   [[VABD_V1_I_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
186 // CHECK:   [[VABD_V3_I_I_I:%.*]] = bitcast <2 x i32> [[VABD_V2_I_I_I]] to <8 x i8>
187 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I_I_I]] to <2 x i32>
188 // CHECK:   [[TMP3:%.*]] = bitcast <2 x i32> [[TMP2]] to <8 x i8>
189 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <2 x i32>
207 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
208 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %c to <8 x i8>
209 // CHECK:   [[VABD_V_I_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
210 // CHECK:   [[VABD_V1_I_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
212 // CHECK:   [[VABD_V3_I_I_I:%.*]] = bitcast <4 x i16> [[VABD_V2_I_I_I]] to <8 x i8>
213 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I_I_I]] to <4 x i16>
214 // CHECK:   [[TMP3:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>
215 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x i16>
224 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
225 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %c to <8 x i8>
226 // CHECK:   [[VABD_V_I_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
227 // CHECK:   [[VABD_V1_I_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
229 // CHECK:   [[VABD_V3_I_I_I:%.*]] = bitcast <2 x i32> [[VABD_V2_I_I_I]] to <8 x i8>
230 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I_I_I]] to <2 x i32>
231 // CHECK:   [[TMP3:%.*]] = bitcast <2 x i32> [[TMP2]] to <8 x i8>
232 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <2 x i32>
249 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
250 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
251 // CHECK:   [[VABD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
252 // CHECK:   [[VABD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
254 // CHECK:   [[VABD_V3_I:%.*]] = bitcast <4 x i16> [[VABD_V2_I]] to <8 x i8>
255 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I]] to <4 x i16>
262 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
263 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
264 // CHECK:   [[VABD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
265 // CHECK:   [[VABD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
267 // CHECK:   [[VABD_V3_I:%.*]] = bitcast <2 x i32> [[VABD_V2_I]] to <8 x i8>
268 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I]] to <2 x i32>
282 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
283 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
284 // CHECK:   [[VABD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
285 // CHECK:   [[VABD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
287 // CHECK:   [[VABD_V3_I:%.*]] = bitcast <4 x i16> [[VABD_V2_I]] to <8 x i8>
288 // CHECK:   [[TMP2:%.*]] = bitcast
295 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
296 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
297 // CHECK:   [[VABD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
298 // CHECK:   [[VABD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
300 // CHECK:   [[VABD_V3_I:%.*]] = bitcast <2 x i32> [[VABD_V2_I]] to <8 x i8>
301 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I]] to <2 x i32>
308 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
309 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
310 // CHECK:   [[VABD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
311 // CHECK:   [[VABD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
313 // CHECK:   [[VABD_V3_I:%.*]] = bitcast <2 x float> [[VABD_V2_I]] to <8 x i8>
314 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I]] to <2 x float>
328 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
329 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
330 // CHECK:   [[VABDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
331 // CHECK:   [[VABDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
333 // CHECK:   [[VABDQ_V3_I:%.*]] = bitcast <8 x i16> [[VABDQ_V2_I]] to <16 x i8>
334 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VABDQ_V3_I]] to <8 x i16>
341 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
342 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
343 // CHECK:   [[VABDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
344 // CHECK:   [[VABDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
346 // CHECK:   [[VABDQ_V3_I:%.*]] = bitcast <4 x i32> [[VABDQ_V2_I]] to <16 x i8>
347 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VABDQ_V3_I]] to <4 x i32>
361 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
362 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
363 // CHECK:   [[VABDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
364 // CHECK:   [[VABDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
366 // CHECK:   [[VABDQ_V3_I:%.*]] = bitcast <8 x i16> [[VABDQ_V2_I]] to <16 x i8>
367 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VABDQ_V3_I]] to <8 x i16>
374 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
375 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
376 // CHECK:   [[VABDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
377 // CHECK:   [[VABDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
379 // CHECK:   [[VABDQ_V3_I:%.*]] = bitcast <4 x i32> [[VABDQ_V2_I]] to <16 x i8>
380 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VABDQ_V3_I]] to <4 x i32>
387 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
388 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
389 // CHECK:   [[VABDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
390 // CHECK:   [[VABDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
392 // CHECK:   [[VABDQ_V3_I:%.*]] = bitcast <4 x float> [[VABDQ_V2_I]] to <16 x i8>
393 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VABDQ_V3_I]] to <4 x float>
409 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
410 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
411 // CHECK:   [[VABD_V_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
412 // CHECK:   [[VABD_V1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
414 // CHECK:   [[VABD_V3_I_I:%.*]] = bitcast <4 x i16> [[VABD_V2_I_I]] to <8 x i8>
415 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I_I]] to <4 x i16>
416 // CHECK:   [[TMP3:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>
417 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x i16>
425 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
426 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
427 // CHECK:   [[VABD_V_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
428 // CHECK:   [[VABD_V1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
430 // CHECK:   [[VABD_V3_I_I:%.*]] = bitcast <2 x i32> [[VABD_V2_I_I]] to <8 x i8>
431 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I_I]] to <2 x i32>
432 // CHECK:   [[TMP3:%.*]] = bitcast <2 x i32> [[TMP2]] to <8 x i8>
433 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <2 x i32>
449 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
450 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
451 // CHECK:   [[VABD_V_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
452 // CHECK:   [[VABD_V1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
454 // CHECK:   [[VABD_V3_I_I:%.*]] = bitcast <4 x i16> [[VABD_V2_I_I]] to <8 x i8>
455 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I_I]] to <4 x i16>
456 // CHECK:   [[TMP3:%.*]] = bitcast <4 x i16> [[TMP2]] to <8 x i8>
457 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x i16>
465 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
466 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
467 // CHECK:   [[VABD_V_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
468 // CHECK:   [[VABD_V1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
470 // CHECK:   [[VABD_V3_I_I:%.*]] = bitcast <2 x i32> [[VABD_V2_I_I]] to <8 x i8>
471 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VABD_V3_I_I]] to <2 x i32>
472 // CHECK:   [[TMP3:%.*]] = bitcast <2 x i32> [[TMP2]] to <8 x i8>
473 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <2 x i32>
489 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
490 // CHECK:   [[VABS_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
498 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
499 // CHECK:   [[VABS_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
507 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
508 // CHECK:   [[VABS_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
523 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
524 // CHECK:   [[VABS_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
532 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
533 // CHECK:   [[VABS_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
541 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
542 // CHECK:   [[VABS_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
678 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
679 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
680 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
681 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
691 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
692 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
693 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
694 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
704 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
705 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
706 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
707 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
717 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
718 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
719 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
720 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
730 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
731 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
732 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
733 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
743 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
744 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
745 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
746 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
766 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
767 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
769 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %b to <8 x i8>
770 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
779 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
780 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
782 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %b to <8 x i8>
783 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
801 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
802 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
804 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %b to <8 x i8>
805 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
814 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
815 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
817 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %b to <8 x i8>
818 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
836 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
837 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
846 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
847 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
864 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
865 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
874 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
875 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
1134 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
1135 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
1136 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %c to <8 x i8>
1138 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[VBSL_V_I]] to <4 x i16>
1145 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
1146 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
1147 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %c to <8 x i8>
1149 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[VBSL_V_I]] to <2 x i32>
1156 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
1157 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
1158 // CHECK:   [[TMP2:%.*]] = bitcast <1 x i64> %c to <8 x i8>
1160 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[VBSL_V_I]] to <1 x i64>
1174 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
1175 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
1176 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %c to <8 x i8>
1178 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[VBSL_V_I]] to <4 x i16>
1185 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
1186 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
1187 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %c to <8 x i8>
1189 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[VBSL_V_I]] to <2 x i32>
1196 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
1197 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
1198 // CHECK:   [[TMP2:%.*]] = bitcast <1 x i64> %c to <8 x i8>
1200 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[VBSL_V_I]] to <1 x i64>
1207 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
1208 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
1209 // CHECK:   [[TMP2:%.*]] = bitcast <2 x float> %c to <8 x i8>
1211 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[VBSL_V_I]] to <2 x float>
1225 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
1226 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
1227 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %c to <8 x i8>
1229 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[VBSL_V_I]] to <4 x i16>
1243 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
1244 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
1245 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i16> %c to <16 x i8>
1247 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[VBSLQ_V_I]] to <8 x i16>
1254 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
1255 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
1256 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i32> %c to <16 x i8>
1258 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[VBSLQ_V_I]] to <4 x i32>
1265 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
1266 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
1267 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i64> %c to <16 x i8>
1269 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[VBSLQ_V_I]] to <2 x i64>
1283 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
1284 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
1285 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i16> %c to <16 x i8>
1287 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[VBSLQ_V_I]] to <8 x i16>
1294 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
1295 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
1296 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i32> %c to <16 x i8>
1298 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[VBSLQ_V_I]] to <4 x i32>
1305 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
1306 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
1307 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i64> %c to <16 x i8>
1309 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[VBSLQ_V_I]] to <2 x i64>
1316 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
1317 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
1318 // CHECK:   [[TMP2:%.*]] = bitcast <4 x float> %c to <16 x i8>
1320 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[VBSLQ_V_I]] to <4 x float>
1334 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
1335 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
1336 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i16> %c to <16 x i8>
1338 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[VBSLQ_V_I]] to <8 x i16>
1346 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
1347 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
1348 // CHECK:   [[VCAGE_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
1349 // CHECK:   [[VCAGE_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
1357 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
1358 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
1359 // CHECK:   [[VCAGEQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
1360 // CHECK:   [[VCAGEQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
1369 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
1370 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
1371 // CHECK:   [[VCAGT_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
1372 // CHECK:   [[VCAGT_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
1380 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
1381 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
1382 // CHECK:   [[VCAGTQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
1383 // CHECK:   [[VCAGTQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
1392 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
1393 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
1394 // CHECK:   [[VCALE_V_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
1395 // CHECK:   [[VCALE_V1_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
1403 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
1404 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
1405 // CHECK:   [[VCALEQ_V_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
1406 // CHECK:   [[VCALEQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
1415 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
1416 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
1417 // CHECK:   [[VCALT_V_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
1418 // CHECK:   [[VCALT_V1_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
1426 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
1427 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
1428 // CHECK:   [[VCALTQ_V_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
1429 // CHECK:   [[VCALTQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
1913 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
1914 // CHECK:   [[VCLS_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
1916 // CHECK:   [[VCLS_V2_I:%.*]] = bitcast <4 x i16> [[VCLS_V1_I]] to <8 x i8>
1917 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VCLS_V2_I]] to <4 x i16>
1924 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
1925 // CHECK:   [[VCLS_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
1927 // CHECK:   [[VCLS_V2_I:%.*]] = bitcast <2 x i32> [[VCLS_V1_I]] to <8 x i8>
1928 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VCLS_V2_I]] to <2 x i32>
1942 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
1943 // CHECK:   [[VCLSQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
1945 // CHECK:   [[VCLSQ_V2_I:%.*]] = bitcast <8 x i16> [[VCLSQ_V1_I]] to <16 x i8>
1946 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[VCLSQ_V2_I]] to <8 x i16>
1953 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
1954 // CHECK:   [[VCLSQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
1956 // CHECK:   [[VCLSQ_V2_I:%.*]] = bitcast <4 x i32> [[VCLSQ_V1_I]] to <16 x i8>
1957 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[VCLSQ_V2_I]] to <4 x i32>
2085 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
2086 // CHECK:   [[VCLZ_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
2088 // CHECK:   [[VCLZ_V2_I:%.*]] = bitcast <4 x i16> [[VCLZ_V1_I]] to <8 x i8>
2089 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VCLZ_V2_I]] to <4 x i16>
2096 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
2097 // CHECK:   [[VCLZ_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
2099 // CHECK:   [[VCLZ_V2_I:%.*]] = bitcast <2 x i32> [[VCLZ_V1_I]] to <8 x i8>
2100 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VCLZ_V2_I]] to <2 x i32>
2114 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
2115 // CHECK:   [[VCLZ_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
2117 // CHECK:   [[VCLZ_V2_I:%.*]] = bitcast <4 x i16> [[VCLZ_V1_I]] to <8 x i8>
2118 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VCLZ_V2_I]] to <4 x i16>
2125 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
2126 // CHECK:   [[VCLZ_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
2128 // CHECK:   [[VCLZ_V2_I:%.*]] = bitcast <2 x i32> [[VCLZ_V1_I]] to <8 x i8>
2129 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VCLZ_V2_I]] to <2 x i32>
2143 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
2144 // CHECK:   [[VCLZQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
2146 // CHECK:   [[VCLZQ_V2_I:%.*]] = bitcast <8 x i16> [[VCLZQ_V1_I]] to <16 x i8>
2147 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[VCLZQ_V2_I]] to <8 x i16>
2154 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
2155 // CHECK:   [[VCLZQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
2157 // CHECK:   [[VCLZQ_V2_I:%.*]] = bitcast <4 x i32> [[VCLZQ_V1_I]] to <16 x i8>
2158 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[VCLZQ_V2_I]] to <4 x i32>
2172 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
2173 // CHECK:   [[VCLZQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
2175 // CHECK:   [[VCLZQ_V2_I:%.*]] = bitcast <8 x i16> [[VCLZQ_V1_I]] to <16 x i8>
2176 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[VCLZQ_V2_I]] to <8 x i16>
2183 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
2184 // CHECK:   [[VCLZQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
2186 // CHECK:   [[VCLZQ_V2_I:%.*]] = bitcast <4 x i32> [[VCLZQ_V1_I]] to <16 x i8>
2187 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[VCLZQ_V2_I]] to <4 x i32>
2323 // CHECK:   [[TMP0:%.*]] = bitcast i64 %a to <8 x i8>
2331 // CHECK:   [[TMP0:%.*]] = bitcast i64 %a to <4 x i16>
2332 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>
2333 // CHECK:   [[VCLZ_V_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
2335 // CHECK:   [[VCLZ_V2_I:%.*]] = bitcast <4 x i16> [[VCLZ_V1_I]] to <8 x i8>
2336 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VCLZ_V2_I]] to <4 x i16>
2343 // CHECK:   [[TMP0:%.*]] = bitcast i64 %a to <2 x i32>
2344 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[TMP0]] to <8 x i8>
2345 // CHECK:   [[VCLZ_V_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
2347 // CHECK:   [[VCLZ_V2_I:%.*]] = bitcast <2 x i32> [[VCLZ_V1_I]] to <8 x i8>
2348 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VCLZ_V2_I]] to <2 x i32>
2355 // CHECK:   [[TMP0:%.*]] = bitcast i64 %a to <4 x half>
2362 // CHECK:   [[TMP0:%.*]] = bitcast i64 %a to <2 x float>
2369 // CHECK:   [[TMP0:%.*]] = bitcast i64 %a to <8 x i8>
2377 // CHECK:   [[TMP0:%.*]] = bitcast i64 %a to <4 x i16>
2378 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>
2379 // CHECK:   [[VCLZ_V_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
2381 // CHECK:   [[VCLZ_V2_I:%.*]] = bitcast <4 x i16> [[VCLZ_V1_I]] to <8 x i8>
2382 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VCLZ_V2_I]] to <4 x i16>
2389 // CHECK:   [[TMP0:%.*]] = bitcast i64 %a to <2 x i32>
2390 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[TMP0]] to <8 x i8>
2391 // CHECK:   [[VCLZ_V_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
2393 // CHECK:   [[VCLZ_V2_I:%.*]] = bitcast <2 x i32> [[VCLZ_V1_I]] to <8 x i8>
2394 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VCLZ_V2_I]] to <2 x i32>
2405 // CHECK:   [[TMP0:%.*]] = bitcast i64 %a to <1 x i64>
2415 // CHECK:   [[TMP0:%.*]] = bitcast i64 %a to <8 x i8>
2423 // CHECK:   [[TMP0:%.*]] = bitcast i64 %a to <4 x i16>
2424 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>
2425 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>
2426 // CHECK:   [[TMP3:%.*]] = bitcast <4 x i16> [[TMP0]] to <8 x i8>
2428 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[VBSL_V_I]] to <4 x i16>
2436 // CHECK:   [[TMP0:%.*]] = bitcast i64 %a to <1 x i64>
2446 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
2447 // CHECK:   [[VCVT_F16_F32_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
2449 // CHECK:   [[VCVT_F16_F322_I:%.*]] = bitcast <4 x i16> [[VCVT_F16_F321_I]] to <8 x i8>
2450 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VCVT_F16_F322_I]] to <4 x half>
2458 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
2459 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
2467 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
2468 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
2476 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
2477 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
2485 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
2486 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
2495 // CHECK:   [[TMP0:%.*]] = bitcast <4 x half> %a to <8 x i8>
2496 // CHECK:   [[VCVT_F32_F16_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
2498 // CHECK:   [[VCVT_F32_F162_I:%.*]] = bitcast <4 x float> [[VCVT_F32_F161_I]] to <16 x i8>
2499 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[VCVT_F32_F162_I]] to <4 x float>
2507 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
2508 // CHECK:   [[VCVT_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
2516 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
2517 // CHECK:   [[VCVT_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
2525 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
2526 // CHECK:   [[VCVT_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
2534 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
2535 // CHECK:   [[VCVT_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
2544 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
2545 // CHECK:   [[VCVT_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
2553 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
2554 // CHECK:   [[VCVT_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
2563 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
2564 // CHECK:   [[VCVT_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
2572 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
2573 // CHECK:   [[VCVT_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
2582 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
2583 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
2591 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
2592 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
2601 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
2602 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
2610 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
2611 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
3209 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
3210 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
3211 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
3212 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
3220 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
3221 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
3222 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
3223 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
3231 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
3232 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
3233 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
3234 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
3242 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
3243 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
3244 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
3245 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
3253 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
3254 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
3255 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
3256 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
3264 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
3265 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
3266 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
3267 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
3275 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
3276 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
3277 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
3278 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
3286 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
3287 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
3288 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
3289 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
3318 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
3319 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
3320 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
3321 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
3329 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
3330 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
3331 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
3332 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
3340 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
3341 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
3342 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
3343 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
3351 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
3352 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
3353 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
3354 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
3362 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
3363 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
3364 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
3365 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
3373 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
3374 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
3375 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
3376 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
3384 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
3385 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
3386 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
3387 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
3395 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
3396 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
3397 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
3398 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
3407 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
3408 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
3409 // CHECK:   [[TMP2:%.*]] = bitcast <2 x float> %c to <8 x i8>
3410 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
3411 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
3412 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x float>
3420 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
3421 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
3422 // CHECK:   [[TMP2:%.*]] = bitcast <4 x float> %c to <16 x i8>
3423 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
3424 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
3425 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <4 x float>
3434 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
3435 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> [[SUB_I]] to <8 x i8>
3436 // CHECK:   [[TMP2:%.*]] = bitcast <2 x float> %c to <8 x i8>
3437 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
3438 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
3439 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x float>
3448 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
3449 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> [[SUB_I]] to <16 x i8>
3450 // CHECK:   [[TMP2:%.*]] = bitcast <4 x float> %c to <16 x i8>
3451 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
3452 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
3453 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <4 x float>
3554 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
3555 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
3563 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
3564 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
3579 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
3580 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
3588 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
3589 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
3604 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
3605 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
3613 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
3614 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
3625 // CHECK:   [[TMP0:%.*]] = bitcast <4 x half>* [[__REINT_242]] to <4 x i16>*
3627 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> [[TMP1]] to <8 x i8>
3628 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
3631 // CHECK:   [[TMP4:%.*]] = bitcast i16* [[__REINT1_242]] to half*
3647 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
3648 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
3656 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
3657 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
3672 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
3673 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
3681 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
3682 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
3697 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
3698 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
3706 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
3707 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
3718 // CHECK:   [[TMP0:%.*]] = bitcast <8 x half>* [[__REINT_244]] to <8 x i16>*
3720 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i16> [[TMP1]] to <16 x i8>
3721 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP2]] to <8 x i16>
3724 // CHECK:   [[TMP4:%.*]] = bitcast i16* [[__REINT1_244]] to half*
3734 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
3735 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
3744 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
3745 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
3753 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
3754 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
3762 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
3763 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
3864 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
3865 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
3866 // CHECK:   [[VHADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
3867 // CHECK:   [[VHADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
3869 // CHECK:   [[VHADD_V3_I:%.*]] = bitcast <4 x i16> [[VHADD_V2_I]] to <8 x i8>
3870 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VHADD_V3_I]] to <4 x i16>
3877 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
3878 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
3879 // CHECK:   [[VHADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
3880 // CHECK:   [[VHADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
3882 // CHECK:   [[VHADD_V3_I:%.*]] = bitcast <2 x i32> [[VHADD_V2_I]] to <8 x i8>
3883 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VHADD_V3_I]] to <2 x i32>
3897 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
3898 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
3899 // CHECK:   [[VHADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
3900 // CHECK:   [[VHADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
3902 // CHECK:   [[VHADD_V3_I:%.*]] = bitcast <4 x i16> [[VHADD_V2_I]] to <8 x i8>
3903 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VHADD_V3_I]] to <4 x i16>
3910 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
3911 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
3912 // CHECK:   [[VHADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
3913 // CHECK:   [[VHADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
3915 // CHECK:   [[VHADD_V3_I:%.*]] = bitcast <2 x i32> [[VHADD_V2_I]] to <8 x i8>
3916 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VHADD_V3_I]] to <2 x i32>
3930 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
3931 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
3932 // CHECK:   [[VHADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
3933 // CHECK:   [[VHADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
3935 // CHECK:   [[VHADDQ_V3_I:%.*]] = bitcast <8 x i16> [[VHADDQ_V2_I]] to <16 x i8>
3936 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VHADDQ_V3_I]] to <8 x i16>
3943 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
3944 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
3945 // CHECK:   [[VHADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
3946 // CHECK:   [[VHADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
3948 // CHECK:   [[VHADDQ_V3_I:%.*]] = bitcast <4 x i32> [[VHADDQ_V2_I]] to <16 x i8>
3949 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VHADDQ_V3_I]] to <4 x i32>
3963 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
3964 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
3965 // CHECK:   [[VHADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
3966 // CHECK:   [[VHADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
3968 // CHECK:   [[VHADDQ_V3_I:%.*]] = bitcast <8 x i16> [[VHADDQ_V2_I]] to <16 x i8>
3969 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VHADDQ_V3_I]] to <8 x i16>
3976 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
3977 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
3978 // CHECK:   [[VHADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
3979 // CHECK:   [[VHADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
3981 // CHECK:   [[VHADDQ_V3_I:%.*]] = bitcast <4 x i32> [[VHADDQ_V2_I]] to <16 x i8>
3982 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VHADDQ_V3_I]] to <4 x i32>
3997 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
3998 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
3999 // CHECK:   [[VHSUB_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
4000 // CHECK:   [[VHSUB_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
4002 // CHECK:   [[VHSUB_V3_I:%.*]] = bitcast <4 x i16> [[VHSUB_V2_I]] to <8 x i8>
4003 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VHSUB_V3_I]] to <4 x i16>
4010 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
4011 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
4012 // CHECK:   [[VHSUB_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
4013 // CHECK:   [[VHSUB_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
4015 // CHECK:   [[VHSUB_V3_I:%.*]] = bitcast <2 x i32> [[VHSUB_V2_I]] to <8 x i8>
4016 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VHSUB_V3_I]] to <2 x i32>
4030 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
4031 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
4032 // CHECK:   [[VHSUB_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
4033 // CHECK:   [[VHSUB_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
4035 // CHECK:   [[VHSUB_V3_I:%.*]] = bitcast <4 x i16> [[VHSUB_V2_I]] to <8 x i8>
4036 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VHSUB_V3_I]] to <4 x i16>
4043 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
4044 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
4045 // CHECK:   [[VHSUB_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
4046 // CHECK:   [[VHSUB_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
4048 // CHECK:   [[VHSUB_V3_I:%.*]] = bitcast <2 x i32> [[VHSUB_V2_I]] to <8 x i8>
4049 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VHSUB_V3_I]] to <2 x i32>
4063 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
4064 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
4065 // CHECK:   [[VHSUBQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
4066 // CHECK:   [[VHSUBQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
4068 // CHECK:   [[VHSUBQ_V3_I:%.*]] = bitcast <8 x i16> [[VHSUBQ_V2_I]] to <16 x i8>
4069 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VHSUBQ_V3_I]] to <8 x i16>
4076 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
4077 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
4078 // CHECK:   [[VHSUBQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
4079 // CHECK:   [[VHSUBQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
4081 // CHECK:   [[VHSUBQ_V3_I:%.*]] = bitcast <4 x i32> [[VHSUBQ_V2_I]] to <16 x i8>
4082 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VHSUBQ_V3_I]] to <4 x i32>
4096 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
4097 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
4098 // CHECK:   [[VHSUBQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
4099 // CHECK:   [[VHSUBQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
4101 // CHECK:   [[VHSUBQ_V3_I:%.*]] = bitcast <8 x i16> [[VHSUBQ_V2_I]] to <16 x i8>
4102 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VHSUBQ_V3_I]] to <8 x i16>
4109 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
4110 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
4111 // CHECK:   [[VHSUBQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
4112 // CHECK:   [[VHSUBQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
4114 // CHECK:   [[VHSUBQ_V3_I:%.*]] = bitcast <4 x i32> [[VHSUBQ_V2_I]] to <16 x i8>
4115 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VHSUBQ_V3_I]] to <4 x i32>
4130 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4138 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
4146 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
4161 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4169 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
4177 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
4185 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
4187 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> [[VLD1]] to <8 x half>
4194 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
4209 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4224 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4232 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
4240 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
4255 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4263 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
4271 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
4279 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
4281 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[VLD1]] to <4 x half>
4288 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
4303 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4321 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4322 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i16*
4332 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
4333 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i32*
4343 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
4344 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i64*
4363 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4364 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i16*
4374 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
4375 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i32*
4385 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
4386 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i64*
4396 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
4397 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i16*
4401 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i16> [[LANE]] to <8 x half>
4408 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
4409 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to float*
4428 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4429 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i16*
4448 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4449 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i16*
4459 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
4460 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i32*
4470 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
4471 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i64*
4490 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4491 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i16*
4501 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
4502 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i32*
4512 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
4513 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i64*
4523 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
4524 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i16*
4528 // CHECK:   [[TMP4:%.*]] = bitcast <4 x i16> [[LANE]] to <4 x half>
4535 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
4536 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to float*
4555 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4556 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to i16*
4575 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4576 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
4577 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
4578 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i16*
4587 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
4588 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
4589 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
4590 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i32*
4599 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
4600 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
4601 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
4619 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4620 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
4621 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
4622 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i16*
4631 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
4632 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
4633 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
4634 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i32*
4643 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
4644 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
4645 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
4655 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
4656 // CHECK:   [[TMP1:%.*]] = bitcast <8 x half> %b to <16 x i8>
4657 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
4658 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i16*
4661 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[VLD1_LANE]] to <8 x half>
4668 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
4669 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
4670 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
4671 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to float*
4688 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4689 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
4690 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
4691 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i16*
4708 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4709 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
4710 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
4711 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i16*
4720 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
4721 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
4722 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
4723 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i32*
4732 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
4733 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
4734 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
4735 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i64*
4752 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4753 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
4754 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
4755 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i16*
4764 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
4765 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
4766 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
4767 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i32*
4776 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
4777 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
4778 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
4779 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i64*
4788 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
4789 // CHECK:   [[TMP1:%.*]] = bitcast <4 x half> %b to <8 x i8>
4790 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
4791 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i16*
4794 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[VLD1_LANE]] to <4 x half>
4801 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
4802 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
4803 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
4804 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to float*
4821 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
4822 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
4823 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
4824 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to i16*
4835 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x16x2_t* [[__RET]] to i8*
4837 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <16 x i8>, <16 x i8> }*
4839 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x16x2_t* %agg.result to i8*
4840 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x16x2_t* [[__RET]] to i8*
4849 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x8x2_t* [[__RET]] to i8*
4850 // CHECK:   [[TMP1:%.*]] = bitcast
4852 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <8 x i16>, <8 x i16> }*
4854 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint16x8x2_t* %agg.result to i8*
4855 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint16x8x2_t* [[__RET]] to i8*
4864 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x4x2_t* [[__RET]] to i8*
4865 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
4867 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i32>, <4 x i32> }*
4869 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint32x4x2_t* %agg.result to i8*
4870 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint32x4x2_t* [[__RET]] to i8*
4879 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x16x2_t* [[__RET]] to i8*
4881 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <16 x i8>, <16 x i8> }*
4883 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x16x2_t* %agg.result to i8*
4884 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x16x2_t* [[__RET]] to i8*
4893 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x8x2_t* [[__RET]] to i8*
4894 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
4896 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <8 x i16>, <8 x i16> }*
4898 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int16x8x2_t* %agg.result to i8*
4899 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int16x8x2_t* [[__RET]] to i8*
4908 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x4x2_t* [[__RET]] to i8*
4909 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
4911 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i32>, <4 x i32> }*
4913 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int32x4x2_t* %agg.result to i8*
4914 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int32x4x2_t* [[__RET]] to i8*
4923 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x8x2_t* [[__RET]] to i8*
4924 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
4926 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <8 x i16>, <8 x i16> }*
4928 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float16x8x2_t* %agg.result to i8*
4929 // CHECK:   [[TMP4:%.*]] = bitcast %struct.float16x8x2_t* [[__RET]] to i8*
4938 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x4x2_t* [[__RET]] to i8*
4939 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*
4941 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x float>, <4 x float> }*
4943 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float32x4x2_t* %agg.result to i8*
4944 // CHECK:   [[TMP4:%.*]] = bitcast %struct.float32x4x2_t* [[__RET]] to i8*
4953 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x16x2_t* [[__RET]] to i8*
4955 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <16 x i8>, <16 x i8> }*
4957 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x16x2_t* %agg.result to i8*
4958 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x16x2_t* [[__RET]] to i8*
4967 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x8x2_t* [[__RET]] to i8*
4968 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
4970 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <8 x i16>, <8 x i16> }*
4972 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly16x8x2_t* %agg.result to i8*
4973 // CHECK:   [[TMP4:%.*]] = bitcast %struct.poly16x8x2_t* [[__RET]] to i8*
4982 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x8x2_t* [[__RET]] to i8*
4984 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8> }*
4986 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x8x2_t* %agg.result to i8*
4987 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x8x2_t* [[__RET]] to i8*
4996 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x4x2_t* [[__RET]] to i8*
4997 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
4999 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16> }*
5001 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint16x4x2_t* %agg.result to i8*
5002 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint16x4x2_t* [[__RET]] to i8*
5011 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x2x2_t* [[__RET]] to i8*
5012 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
5014 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <2 x i32>, <2 x i32> }*
5016 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint32x2x2_t* %agg.result to i8*
5017 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint32x2x2_t* [[__RET]] to i8*
5026 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint64x1x2_t* [[__RET]] to i8*
5027 // CHECK:   [[TMP1:%.*]] = bitcast i64* %a to i8*
5029 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <1 x i64>, <1 x i64> }*
5031 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint64x1x2_t* %agg.result to i8*
5032 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint64x1x2_t* [[__RET]] to i8*
5041 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x8x2_t* [[__RET]] to i8*
5043 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8> }*
5045 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x8x2_t* %agg.result to i8*
5046 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x8x2_t* [[__RET]] to i8*
5055 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x4x2_t* [[__RET]] to i8*
5056 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
5058 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16> }*
5060 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int16x4x2_t* %agg.result to i8*
5061 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int16x4x2_t* [[__RET]] to i8*
5070 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x2x2_t* [[__RET]] to i8*
5071 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
5073 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <2 x i32>, <2 x i32> }*
5075 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int32x2x2_t* %agg.result to i8*
5076 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int32x2x2_t* [[__RET]] to i8*
5085 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int64x1x2_t* [[__RET]] to i8*
5086 // CHECK:   [[TMP1:%.*]] = bitcast i64* %a to i8*
5088 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <1 x i64>, <1 x i64> }*
5090 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int64x1x2_t* %agg.result to i8*
5091 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int64x1x2_t* [[__RET]] to i8*
5100 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x4x2_t* [[__RET]] to i8*
5101 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
5103 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16> }*
5105 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float16x4x2_t* %agg.result to i8*
5106 // CHECK:   [[TMP4:%.*]] = bitcast %struct.float16x4x2_t* [[__RET]] to i8*
5115 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x2x2_t* [[__RET]] to i8*
5116 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*
5118 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <2 x float>, <2 x float> }*
5120 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float32x2x2_t* %agg.result to i8*
5121 // CHECK:   [[TMP4:%.*]] = bitcast %struct.float32x2x2_t* [[__RET]] to i8*
5130 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x8x2_t* [[__RET]] to i8*
5132 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8> }*
5134 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x8x2_t* %agg.result to i8*
5135 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x8x2_t* [[__RET]] to i8*
5144 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x4x2_t* [[__RET]] to i8*
5145 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
5147 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16> }*
5149 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly16x4x2_t* %agg.result to i8*
5150 // CHECK:   [[TMP4:%.*]] = bitcast %struct.poly16x4x2_t* [[__RET]] to i8*
5160 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x8x2_t* [[__RET]] to i8*
5168 // CHECK:   [[TMP5:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8> }*
5170 // CHECK:   [[TMP6:%.*]] = bitcast %struct.uint8x8x2_t* %agg.result to i8*
5171 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint8x8x2_t* [[__RET]] to i8*
5180 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x4x2_t* [[__RET]] to i8*
5181 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
5189 // CHECK:   [[TMP6:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16> }*
5191 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint16x4x2_t* %agg.result to i8*
5192 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint16x4x2_t* [[__RET]] to i8*
5201 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x2x2_t* [[__RET]] to i8*
5202 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
5210 // CHECK:   [[TMP6:%.*]] = bitcast i8* [[TMP0]] to { <2 x i32>, <2 x i32> }*
5212 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint32x2x2_t* %agg.result to i8*
5213 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint32x2x2_t* [[__RET]] to i8*
5222 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint64x1x2_t* [[__RET]] to i8*
5223 // CHECK:   [[TMP1:%.*]] = bitcast i64* %a to i8*
5225 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <1 x i64>, <1 x i64> }*
5227 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint64x1x2_t* %agg.result to i8*
5228 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint64x1x2_t* [[__RET]] to i8*
5237 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x8x2_t* [[__RET]] to i8*
5245 // CHECK:   [[TMP5:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8> }*
5247 // CHECK:   [[TMP6:%.*]] = bitcast %struct.int8x8x2_t* %agg.result to i8*
5248 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int8x8x2_t* [[__RET]] to i8*
5257 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x4x2_t* [[__RET]] to i8*
5258 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
5266 // CHECK:   [[TMP6:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16> }*
5268 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int16x4x2_t* %agg.result to i8*
5269 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int16x4x2_t* [[__RET]] to i8*
5278 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x2x2_t* [[__RET]] to i8*
5279 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
5287 // CHECK:   [[TMP6:%.*]] = bitcast i8* [[TMP0]] to { <2 x i32>, <2 x i32> }*
5289 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int32x2x2_t* %agg.result to i8*
5290 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int32x2x2_t* [[__RET]] to i8*
5299 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int64x1x2_t* [[__RET]] to i8*
5300 // CHECK:   [[TMP1:%.*]] = bitcast i64* %a to i8*
5302 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <1 x i64>, <1 x i64> }*
5304 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int64x1x2_t* %agg.result to i8*
5305 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int64x1x2_t* [[__RET]] to i8*
5314 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x4x2_t* [[__RET]] to i8*
5315 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
5323 // CHECK:   [[TMP6:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16> }*
5325 // CHECK:   [[TMP7:%.*]] = bitcast %struct.float16x4x2_t* %agg.result to i8*
5326 // CHECK:   [[TMP8:%.*]] = bitcast %struct.float16x4x2_t* [[__RET]] to i8*
5335 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x2x2_t* [[__RET]] to i8*
5336 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*
5344 // CHECK:   [[TMP6:%.*]] = bitcast i8* [[TMP0]] to { <2 x float>, <2 x float> }*
5346 // CHECK:   [[TMP7:%.*]] = bitcast %struct.float32x2x2_t* %agg.result to i8*
5347 // CHECK:   [[TMP8:%.*]] = bitcast %struct.float32x2x2_t* [[__RET]] to i8*
5356 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x8x2_t* [[__RET]] to i8*
5364 // CHECK:   [[TMP5:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8> }*
5366 // CHECK:   [[TMP6:%.*]] = bitcast %struct.poly8x8x2_t* %agg.result to i8*
5367 // CHECK:   [[TMP7:%.*]] = bitcast %struct.poly8x8x2_t* [[__RET]] to i8*
5376 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x4x2_t* [[__RET]] to i8*
5377 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
5385 // CHECK:   [[TMP6:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16> }*
5387 // CHECK:   [[TMP7:%.*]] = bitcast %struct.poly16x4x2_t* %agg.result to i8*
5388 // CHECK:   [[TMP8:%.*]] = bitcast %struct.poly16x4x2_t* [[__RET]] to i8*
5401 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
5403 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x8x2_t* [[__S1]] to i8*
5404 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x8x2_t* [[B]] to i8*
5406 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint16x8x2_t* [[__RET]] to i8*
5407 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
5411 // CHECK:   [[TMP6:%.*]] = bitcast <8 x i16> [[TMP5]] to <16 x i8>
5415 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i16> [[TMP7]] to <16 x i8>
5416 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
5417 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
5419 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <8 x i16>, <8 x i16> }*
5421 // CHECK:   [[TMP12:%.*]] = bitcast %struct.uint16x8x2_t* %agg.result to i8*
5422 // CHECK:   [[TMP13:%.*]] = bitcast %struct.uint16x8x2_t* [[__RET]] to i8*
5434 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i32>]* [[COERCE_DIVE]] to [4 x i64]*
5436 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x4x2_t* [[__S1]] to i8*
5437 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x4x2_t* [[B]] to i8*
5439 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint32x4x2_t* [[__RET]] to i8*
5440 // CHECK:   [[TMP4:%.*]] = bitcast i32* %a to i8*
5444 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i32> [[TMP5]] to <16 x i8>
5448 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i32> [[TMP7]] to <16 x i8>
5449 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP6]] to <4 x i32>
5450 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP8]] to <4 x i32>
5452 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <4 x i32>, <4 x i32> }*
5454 // CHECK:   [[TMP12:%.*]] = bitcast %struct.uint32x4x2_t* %agg.result to i8*
5455 // CHECK:   [[TMP13:%.*]] = bitcast %struct.uint32x4x2_t* [[__RET]] to i8*
5467 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
5469 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x8x2_t* [[__S1]] to i8*
5470 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x8x2_t* [[B]] to i8*
5472 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int16x8x2_t* [[__RET]] to i8*
5473 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
5477 // CHECK:   [[TMP6:%.*]] = bitcast <8 x i16> [[TMP5]] to <16 x i8>
5481 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i16> [[TMP7]] to <16 x i8>
5482 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
5483 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
5485 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <8 x i16>, <8 x i16> }*
5487 // CHECK:   [[TMP12:%.*]] = bitcast %struct.int16x8x2_t* %agg.result to i8*
5488 // CHECK:   [[TMP13:%.*]] = bitcast %struct.int16x8x2_t* [[__RET]] to i8*
5500 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i32>]* [[COERCE_DIVE]] to [4 x i64]*
5502 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x4x2_t* [[__S1]] to i8*
5503 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x4x2_t* [[B]] to i8*
5505 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int32x4x2_t* [[__RET]] to i8*
5506 // CHECK:   [[TMP4:%.*]] = bitcast i32* %a to i8*
5510 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i32> [[TMP5]] to <16 x i8>
5514 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i32> [[TMP7]] to <16 x i8>
5515 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP6]] to <4 x i32>
5516 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP8]] to <4 x i32>
5518 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <4 x i32>, <4 x i32> }*
5520 // CHECK:   [[TMP12:%.*]] = bitcast %struct.int32x4x2_t* %agg.result to i8*
5521 // CHECK:   [[TMP13:%.*]] = bitcast %struct.int32x4x2_t* [[__RET]] to i8*
5533 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x half>]* [[COERCE_DIVE]] to [4 x i64]*
5535 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x8x2_t* [[__S1]] to i8*
5536 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x8x2_t* [[B]] to i8*
5538 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float16x8x2_t* [[__RET]] to i8*
5539 // CHECK:   [[TMP4:%.*]] = bitcast half* %a to i8*
5543 // CHECK:   [[TMP6:%.*]] = bitcast <8 x half> [[TMP5]] to <16 x i8>
5547 // CHECK:   [[TMP8:%.*]] = bitcast <8 x half> [[TMP7]] to <16 x i8>
5548 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
5549 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
5551 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <8 x i16>, <8 x i16> }*
5553 // CHECK:   [[TMP12:%.*]] = bitcast %struct.float16x8x2_t* %agg.result to i8*
5554 // CHECK:   [[TMP13:%.*]] = bitcast %struct.float16x8x2_t* [[__RET]] to i8*
5566 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x float>]* [[COERCE_DIVE]] to [4 x i64]*
5568 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x4x2_t* [[__S1]] to i8*
5569 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x4x2_t* [[B]] to i8*
5571 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float32x4x2_t* [[__RET]] to i8*
5572 // CHECK:   [[TMP4:%.*]] = bitcast float* %a to i8*
5576 // CHECK:   [[TMP6:%.*]] = bitcast <4 x float> [[TMP5]] to <16 x i8>
5580 // CHECK:   [[TMP8:%.*]] = bitcast <4 x float> [[TMP7]] to <16 x i8>
5581 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP6]] to <4 x float>
5582 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP8]] to <4 x float>
5584 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <4 x float>, <4 x float> }*
5586 // CHECK:   [[TMP12:%.*]] = bitcast %struct.float32x4x2_t* %agg.result to i8*
5587 // CHECK:   [[TMP13:%.*]] = bitcast %struct.float32x4x2_t* [[__RET]] to i8*
5599 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
5601 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x8x2_t* [[__S1]] to i8*
5602 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x8x2_t* [[B]] to i8*
5604 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly16x8x2_t* [[__RET]] to i8*
5605 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
5609 // CHECK:   [[TMP6:%.*]] = bitcast <8 x i16> [[TMP5]] to <16 x i8>
5613 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i16> [[TMP7]] to <16 x i8>
5614 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
5615 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
5617 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <8 x i16>, <8 x i16> }*
5619 // CHECK:   [[TMP12:%.*]] = bitcast %struct.poly16x8x2_t* %agg.result to i8*
5620 // CHECK:   [[TMP13:%.*]] = bitcast %struct.poly16x8x2_t* [[__RET]] to i8*
5632 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
5634 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint8x8x2_t* [[__S1]] to i8*
5635 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x8x2_t* [[B]] to i8*
5637 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x8x2_t* [[__RET]] to i8*
5645 // CHECK:   [[TMP6:%.*]] = bitcast i8* [[TMP3]] to { <8 x i8>, <8 x i8> }*
5647 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint8x8x2_t* %agg.result to i8*
5648 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint8x8x2_t* [[__RET]] to i8*
5660 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i16>]* [[COERCE_DIVE]] to [2 x i64]*
5662 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x4x2_t* [[__S1]] to i8*
5663 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x4x2_t* [[B]] to i8*
5665 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint16x4x2_t* [[__RET]] to i8*
5666 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
5670 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i16> [[TMP5]] to <8 x i8>
5674 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i16> [[TMP7]] to <8 x i8>
5675 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
5676 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
5678 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <4 x i16>, <4 x i16> }*
5680 // CHECK:   [[TMP12:%.*]] = bitcast %struct.uint16x4x2_t* %agg.result to i8*
5681 // CHECK:   [[TMP13:%.*]] = bitcast %struct.uint16x4x2_t* [[__RET]] to i8*
5693 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <2 x i32>]* [[COERCE_DIVE]] to [2 x i64]*
5695 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x2x2_t* [[__S1]] to i8*
5696 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x2x2_t* [[B]] to i8*
5698 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint32x2x2_t* [[__RET]] to i8*
5699 // CHECK:   [[TMP4:%.*]] = bitcast i32* %a to i8*
5703 // CHECK:   [[TMP6:%.*]] = bitcast <2 x i32> [[TMP5]] to <8 x i8>
5707 // CHECK:   [[TMP8:%.*]] = bitcast <2 x i32> [[TMP7]] to <8 x i8>
5708 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP6]] to <2 x i32>
5709 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP8]] to <2 x i32>
5711 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <2 x i32>, <2 x i32> }*
5713 // CHECK:   [[TMP12:%.*]] = bitcast %struct.uint32x2x2_t* %agg.result to i8*
5714 // CHECK:   [[TMP13:%.*]] = bitcast %struct.uint32x2x2_t* [[__RET]] to i8*
5726 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
5728 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int8x8x2_t* [[__S1]] to i8*
5729 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x8x2_t* [[B]] to i8*
5731 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x8x2_t* [[__RET]] to i8*
5739 // CHECK:   [[TMP6:%.*]] = bitcast i8* [[TMP3]] to { <8 x i8>, <8 x i8> }*
5741 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int8x8x2_t* %agg.result to i8*
5742 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int8x8x2_t* [[__RET]] to i8*
5754 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i16>]* [[COERCE_DIVE]] to [2 x i64]*
5756 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x4x2_t* [[__S1]] to i8*
5757 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x4x2_t* [[B]] to i8*
5759 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int16x4x2_t* [[__RET]] to i8*
5760 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
5764 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i16> [[TMP5]] to <8 x i8>
5768 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i16> [[TMP7]] to <8 x i8>
5769 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
5770 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
5772 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <4 x i16>, <4 x i16> }*
5774 // CHECK:   [[TMP12:%.*]] = bitcast %struct.int16x4x2_t* %agg.result to i8*
5775 // CHECK:   [[TMP13:%.*]] = bitcast %struct.int16x4x2_t* [[__RET]] to i8*
5787 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <2 x i32>]* [[COERCE_DIVE]] to [2 x i64]*
5789 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x2x2_t* [[__S1]] to i8*
5790 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x2x2_t* [[B]] to i8*
5792 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int32x2x2_t* [[__RET]] to i8*
5793 // CHECK:   [[TMP4:%.*]] = bitcast i32* %a to i8*
5797 // CHECK:   [[TMP6:%.*]] = bitcast <2 x i32> [[TMP5]] to <8 x i8>
5801 // CHECK:   [[TMP8:%.*]] = bitcast <2 x i32> [[TMP7]] to <8 x i8>
5802 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP6]] to <2 x i32>
5803 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP8]] to <2 x i32>
5805 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <2 x i32>, <2 x i32> }*
5807 // CHECK:   [[TMP12:%.*]] = bitcast %struct.int32x2x2_t* %agg.result to i8*
5808 // CHECK:   [[TMP13:%.*]] = bitcast %struct.int32x2x2_t* [[__RET]] to i8*
5820 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x half>]* [[COERCE_DIVE]] to [2 x i64]*
5822 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x4x2_t* [[__S1]] to i8*
5823 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x4x2_t* [[B]] to i8*
5825 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float16x4x2_t* [[__RET]] to i8*
5826 // CHECK:   [[TMP4:%.*]] = bitcast half* %a to i8*
5830 // CHECK:   [[TMP6:%.*]] = bitcast <4 x half> [[TMP5]] to <8 x i8>
5834 // CHECK:   [[TMP8:%.*]] = bitcast <4 x half> [[TMP7]] to <8 x i8>
5835 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
5836 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
5838 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <4 x i16>, <4 x i16> }*
5840 // CHECK:   [[TMP12:%.*]] = bitcast %struct.float16x4x2_t* %agg.result to i8*
5841 // CHECK:   [[TMP13:%.*]] = bitcast %struct.float16x4x2_t* [[__RET]] to i8*
5853 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <2 x float>]* [[COERCE_DIVE]] to [2 x i64]*
5855 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x2x2_t* [[__S1]] to i8*
5856 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x2x2_t* [[B]] to i8*
5858 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float32x2x2_t* [[__RET]] to i8*
5859 // CHECK:   [[TMP4:%.*]] = bitcast float* %a to i8*
5863 // CHECK:   [[TMP6:%.*]] = bitcast <2 x float> [[TMP5]] to <8 x i8>
5867 // CHECK:   [[TMP8:%.*]] = bitcast <2 x float> [[TMP7]] to <8 x i8>
5868 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP6]] to <2 x float>
5869 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP8]] to <2 x float>
5871 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <2 x float>, <2 x float> }*
5873 // CHECK:   [[TMP12:%.*]] = bitcast %struct.float32x2x2_t* %agg.result to i8*
5874 // CHECK:   [[TMP13:%.*]] = bitcast %struct.float32x2x2_t* [[__RET]] to i8*
5886 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
5888 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly8x8x2_t* [[__S1]] to i8*
5889 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x8x2_t* [[B]] to i8*
5891 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x8x2_t* [[__RET]] to i8*
5899 // CHECK:   [[TMP6:%.*]] = bitcast i8* [[TMP3]] to { <8 x i8>, <8 x i8> }*
5901 // CHECK:   [[TMP7:%.*]] = bitcast %struct.poly8x8x2_t* %agg.result to i8*
5902 // CHECK:   [[TMP8:%.*]] = bitcast %struct.poly8x8x2_t* [[__RET]] to i8*
5914 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i16>]* [[COERCE_DIVE]] to [2 x i64]*
5916 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x4x2_t* [[__S1]] to i8*
5917 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x4x2_t* [[B]] to i8*
5919 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly16x4x2_t* [[__RET]] to i8*
5920 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
5924 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i16> [[TMP5]] to <8 x i8>
5928 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i16> [[TMP7]] to <8 x i8>
5929 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
5930 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
5932 // CHECK:   [[TMP11:%.*]] = bitcast i8* [[TMP3]] to { <4 x i16>, <4 x i16> }*
5934 // CHECK:   [[TMP12:%.*]] = bitcast %struct.poly16x4x2_t* %agg.result to i8*
5935 // CHECK:   [[TMP13:%.*]] = bitcast %struct.poly16x4x2_t* [[__RET]] to i8*
5945 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x16x3_t* [[__RET]] to i8*
5947 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <16 x i8>, <16 x i8>, <16 x i8> }*
5949 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x16x3_t* %agg.result to i8*
5950 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x16x3_t* [[__RET]] to i8*
5959 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x8x3_t* [[__RET]] to i8*
5960 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
5962 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <8 x i16>, <8 x i16>, <8 x i16> }*
5964 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint16x8x3_t* %agg.result to i8*
5965 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint16x8x3_t* [[__RET]] to i8*
5974 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x4x3_t* [[__RET]] to i8*
5975 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
5977 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i32>, <4 x i32>, <4 x i32> }*
5979 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint32x4x3_t* %agg.result to i8*
5980 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint32x4x3_t* [[__RET]] to i8*
5989 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x16x3_t* [[__RET]] to i8*
5991 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <16 x i8>, <16 x i8>, <16 x i8> }*
5993 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x16x3_t* %agg.result to i8*
5994 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x16x3_t* [[__RET]] to i8*
6003 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x8x3_t* [[__RET]] to i8*
6004 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
6006 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <8 x i16>, <8 x i16>, <8 x i16> }*
6008 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int16x8x3_t* %agg.result to i8*
6009 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int16x8x3_t* [[__RET]] to i8*
6018 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x4x3_t* [[__RET]] to i8*
6019 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
6021 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i32>, <4 x i32>, <4 x i32> }*
6023 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int32x4x3_t* %agg.result to i8*
6024 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int32x4x3_t* [[__RET]] to i8*
6033 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x8x3_t* [[__RET]] to i8*
6034 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
6036 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <8 x i16>, <8 x i16>, <8 x i16> }*
6038 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float16x8x3_t* %agg.result to i8*
6039 // CHECK:   [[TMP4:%.*]] = bitcast %struct.float16x8x3_t* [[__RET]] to i8*
6048 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x4x3_t* [[__RET]] to i8*
6049 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*
6051 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x float>, <4 x float>, <4 x float> }*
6053 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float32x4x3_t* %agg.result to i8*
6054 // CHECK:   [[TMP4:%.*]] = bitcast %struct.float32x4x3_t* [[__RET]] to i8*
6063 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x16x3_t* [[__RET]] to i8*
6065 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <16 x i8>, <16 x i8>, <16 x i8> }*
6067 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x16x3_t* %agg.result to i8*
6068 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x16x3_t* [[__RET]] to i8*
6077 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x8x3_t* [[__RET]] to i8*
6078 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
6080 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <8 x i16>, <8 x i16>, <8 x i16> }*
6082 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly16x8x3_t* %agg.result to i8*
6083 // CHECK:   [[TMP4:%.*]] = bitcast %struct.poly16x8x3_t* [[__RET]] to i8*
6092 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x8x3_t* [[__RET]] to i8*
6094 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8>, <8 x i8> }*
6096 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x8x3_t* %agg.result to i8*
6097 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x8x3_t* [[__RET]] to i8*
6106 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x4x3_t* [[__RET]] to i8*
6107 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
6109 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16> }*
6111 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint16x4x3_t* %agg.result to i8*
6112 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint16x4x3_t* [[__RET]] to i8*
6121 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x2x3_t* [[__RET]] to i8*
6122 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
6124 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <2 x i32>, <2 x i32>, <2 x i32> }*
6126 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint32x2x3_t* %agg.result to i8*
6127 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint32x2x3_t* [[__RET]] to i8*
6136 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint64x1x3_t* [[__RET]] to i8*
6137 // CHECK:   [[TMP1:%.*]] = bitcast i64* %a to i8*
6139 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <1 x i64>, <1 x i64>, <1 x i64> }*
6141 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint64x1x3_t* %agg.result to i8*
6142 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint64x1x3_t* [[__RET]] to i8*
6151 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x8x3_t* [[__RET]] to i8*
6153 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8>, <8 x i8> }*
6155 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x8x3_t* %agg.result to i8*
6156 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x8x3_t* [[__RET]] to i8*
6165 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x4x3_t* [[__RET]] to i8*
6166 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
6168 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16> }*
6170 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int16x4x3_t* %agg.result to i8*
6171 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int16x4x3_t* [[__RET]] to i8*
6180 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x2x3_t* [[__RET]] to i8*
6181 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
6183 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <2 x i32>, <2 x i32>, <2 x i32> }*
6185 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int32x2x3_t* %agg.result to i8*
6186 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int32x2x3_t* [[__RET]] to i8*
6195 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int64x1x3_t* [[__RET]] to i8*
6196 // CHECK:   [[TMP1:%.*]] = bitcast i64* %a to i8*
6198 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <1 x i64>, <1 x i64>, <1 x i64> }*
6200 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int64x1x3_t* %agg.result to i8*
6201 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int64x1x3_t* [[__RET]] to i8*
6210 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x4x3_t* [[__RET]] to i8*
6211 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
6213 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16> }*
6215 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float16x4x3_t* %agg.result to i8*
6216 // CHECK:   [[TMP4:%.*]] = bitcast %struct.float16x4x3_t* [[__RET]] to i8*
6225 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x2x3_t* [[__RET]] to i8*
6226 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*
6228 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <2 x float>, <2 x float>, <2 x float> }*
6230 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float32x2x3_t* %agg.result to i8*
6231 // CHECK:   [[TMP4:%.*]] = bitcast %struct.float32x2x3_t* [[__RET]] to i8*
6240 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x8x3_t* [[__RET]] to i8*
6242 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8>, <8 x i8> }*
6244 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x8x3_t* %agg.result to i8*
6245 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x8x3_t* [[__RET]] to i8*
6254 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x4x3_t* [[__RET]] to i8*
6255 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
6257 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16> }*
6259 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly16x4x3_t* %agg.result to i8*
6260 // CHECK:   [[TMP4:%.*]] = bitcast %struct.poly16x4x3_t* [[__RET]] to i8*
6270 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x8x3_t* [[__RET]] to i8*
6281 // CHECK:   [[TMP7:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8>, <8 x i8> }*
6283 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint8x8x3_t* %agg.result to i8*
6284 // CHECK:   [[TMP9:%.*]] = bitcast %struct.uint8x8x3_t* [[__RET]] to i8*
6293 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x4x3_t* [[__RET]] to i8*
6294 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
6305 // CHECK:   [[TMP8:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16> }*
6307 // CHECK:   [[TMP9:%.*]] = bitcast %struct.uint16x4x3_t* %agg.result to i8*
6308 // CHECK:   [[TMP10:%.*]] = bitcast %struct.uint16x4x3_t* [[__RET]] to i8*
6317 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x2x3_t* [[__RET]] to i8*
6318 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
6329 // CHECK:   [[TMP8:%.*]] = bitcast i8* [[TMP0]] to { <2 x i32>, <2 x i32>, <2 x i32> }*
6331 // CHECK:   [[TMP9:%.*]] = bitcast %struct.uint32x2x3_t* %agg.result to i8*
6332 // CHECK:   [[TMP10:%.*]] = bitcast %struct.uint32x2x3_t* [[__RET]] to i8*
6341 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint64x1x3_t* [[__RET]] to i8*
6342 // CHECK:   [[TMP1:%.*]] = bitcast i64* %a to i8*
6344 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <1 x i64>, <1 x i64>, <1 x i64> }*
6346 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint64x1x3_t* %agg.result to i8*
6347 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint64x1x3_t* [[__RET]] to i8*
6356 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x8x3_t* [[__RET]] to i8*
6367 // CHECK:   [[TMP7:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8>, <8 x i8> }*
6369 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int8x8x3_t* %agg.result to i8*
6370 // CHECK:   [[TMP9:%.*]] = bitcast %struct.int8x8x3_t* [[__RET]] to i8*
6379 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x4x3_t* [[__RET]] to i8*
6380 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
6391 // CHECK:   [[TMP8:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16> }*
6393 // CHECK:   [[TMP9:%.*]] = bitcast %struct.int16x4x3_t* %agg.result to i8*
6394 // CHECK:   [[TMP10:%.*]] = bitcast %struct.int16x4x3_t* [[__RET]] to i8*
6403 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x2x3_t* [[__RET]] to i8*
6404 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
6415 // CHECK:   [[TMP8:%.*]] = bitcast i8* [[TMP0]] to { <2 x i32>, <2 x i32>, <2 x i32> }*
6417 // CHECK:   [[TMP9:%.*]] = bitcast %struct.int32x2x3_t* %agg.result to i8*
6418 // CHECK:   [[TMP10:%.*]] = bitcast %struct.int32x2x3_t* [[__RET]] to i8*
6427 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int64x1x3_t* [[__RET]] to i8*
6428 // CHECK:   [[TMP1:%.*]] = bitcast i64* %a to i8*
6430 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <1 x i64>, <1 x i64>, <1 x i64> }*
6432 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int64x1x3_t* %agg.result to i8*
6433 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int64x1x3_t* [[__RET]] to i8*
6442 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x4x3_t* [[__RET]] to i8*
6443 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
6454 // CHECK:   [[TMP8:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16> }*
6456 // CHECK:   [[TMP9:%.*]] = bitcast %struct.float16x4x3_t* %agg.result to i8*
6457 // CHECK:   [[TMP10:%.*]] = bitcast %struct.float16x4x3_t* [[__RET]] to i8*
6466 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x2x3_t* [[__RET]] to i8*
6467 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*
6478 // CHECK:   [[TMP8:%.*]] = bitcast i8* [[TMP0]] to { <2 x float>, <2 x float>, <2 x float> }*
6480 // CHECK:   [[TMP9:%.*]] = bitcast %struct.float32x2x3_t* %agg.result to i8*
6481 // CHECK:   [[TMP10:%.*]] = bitcast %struct.float32x2x3_t* [[__RET]] to i8*
6490 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x8x3_t* [[__RET]] to i8*
6501 // CHECK:   [[TMP7:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8>, <8 x i8> }*
6503 // CHECK:   [[TMP8:%.*]] = bitcast %struct.poly8x8x3_t* %agg.result to i8*
6504 // CHECK:   [[TMP9:%.*]] = bitcast %struct.poly8x8x3_t* [[__RET]] to i8*
6513 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x4x3_t* [[__RET]] to i8*
6514 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
6525 // CHECK:   [[TMP8:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16> }*
6527 // CHECK:   [[TMP9:%.*]] = bitcast %struct.poly16x4x3_t* %agg.result to i8*
6528 // CHECK:   [[TMP10:%.*]] = bitcast %struct.poly16x4x3_t* [[__RET]] to i8*
6541 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i16>]* [[COERCE_DIVE]] to [6 x i64]*
6543 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x8x3_t* [[__S1]] to i8*
6544 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x8x3_t* [[B]] to i8*
6546 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint16x8x3_t* [[__RET]] to i8*
6547 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
6551 // CHECK:   [[TMP6:%.*]] = bitcast <8 x i16> [[TMP5]] to <16 x i8>
6555 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i16> [[TMP7]] to <16 x i8>
6559 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i16> [[TMP9]] to <16 x i8>
6560 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
6561 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
6562 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP10]] to <8 x i16>
6564 // CHECK:   [[TMP14:%.*]] = bitcast i8* [[TMP3]] to { <8 x i16>, <8 x i16>, <8 x i16> }*
6566 // CHECK:   [[TMP15:%.*]] = bitcast %struct.uint16x8x3_t* %agg.result to i8*
6567 // CHECK:   [[TMP16:%.*]] = bitcast %struct.uint16x8x3_t* [[__RET]] to i8*
6579 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i32>]* [[COERCE_DIVE]] to [6 x i64]*
6581 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x4x3_t* [[__S1]] to i8*
6582 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x4x3_t* [[B]] to i8*
6584 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint32x4x3_t* [[__RET]] to i8*
6585 // CHECK:   [[TMP4:%.*]] = bitcast i32* %a to i8*
6589 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i32> [[TMP5]] to <16 x i8>
6593 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i32> [[TMP7]] to <16 x i8>
6597 // CHECK:   [[TMP10:%.*]] = bitcast <4 x i32> [[TMP9]] to <16 x i8>
6598 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP6]] to <4 x i32>
6599 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP8]] to <4 x i32>
6600 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP10]] to <4 x i32>
6602 // CHECK:   [[TMP14:%.*]] = bitcast i8* [[TMP3]] to { <4 x i32>, <4 x i32>, <4 x i32> }*
6604 // CHECK:   [[TMP15:%.*]] = bitcast %struct.uint32x4x3_t* %agg.result to i8*
6605 // CHECK:   [[TMP16:%.*]] = bitcast %struct.uint32x4x3_t* [[__RET]] to i8*
6617 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i16>]* [[COERCE_DIVE]] to [6 x i64]*
6619 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x8x3_t* [[__S1]] to i8*
6620 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x8x3_t* [[B]] to i8*
6622 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int16x8x3_t* [[__RET]] to i8*
6623 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
6627 // CHECK:   [[TMP6:%.*]] = bitcast <8 x i16> [[TMP5]] to <16 x i8>
6631 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i16> [[TMP7]] to <16 x i8>
6635 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i16> [[TMP9]] to <16 x i8>
6636 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
6637 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
6638 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP10]] to <8 x i16>
6640 // CHECK:   [[TMP14:%.*]] = bitcast i8* [[TMP3]] to { <8 x i16>, <8 x i16>, <8 x i16> }*
6642 // CHECK:   [[TMP15:%.*]] = bitcast %struct.int16x8x3_t* %agg.result to i8*
6643 // CHECK:   [[TMP16:%.*]] = bitcast %struct.int16x8x3_t* [[__RET]] to i8*
6655 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i32>]* [[COERCE_DIVE]] to [6 x i64]*
6657 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x4x3_t* [[__S1]] to i8*
6658 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x4x3_t* [[B]] to i8*
6660 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int32x4x3_t* [[__RET]] to i8*
6661 // CHECK:   [[TMP4:%.*]] = bitcast i32* %a to i8*
6665 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i32> [[TMP5]] to <16 x i8>
6669 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i32> [[TMP7]] to <16 x i8>
6673 // CHECK:   [[TMP10:%.*]] = bitcast <4 x i32> [[TMP9]] to <16 x i8>
6674 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP6]] to <4 x i32>
6675 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP8]] to <4 x i32>
6676 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP10]] to <4 x i32>
6678 // CHECK:   [[TMP14:%.*]] = bitcast i8* [[TMP3]] to { <4 x i32>, <4 x i32>, <4 x i32> }*
6680 // CHECK:   [[TMP15:%.*]] = bitcast %struct.int32x4x3_t* %agg.result to i8*
6681 // CHECK:   [[TMP16:%.*]] = bitcast %struct.int32x4x3_t* [[__RET]] to i8*
6693 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x half>]* [[COERCE_DIVE]] to [6 x i64]*
6695 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x8x3_t* [[__S1]] to i8*
6696 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x8x3_t* [[B]] to i8*
6698 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float16x8x3_t* [[__RET]] to i8*
6699 // CHECK:   [[TMP4:%.*]] = bitcast half* %a to i8*
6703 // CHECK:   [[TMP6:%.*]] = bitcast <8 x half> [[TMP5]] to <16 x i8>
6707 // CHECK:   [[TMP8:%.*]] = bitcast <8 x half> [[TMP7]] to <16 x i8>
6711 // CHECK:   [[TMP10:%.*]] = bitcast <8 x half> [[TMP9]] to <16 x i8>
6712 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
6713 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
6714 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP10]] to <8 x i16>
6716 // CHECK:   [[TMP14:%.*]] = bitcast i8* [[TMP3]] to { <8 x i16>, <8 x i16>, <8 x i16> }*
6718 // CHECK:   [[TMP15:%.*]] = bitcast %struct.float16x8x3_t* %agg.result to i8*
6719 // CHECK:   [[TMP16:%.*]] = bitcast %struct.float16x8x3_t* [[__RET]] to i8*
6731 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x float>]* [[COERCE_DIVE]] to [6 x i64]*
6733 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x4x3_t* [[__S1]] to i8*
6734 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x4x3_t* [[B]] to i8*
6736 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float32x4x3_t* [[__RET]] to i8*
6737 // CHECK:   [[TMP4:%.*]] = bitcast float* %a to i8*
6741 // CHECK:   [[TMP6:%.*]] = bitcast <4 x float> [[TMP5]] to <16 x i8>
6745 // CHECK:   [[TMP8:%.*]] = bitcast <4 x float> [[TMP7]] to <16 x i8>
6749 // CHECK:   [[TMP10:%.*]] = bitcast <4 x float> [[TMP9]] to <16 x i8>
6750 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP6]] to <4 x float>
6751 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP8]] to <4 x float>
6752 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP10]] to <4 x float>
6754 // CHECK:   [[TMP14:%.*]] = bitcast i8* [[TMP3]] to { <4 x float>, <4 x float>, <4 x float> }*
6756 // CHECK:   [[TMP15:%.*]] = bitcast %struct.float32x4x3_t* %agg.result to i8*
6757 // CHECK:   [[TMP16:%.*]] = bitcast %struct.float32x4x3_t* [[__RET]] to i8*
6769 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i16>]* [[COERCE_DIVE]] to [6 x i64]*
6771 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x8x3_t* [[__S1]] to i8*
6772 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x8x3_t* [[B]] to i8*
6774 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly16x8x3_t* [[__RET]] to i8*
6775 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
6779 // CHECK:   [[TMP6:%.*]] = bitcast <8 x i16> [[TMP5]] to <16 x i8>
6783 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i16> [[TMP7]] to <16 x i8>
6787 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i16> [[TMP9]] to <16 x i8>
6788 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
6789 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
6790 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP10]] to <8 x i16>
6792 // CHECK:   [[TMP14:%.*]] = bitcast i8* [[TMP3]] to { <8 x i16>, <8 x i16>, <8 x i16> }*
6794 // CHECK:   [[TMP15:%.*]] = bitcast %struct.poly16x8x3_t* %agg.result to i8*
6795 // CHECK:   [[TMP16:%.*]] = bitcast %struct.poly16x8x3_t* [[__RET]] to i8*
6807 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
6809 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint8x8x3_t* [[__S1]] to i8*
6810 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x8x3_t* [[B]] to i8*
6812 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x8x3_t* [[__RET]] to i8*
6823 // CHECK:   [[TMP7:%.*]] = bitcast i8* [[TMP3]] to { <8 x i8>, <8 x i8>, <8 x i8> }*
6825 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint8x8x3_t* %agg.result to i8*
6826 // CHECK:   [[TMP9:%.*]] = bitcast %struct.uint8x8x3_t* [[__RET]] to i8*
6838 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i16>]* [[COERCE_DIVE]] to [3 x i64]*
6840 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x4x3_t* [[__S1]] to i8*
6841 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x4x3_t* [[B]] to i8*
6843 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint16x4x3_t* [[__RET]] to i8*
6844 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
6848 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i16> [[TMP5]] to <8 x i8>
6852 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i16> [[TMP7]] to <8 x i8>
6856 // CHECK:   [[TMP10:%.*]] = bitcast <4 x i16> [[TMP9]] to <8 x i8>
6857 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
6858 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
6859 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP10]] to <4 x i16>
6861 // CHECK:   [[TMP14:%.*]] = bitcast
6863 // CHECK:   [[TMP15:%.*]] = bitcast %struct.uint16x4x3_t* %agg.result to i8*
6864 // CHECK:   [[TMP16:%.*]] = bitcast %struct.uint16x4x3_t* [[__RET]] to i8*
6876 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <2 x i32>]* [[COERCE_DIVE]] to [3 x i64]*
6878 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x2x3_t* [[__S1]] to i8*
6879 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x2x3_t* [[B]] to i8*
6881 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint32x2x3_t* [[__RET]] to i8*
6882 // CHECK:   [[TMP4:%.*]] = bitcast i32* %a to i8*
6886 // CHECK:   [[TMP6:%.*]] = bitcast <2 x i32> [[TMP5]] to <8 x i8>
6890 // CHECK:   [[TMP8:%.*]] = bitcast <2 x i32> [[TMP7]] to <8 x i8>
6894 // CHECK:   [[TMP10:%.*]] = bitcast <2 x i32> [[TMP9]] to <8 x i8>
6895 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP6]] to <2 x i32>
6896 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP8]] to <2 x i32>
6897 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP10]] to <2 x i32>
6899 // CHECK:   [[TMP14:%.*]] = bitcast i8* [[TMP3]] to { <2 x i32>, <2 x i32>, <2 x i32> }*
6901 // CHECK:   [[TMP15:%.*]] = bitcast %struct.uint32x2x3_t* %agg.result to i8*
6902 // CHECK:   [[TMP16:%.*]] = bitcast %struct.uint32x2x3_t* [[__RET]] to i8*
6914 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
6916 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int8x8x3_t* [[__S1]] to i8*
6917 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x8x3_t* [[B]] to i8*
6919 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x8x3_t* [[__RET]] to i8*
6930 // CHECK:   [[TMP7:%.*]] = bitcast i8* [[TMP3]] to { <8 x i8>, <8 x i8>, <8 x i8> }*
6932 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int8x8x3_t* %agg.result to i8*
6933 // CHECK:   [[TMP9:%.*]] = bitcast %struct.int8x8x3_t* [[__RET]] to i8*
6945 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i16>]* [[COERCE_DIVE]] to [3 x i64]*
6947 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x4x3_t* [[__S1]] to i8*
6948 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x4x3_t* [[B]] to i8*
6950 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int16x4x3_t* [[__RET]] to i8*
6951 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
6955 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i16> [[TMP5]] to <8 x i8>
6959 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i16> [[TMP7]] to <8 x i8>
6963 // CHECK:   [[TMP10:%.*]] = bitcast <4 x i16> [[TMP9]] to <8 x i8>
6964 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
6965 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
6966 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP10]] to <4 x i16>
6968 // CHECK:   [[TMP14:%.*]] = bitcast i8* [[TMP3]] to { <4 x i16>, <4 x i16>, <4 x i16> }*
6970 // CHECK:   [[TMP15:%.*]] = bitcast %struct.int16x4x3_t* %agg.result to i8*
6971 // CHECK:   [[TMP16:%.*]] = bitcast %struct.int16x4x3_t* [[__RET]] to i8*
6983 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <2 x i32>]* [[COERCE_DIVE]] to [3 x i64]*
6985 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x2x3_t* [[__S1]] to i8*
6986 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x2x3_t* [[B]] to i8*
6988 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int32x2x3_t* [[__RET]] to i8*
6989 // CHECK:   [[TMP4:%.*]] = bitcast i32* %a to i8*
6993 // CHECK:   [[TMP6:%.*]] = bitcast <2 x i32> [[TMP5]] to <8 x i8>
6997 // CHECK:   [[TMP8:%.*]] = bitcast <2 x i32> [[TMP7]] to <8 x i8>
7001 // CHECK:   [[TMP10:%.*]] = bitcast <2 x i32> [[TMP9]] to <8 x i8>
7002 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP6]] to <2 x i32>
7003 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP8]] to <2 x i32>
7004 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP10]] to <2 x i32>
7006 // CHECK:   [[TMP14:%.*]] = bitcast i8* [[TMP3]] to { <2 x i32>, <2 x i32>, <2 x i32> }*
7008 // CHECK:   [[TMP15:%.*]] = bitcast %struct.int32x2x3_t* %agg.result to i8*
7009 // CHECK:   [[TMP16:%.*]] = bitcast %struct.int32x2x3_t* [[__RET]] to i8*
7021 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x half>]* [[COERCE_DIVE]] to [3 x i64]*
7023 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x4x3_t* [[__S1]] to i8*
7024 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x4x3_t* [[B]] to i8*
7026 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float16x4x3_t* [[__RET]] to i8*
7027 // CHECK:   [[TMP4:%.*]] = bitcast half* %a to i8*
7031 // CHECK:   [[TMP6:%.*]] = bitcast <4 x half> [[TMP5]] to <8 x i8>
7035 // CHECK:   [[TMP8:%.*]] = bitcast <4 x half> [[TMP7]] to <8 x i8>
7039 // CHECK:   [[TMP10:%.*]] = bitcast <4 x half> [[TMP9]] to <8 x i8>
7040 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
7041 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
7042 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP10]] to <4 x i16>
7044 // CHECK:   [[TMP14:%.*]] = bitcast i8* [[TMP3]] to { <4 x i16>, <4 x i16>, <4 x i16> }*
7046 // CHECK:   [[TMP15:%.*]] = bitcast %struct.float16x4x3_t* %agg.result to i8*
7047 // CHECK:   [[TMP16:%.*]] = bitcast %struct.float16x4x3_t* [[__RET]] to i8*
7059 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <2 x float>]* [[COERCE_DIVE]] to [3 x i64]*
7061 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x2x3_t* [[__S1]] to i8*
7062 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x2x3_t* [[B]] to i8*
7064 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float32x2x3_t* [[__RET]] to i8*
7065 // CHECK:   [[TMP4:%.*]] = bitcast float* %a to i8*
7069 // CHECK:   [[TMP6:%.*]] = bitcast <2 x float> [[TMP5]] to <8 x i8>
7073 // CHECK:   [[TMP8:%.*]] = bitcast <2 x float> [[TMP7]] to <8 x i8>
7077 // CHECK:   [[TMP10:%.*]] = bitcast <2 x float> [[TMP9]] to <8 x i8>
7078 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP6]] to <2 x float>
7079 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP8]] to <2 x float>
7080 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP10]] to <2 x float>
7082 // CHECK:   [[TMP14:%.*]] = bitcast i8* [[TMP3]] to { <2 x float>, <2 x float>, <2 x float> }*
7084 // CHECK:   [[TMP15:%.*]] = bitcast %struct.float32x2x3_t* %agg.result to i8*
7085 // CHECK:   [[TMP16:%.*]] = bitcast %struct.float32x2x3_t* [[__RET]] to i8*
7097 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
7099 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly8x8x3_t* [[__S1]] to i8*
7100 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x8x3_t* [[B]] to i8*
7102 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x8x3_t* [[__RET]] to i8*
7113 // CHECK:   [[TMP7:%.*]] = bitcast i8* [[TMP3]] to { <8 x i8>, <8 x i8>, <8 x i8> }*
7115 // CHECK:   [[TMP8:%.*]] = bitcast %struct.poly8x8x3_t* %agg.result to i8*
7116 // CHECK:   [[TMP9:%.*]] = bitcast %struct.poly8x8x3_t* [[__RET]] to i8*
7128 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i16>]* [[COERCE_DIVE]] to [3 x i64]*
7130 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x4x3_t* [[__S1]] to i8*
7131 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x4x3_t* [[B]] to i8*
7133 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly16x4x3_t* [[__RET]] to i8*
7134 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
7138 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i16> [[TMP5]] to <8 x i8>
7142 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i16> [[TMP7]] to <8 x i8>
7146 // CHECK:   [[TMP10:%.*]] = bitcast <4 x i16> [[TMP9]] to <8 x i8>
7147 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
7148 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
7149 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP10]] to <4 x i16>
7151 // CHECK:   [[TMP14:%.*]] = bitcast i8* [[TMP3]] to { <4 x i16>, <4 x i16>, <4 x i16> }*
7153 // CHECK:   [[TMP15:%.*]] = bitcast %struct.poly16x4x3_t* %agg.result to i8*
7154 // CHECK:   [[TMP16:%.*]] = bitcast %struct.poly16x4x3_t* [[__RET]] to i8*
7164 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x16x4_t* [[__RET]] to i8*
7166 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <16 x i8>, <16 x i8>, <16 x i8>, <16 x i8> }*
7168 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x16x4_t* %agg.result to i8*
7169 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x16x4_t* [[__RET]] to i8*
7178 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x8x4_t* [[__RET]] to i8*
7179 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
7181 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <8 x i16>, <8 x i16>, <8 x i16>, <8 x i16> }*
7183 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint16x8x4_t* %agg.result to i8*
7184 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint16x8x4_t* [[__RET]] to i8*
7193 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x4x4_t* [[__RET]] to i8*
7194 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
7196 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i32>, <4 x i32>, <4 x i32>, <4 x i32> }*
7198 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint32x4x4_t* %agg.result to i8*
7199 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint32x4x4_t* [[__RET]] to i8*
7208 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x16x4_t* [[__RET]] to i8*
7210 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <16 x i8>, <16 x i8>, <16 x i8>, <16 x i8> }*
7212 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x16x4_t* %agg.result to i8*
7213 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x16x4_t* [[__RET]] to i8*
7222 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x8x4_t* [[__RET]] to i8*
7223 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
7225 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <8 x i16>, <8 x i16>, <8 x i16>, <8 x i16> }*
7227 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int16x8x4_t* %agg.result to i8*
7228 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int16x8x4_t* [[__RET]] to i8*
7237 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x4x4_t* [[__RET]] to i8*
7238 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
7240 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i32>, <4 x i32>, <4 x i32>, <4 x i32> }*
7242 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int32x4x4_t* %agg.result to i8*
7243 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int32x4x4_t* [[__RET]] to i8*
7252 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x8x4_t* [[__RET]] to i8*
7253 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
7255 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <8 x i16>, <8 x i16>, <8 x i16>, <8 x i16> }*
7257 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float16x8x4_t* %agg.result to i8*
7258 // CHECK:   [[TMP4:%.*]] = bitcast %struct.float16x8x4_t* [[__RET]] to i8*
7267 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x4x4_t* [[__RET]] to i8*
7268 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*
7270 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x float>, <4 x float>, <4 x float>, <4 x float> }*
7272 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float32x4x4_t* %agg.result to i8*
7273 // CHECK:   [[TMP4:%.*]] = bitcast %struct.float32x4x4_t* [[__RET]] to i8*
7282 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x16x4_t* [[__RET]] to i8*
7284 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <16 x i8>, <16 x i8>, <16 x i8>, <16 x i8> }*
7286 bitcast %struct.poly8x16x4_t* %agg.result to i8*
7287 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x16x4_t* [[__RET]] to i8*
7296 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x8x4_t* [[__RET]] to i8*
7297 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
7299 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <8 x i16>, <8 x i16>, <8 x i16>, <8 x i16> }*
7301 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly16x8x4_t* %agg.result to i8*
7302 // CHECK:   [[TMP4:%.*]] = bitcast %struct.poly16x8x4_t* [[__RET]] to i8*
7311 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x8x4_t* [[__RET]] to i8*
7313 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> }*
7315 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x8x4_t* %agg.result to i8*
7316 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x8x4_t* [[__RET]] to i8*
7325 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x4x4_t* [[__RET]] to i8*
7326 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
7328 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16> }*
7330 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint16x4x4_t* %agg.result to i8*
7331 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint16x4x4_t* [[__RET]] to i8*
7340 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x2x4_t* [[__RET]] to i8*
7341 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
7343 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <2 x i32>, <2 x i32>, <2 x i32>, <2 x i32> }*
7345 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint32x2x4_t* %agg.result to i8*
7346 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint32x2x4_t* [[__RET]] to i8*
7355 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint64x1x4_t* [[__RET]] to i8*
7356 // CHECK:   [[TMP1:%.*]] = bitcast i64* %a to i8*
7358 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <1 x i64>, <1 x i64>, <1 x i64>, <1 x i64> }*
7360 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint64x1x4_t* %agg.result to i8*
7361 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint64x1x4_t* [[__RET]] to i8*
7370 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x8x4_t* [[__RET]] to i8*
7372 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> }*
7374 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x8x4_t* %agg.result to i8*
7375 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x8x4_t* [[__RET]] to i8*
7384 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x4x4_t* [[__RET]] to i8*
7385 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
7387 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16> }*
7389 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int16x4x4_t* %agg.result to i8*
7390 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int16x4x4_t* [[__RET]] to i8*
7399 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x2x4_t* [[__RET]] to i8*
7400 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
7402 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <2 x i32>, <2 x i32>, <2 x i32>, <2 x i32> }*
7404 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int32x2x4_t* %agg.result to i8*
7405 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int32x2x4_t* [[__RET]] to i8*
7414 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int64x1x4_t* [[__RET]] to i8*
7415 // CHECK:   [[TMP1:%.*]] = bitcast i64* %a to i8*
7417 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <1 x i64>, <1 x i64>, <1 x i64>, <1 x i64> }*
7419 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int64x1x4_t* %agg.result to i8*
7420 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int64x1x4_t* [[__RET]] to i8*
7429 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x4x4_t* [[__RET]] to i8*
7430 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
7432 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16> }*
7434 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float16x4x4_t* %agg.result to i8*
7435 // CHECK:   [[TMP4:%.*]] = bitcast %struct.float16x4x4_t* [[__RET]] to i8*
7444 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x2x4_t* [[__RET]] to i8*
7445 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*
7447 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <2 x float>, <2 x float>, <2 x float>, <2 x float> }*
7449 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float32x2x4_t* %agg.result to i8*
7450 // CHECK:   [[TMP4:%.*]] = bitcast %struct.float32x2x4_t* [[__RET]] to i8*
7459 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x8x4_t* [[__RET]] to i8*
7461 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> }*
7463 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x8x4_t* %agg.result to i8*
7464 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x8x4_t* [[__RET]] to i8*
7473 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x4x4_t* [[__RET]] to i8*
7474 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
7476 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16> }*
7478 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly16x4x4_t* %agg.result to i8*
7479 // CHECK:   [[TMP4:%.*]] = bitcast %struct.poly16x4x4_t* [[__RET]] to i8*
7489 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x8x4_t* [[__RET]] to i8*
7503 // CHECK:   [[TMP9:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> }*
7505 // CHECK:   [[TMP10:%.*]] = bitcast %struct.uint8x8x4_t* %agg.result to i8*
7506 // CHECK:   [[TMP11:%.*]] = bitcast %struct.uint8x8x4_t* [[__RET]] to i8*
7515 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x4x4_t* [[__RET]] to i8*
7516 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
7530 // CHECK:   [[TMP10:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16> }*
7532 // CHECK:   [[TMP11:%.*]] = bitcast %struct.uint16x4x4_t* %agg.result to i8*
7533 // CHECK:   [[TMP12:%.*]] = bitcast %struct.uint16x4x4_t* [[__RET]] to i8*
7542 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x2x4_t* [[__RET]] to i8*
7543 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
7557 // CHECK:   [[TMP10:%.*]] = bitcast i8* [[TMP0]] to { <2 x i32>, <2 x i32>, <2 x i32>, <2 x i32> }*
7559 // CHECK:   [[TMP11:%.*]] = bitcast %struct.uint32x2x4_t* %agg.result to i8*
7560 // CHECK:   [[TMP12:%.*]] = bitcast %struct.uint32x2x4_t* [[__RET]] to i8*
7569 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint64x1x4_t* [[__RET]] to i8*
7570 // CHECK:   [[TMP1:%.*]] = bitcast i64* %a to i8*
7572 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <1 x i64>, <1 x i64>, <1 x i64>, <1 x i64> }*
7574 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint64x1x4_t* %agg.result to i8*
7575 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint64x1x4_t* [[__RET]] to i8*
7584 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x8x4_t* [[__RET]] to i8*
7598 // CHECK:   [[TMP9:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> }*
7600 // CHECK:   [[TMP10:%.*]] = bitcast %struct.int8x8x4_t* %agg.result to i8*
7601 // CHECK:   [[TMP11:%.*]] = bitcast %struct.int8x8x4_t* [[__RET]] to i8*
7610 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x4x4_t* [[__RET]] to i8*
7611 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
7625 // CHECK:   [[TMP10:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16> }*
7627 // CHECK:   [[TMP11:%.*]] = bitcast %struct.int16x4x4_t* %agg.result to i8*
7628 // CHECK:   [[TMP12:%.*]] = bitcast %struct.int16x4x4_t* [[__RET]] to i8*
7637 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x2x4_t* [[__RET]] to i8*
7638 // CHECK:   [[TMP1:%.*]] = bitcast i32* %a to i8*
7652 // CHECK:   [[TMP10:%.*]] = bitcast i8* [[TMP0]] to { <2 x i32>, <2 x i32>, <2 x i32>, <2 x i32> }*
7654 // CHECK:   [[TMP11:%.*]] = bitcast %struct.int32x2x4_t* %agg.result to i8*
7655 // CHECK:   [[TMP12:%.*]] = bitcast %struct.int32x2x4_t* [[__RET]] to i8*
7664 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int64x1x4_t* [[__RET]] to i8*
7665 // CHECK:   [[TMP1:%.*]] = bitcast i64* %a to i8*
7667 // CHECK:   [[TMP2:%.*]] = bitcast i8* [[TMP0]] to { <1 x i64>, <1 x i64>, <1 x i64>, <1 x i64> }*
7669 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int64x1x4_t* %agg.result to i8*
7670 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int64x1x4_t* [[__RET]] to i8*
7679 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float16x4x4_t* [[__RET]] to i8*
7680 // CHECK:   [[TMP1:%.*]] = bitcast half* %a to i8*
7694 // CHECK:   [[TMP10:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16> }*
7696 // CHECK:   [[TMP11:%.*]] = bitcast %struct.float16x4x4_t* %agg.result to i8*
7697 // CHECK:   [[TMP12:%.*]] = bitcast %struct.float16x4x4_t* [[__RET]] to i8*
7706 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x2x4_t* [[__RET]] to i8*
7707 // CHECK:   [[TMP1:%.*]] = bitcast float* %a to i8*
7721 // CHECK:   [[TMP10:%.*]] = bitcast i8* [[TMP0]] to { <2 x float>, <2 x float>, <2 x float>, <2 x float> }*
7723 // CHECK:   [[TMP11:%.*]] = bitcast %struct.float32x2x4_t* %agg.result to i8*
7724 // CHECK:   [[TMP12:%.*]] = bitcast %struct.float32x2x4_t* [[__RET]] to i8*
7733 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x8x4_t* [[__RET]] to i8*
7747 // CHECK:   [[TMP9:%.*]] = bitcast i8* [[TMP0]] to { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> }*
7749 // CHECK:   [[TMP10:%.*]] = bitcast %struct.poly8x8x4_t* %agg.result to i8*
7750 // CHECK:   [[TMP11:%.*]] = bitcast %struct.poly8x8x4_t* [[__RET]] to i8*
7759 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x4x4_t* [[__RET]] to i8*
7760 // CHECK:   [[TMP1:%.*]] = bitcast i16* %a to i8*
7774 // CHECK:   [[TMP10:%.*]] = bitcast i8* [[TMP0]] to { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16> }*
7776 // CHECK:   [[TMP11:%.*]] = bitcast %struct.poly16x4x4_t* %agg.result to i8*
7777 // CHECK:   [[TMP12:%.*]] = bitcast %struct.poly16x4x4_t* [[__RET]] to i8*
7790 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i16>]* [[COERCE_DIVE]] to [8 x i64]*
7792 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x8x4_t* [[__S1]] to i8*
7793 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x8x4_t* [[B]] to i8*
7795 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint16x8x4_t* [[__RET]] to i8*
7796 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
7800 // CHECK:   [[TMP6:%.*]] = bitcast <8 x i16> [[TMP5]] to <16 x i8>
7804 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i16> [[TMP7]] to <16 x i8>
7808 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i16> [[TMP9]] to <16 x i8>
7812 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i16> [[TMP11]] to <16 x i8>
7813 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
7814 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
7815 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP10]] to <8 x i16>
7816 // CHECK:   [[TMP16:%.*]] = bitcast <16 x i8> [[TMP12]] to <8 x i16>
7818 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <8 x i16>, <8 x i16>, <8 x i16>, <8 x i16> }*
7820 // CHECK:   [[TMP18:%.*]] = bitcast %struct.uint16x8x4_t* %agg.result to i8*
7821 // CHECK:   [[TMP19:%.*]] = bitcast %struct.uint16x8x4_t* [[__RET]] to i8*
7833 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i32>]* [[COERCE_DIVE]] to [8 x i64]*
7835 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x4x4_t* [[__S1]] to i8*
7836 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x4x4_t* [[B]] to i8*
7838 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint32x4x4_t* [[__RET]] to i8*
7839 // CHECK:   [[TMP4:%.*]] = bitcast i32* %a to i8*
7843 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i32> [[TMP5]] to <16 x i8>
7847 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i32> [[TMP7]] to <16 x i8>
7851 // CHECK:   [[TMP10:%.*]] = bitcast <4 x i32> [[TMP9]] to <16 x i8>
7855 // CHECK:   [[TMP12:%.*]] = bitcast <4 x i32> [[TMP11]] to <16 x i8>
7856 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP6]] to <4 x i32>
7857 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP8]] to <4 x i32>
7858 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP10]] to <4 x i32>
7859 // CHECK:   [[TMP16:%.*]] = bitcast <16 x i8> [[TMP12]] to <4 x i32>
7861 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <4 x i32>, <4 x i32>, <4 x i32>, <4 x i32> }*
7863 // CHECK:   [[TMP18:%.*]] = bitcast %struct.uint32x4x4_t* %agg.result to i8*
7864 // CHECK:   [[TMP19:%.*]] = bitcast %struct.uint32x4x4_t* [[__RET]] to i8*
7876 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i16>]* [[COERCE_DIVE]] to [8 x i64]*
7878 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x8x4_t* [[__S1]] to i8*
7879 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x8x4_t* [[B]] to i8*
7881 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int16x8x4_t* [[__RET]] to i8*
7882 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
7886 // CHECK:   [[TMP6:%.*]] = bitcast <8 x i16> [[TMP5]] to <16 x i8>
7890 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i16> [[TMP7]] to <16 x i8>
7894 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i16> [[TMP9]] to <16 x i8>
7898 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i16> [[TMP11]] to <16 x i8>
7899 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
7900 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
7901 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP10]] to <8 x i16>
7902 // CHECK:   [[TMP16:%.*]] = bitcast <16 x i8> [[TMP12]] to <8 x i16>
7904 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <8 x i16>, <8 x i16>, <8 x i16>, <8 x i16> }*
7906 // CHECK:   [[TMP18:%.*]] = bitcast %struct.int16x8x4_t* %agg.result to i8*
7907 // CHECK:   [[TMP19:%.*]] = bitcast %struct.int16x8x4_t* [[__RET]] to i8*
7919 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i32>]* [[COERCE_DIVE]] to [8 x i64]*
7921 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x4x4_t* [[__S1]] to i8*
7922 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x4x4_t* [[B]] to i8*
7924 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int32x4x4_t* [[__RET]] to i8*
7925 // CHECK:   [[TMP4:%.*]] = bitcast i32* %a to i8*
7929 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i32> [[TMP5]] to <16 x i8>
7933 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i32> [[TMP7]] to <16 x i8>
7937 // CHECK:   [[TMP10:%.*]] = bitcast <4 x i32> [[TMP9]] to <16 x i8>
7941 // CHECK:   [[TMP12:%.*]] = bitcast <4 x i32> [[TMP11]] to <16 x i8>
7942 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP6]] to <4 x i32>
7943 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP8]] to <4 x i32>
7944 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP10]] to <4 x i32>
7945 // CHECK:   [[TMP16:%.*]] = bitcast <16 x i8> [[TMP12]] to <4 x i32>
7947 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <4 x i32>, <4 x i32>, <4 x i32>, <4 x i32> }*
7949 // CHECK:   [[TMP18:%.*]] = bitcast %struct.int32x4x4_t* %agg.result to i8*
7950 // CHECK:   [[TMP19:%.*]] = bitcast %struct.int32x4x4_t* [[__RET]] to i8*
7962 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x half>]* [[COERCE_DIVE]] to [8 x i64]*
7964 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x8x4_t* [[__S1]] to i8*
7965 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x8x4_t* [[B]] to i8*
7967 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float16x8x4_t* [[__RET]] to i8*
7968 // CHECK:   [[TMP4:%.*]] = bitcast half* %a to i8*
7972 // CHECK:   [[TMP6:%.*]] = bitcast <8 x half> [[TMP5]] to <16 x i8>
7976 // CHECK:   [[TMP8:%.*]] = bitcast <8 x half> [[TMP7]] to <16 x i8>
7980 // CHECK:   [[TMP10:%.*]] = bitcast <8 x half> [[TMP9]] to <16 x i8>
7984 // CHECK:   [[TMP12:%.*]] = bitcast <8 x half> [[TMP11]] to <16 x i8>
7985 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
7986 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
7987 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP10]] to <8 x i16>
7988 // CHECK:   [[TMP16:%.*]] = bitcast <16 x i8> [[TMP12]] to <8 x i16>
7990 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <8 x i16>, <8 x i16>, <8 x i16>, <8 x i16> }*
7992 // CHECK:   [[TMP18:%.*]] = bitcast %struct.float16x8x4_t* %agg.result to i8*
7993 // CHECK:   [[TMP19:%.*]] = bitcast %struct.float16x8x4_t* [[__RET]] to i8*
8005 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x float>]* [[COERCE_DIVE]] to [8 x i64]*
8007 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x4x4_t* [[__S1]] to i8*
8008 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x4x4_t* [[B]] to i8*
8010 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float32x4x4_t* [[__RET]] to i8*
8011 // CHECK:   [[TMP4:%.*]] = bitcast float* %a to i8*
8015 // CHECK:   [[TMP6:%.*]] = bitcast <4 x float> [[TMP5]] to <16 x i8>
8019 // CHECK:   [[TMP8:%.*]] = bitcast <4 x float> [[TMP7]] to <16 x i8>
8023 // CHECK:   [[TMP10:%.*]] = bitcast <4 x float> [[TMP9]] to <16 x i8>
8027 // CHECK:   [[TMP12:%.*]] = bitcast <4 x float> [[TMP11]] to <16 x i8>
8028 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP6]] to <4 x float>
8029 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP8]] to <4 x float>
8030 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP10]] to <4 x float>
8031 // CHECK:   [[TMP16:%.*]] = bitcast <16 x i8> [[TMP12]] to <4 x float>
8033 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <4 x float>, <4 x float>, <4 x float>, <4 x float> }*
8035 // CHECK:   [[TMP18:%.*]] = bitcast %struct.float32x4x4_t* %agg.result to i8*
8036 // CHECK:   [[TMP19:%.*]] = bitcast %struct.float32x4x4_t* [[__RET]] to i8*
8048 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i16>]* [[COERCE_DIVE]] to [8 x i64]*
8050 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x8x4_t* [[__S1]] to i8*
8051 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x8x4_t* [[B]] to i8*
8053 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly16x8x4_t* [[__RET]] to i8*
8054 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
8058 // CHECK:   [[TMP6:%.*]] = bitcast <8 x i16> [[TMP5]] to <16 x i8>
8062 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i16> [[TMP7]] to <16 x i8>
8066 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i16> [[TMP9]] to <16 x i8>
8070 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i16> [[TMP11]] to <16 x i8>
8071 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP6]] to <8 x i16>
8072 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP8]] to <8 x i16>
8073 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP10]] to <8 x i16>
8074 // CHECK:   [[TMP16:%.*]] = bitcast <16 x i8> [[TMP12]] to <8 x i16>
8076 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <8 x i16>, <8 x i16>, <8 x i16>, <8 x i16> }*
8078 // CHECK:   [[TMP18:%.*]] = bitcast %struct.poly16x8x4_t* %agg.result to i8*
8079 // CHECK:   [[TMP19:%.*]] = bitcast %struct.poly16x8x4_t* [[__RET]] to i8*
8091 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
8093 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint8x8x4_t* [[__S1]] to i8*
8094 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x8x4_t* [[B]] to i8*
8096 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x8x4_t* [[__RET]] to i8*
8110 // CHECK:   [[TMP8:%.*]] = bitcast i8* [[TMP3]] to { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> }*
8112 // CHECK:   [[TMP9:%.*]] = bitcast %struct.uint8x8x4_t* %agg.result to i8*
8113 // CHECK:   [[TMP10:%.*]] = bitcast %struct.uint8x8x4_t* [[__RET]] to i8*
8125 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
8127 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x4x4_t* [[__S1]] to i8*
8128 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x4x4_t* [[B]] to i8*
8130 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint16x4x4_t* [[__RET]] to i8*
8131 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
8135 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i16> [[TMP5]] to <8 x i8>
8139 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i16> [[TMP7]] to <8 x i8>
8143 // CHECK:   [[TMP10:%.*]] = bitcast <4 x i16> [[TMP9]] to <8 x i8>
8147 // CHECK:   [[TMP12:%.*]] = bitcast <4 x i16> [[TMP11]] to <8 x i8>
8148 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
8149 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
8150 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP10]] to <4 x i16>
8151 // CHECK:   [[TMP16:%.*]] = bitcast <8 x i8> [[TMP12]] to <4 x i16>
8153 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16> }*
8155 // CHECK:   [[TMP18:%.*]] = bitcast %struct.uint16x4x4_t* %agg.result to i8*
8156 // CHECK:   [[TMP19:%.*]] = bitcast %struct.uint16x4x4_t* [[__RET]] to i8*
8168 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <2 x i32>]* [[COERCE_DIVE]] to [4 x i64]*
8170 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x2x4_t* [[__S1]] to i8*
8171 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x2x4_t* [[B]] to i8*
8173 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint32x2x4_t* [[__RET]] to i8*
8174 // CHECK:   [[TMP4:%.*]] = bitcast i32* %a to i8*
8178 // CHECK:   [[TMP6:%.*]] = bitcast <2 x i32> [[TMP5]] to <8 x i8>
8182 // CHECK:   [[TMP8:%.*]] = bitcast <2 x i32> [[TMP7]] to <8 x i8>
8186 // CHECK:   [[TMP10:%.*]] = bitcast <2 x i32> [[TMP9]] to <8 x i8>
8190 // CHECK:   [[TMP12:%.*]] = bitcast <2 x i32> [[TMP11]] to <8 x i8>
8191 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP6]] to <2 x i32>
8192 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP8]] to <2 x i32>
8193 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP10]] to <2 x i32>
8194 // CHECK:   [[TMP16:%.*]] = bitcast <8 x i8> [[TMP12]] to <2 x i32>
8196 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <2 x i32>, <2 x i32>, <2 x i32>, <2 x i32> }*
8198 // CHECK:   [[TMP18:%.*]] = bitcast %struct.uint32x2x4_t* %agg.result to i8*
8199 // CHECK:   [[TMP19:%.*]] = bitcast %struct.uint32x2x4_t* [[__RET]] to i8*
8211 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
8213 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int8x8x4_t* [[__S1]] to i8*
8214 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x8x4_t* [[B]] to i8*
8216 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x8x4_t* [[__RET]] to i8*
8230 // CHECK:   [[TMP8:%.*]] = bitcast i8* [[TMP3]] to { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> }*
8232 // CHECK:   [[TMP9:%.*]] = bitcast %struct.int8x8x4_t* %agg.result to i8*
8233 // CHECK:   [[TMP10:%.*]] = bitcast %struct.int8x8x4_t* [[__RET]] to i8*
8245 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
8247 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x4x4_t* [[__S1]] to i8*
8248 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x4x4_t* [[B]] to i8*
8250 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int16x4x4_t* [[__RET]] to i8*
8251 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
8255 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i16> [[TMP5]] to <8 x i8>
8259 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i16> [[TMP7]] to <8 x i8>
8263 // CHECK:   [[TMP10:%.*]] = bitcast <4 x i16> [[TMP9]] to <8 x i8>
8267 // CHECK:   [[TMP12:%.*]] = bitcast <4 x i16> [[TMP11]] to <8 x i8>
8268 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
8269 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
8270 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP10]] to <4 x i16>
8271 // CHECK:   [[TMP16:%.*]] = bitcast <8 x i8> [[TMP12]] to <4 x i16>
8273 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16> }*
8275 // CHECK:   [[TMP18:%.*]] = bitcast %struct.int16x4x4_t* %agg.result to i8*
8276 // CHECK:   [[TMP19:%.*]] = bitcast %struct.int16x4x4_t* [[__RET]] to i8*
8288 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <2 x i32>]* [[COERCE_DIVE]] to [4 x i64]*
8290 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x2x4_t* [[__S1]] to i8*
8291 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x2x4_t* [[B]] to i8*
8293 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int32x2x4_t* [[__RET]] to i8*
8294 // CHECK:   [[TMP4:%.*]] = bitcast i32* %a to i8*
8298 // CHECK:   [[TMP6:%.*]] = bitcast <2 x i32> [[TMP5]] to <8 x i8>
8302 // CHECK:   [[TMP8:%.*]] = bitcast <2 x i32> [[TMP7]] to <8 x i8>
8306 // CHECK:   [[TMP10:%.*]] = bitcast <2 x i32> [[TMP9]] to <8 x i8>
8310 // CHECK:   [[TMP12:%.*]] = bitcast <2 x i32> [[TMP11]] to <8 x i8>
8311 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP6]] to <2 x i32>
8312 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP8]] to <2 x i32>
8313 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP10]] to <2 x i32>
8314 // CHECK:   [[TMP16:%.*]] = bitcast <8 x i8> [[TMP12]] to <2 x i32>
8316 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <2 x i32>, <2 x i32>, <2 x i32>, <2 x i32> }*
8318 // CHECK:   [[TMP18:%.*]] = bitcast %struct.int32x2x4_t* %agg.result to i8*
8319 // CHECK:   [[TMP19:%.*]] = bitcast %struct.int32x2x4_t* [[__RET]] to i8*
8331 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x half>]* [[COERCE_DIVE]] to [4 x i64]*
8333 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x4x4_t* [[__S1]] to i8*
8334 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x4x4_t* [[B]] to i8*
8336 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float16x4x4_t* [[__RET]] to i8*
8337 // CHECK:   [[TMP4:%.*]] = bitcast half* %a to i8*
8341 // CHECK:   [[TMP6:%.*]] = bitcast <4 x half> [[TMP5]] to <8 x i8>
8345 // CHECK:   [[TMP8:%.*]] = bitcast <4 x half> [[TMP7]] to <8 x i8>
8349 // CHECK:   [[TMP10:%.*]] = bitcast <4 x half> [[TMP9]] to <8 x i8>
8353 // CHECK:   [[TMP12:%.*]] = bitcast <4 x half> [[TMP11]] to <8 x i8>
8354 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
8355 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
8356 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP10]] to <4 x i16>
8357 // CHECK:   [[TMP16:%.*]] = bitcast <8 x i8> [[TMP12]] to <4 x i16>
8359 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16> }*
8361 // CHECK:   [[TMP18:%.*]] = bitcast %struct.float16x4x4_t* %agg.result to i8*
8362 // CHECK:   [[TMP19:%.*]] = bitcast %struct.float16x4x4_t* [[__RET]] to i8*
8374 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <2 x float>]* [[COERCE_DIVE]] to [4 x i64]*
8376 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x2x4_t* [[__S1]] to i8*
8377 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x2x4_t* [[B]] to i8*
8379 // CHECK:   [[TMP3:%.*]] = bitcast %struct.float32x2x4_t* [[__RET]] to i8*
8380 // CHECK:   [[TMP4:%.*]] = bitcast float* %a to i8*
8384 // CHECK:   [[TMP6:%.*]] = bitcast <2 x float> [[TMP5]] to <8 x i8>
8388 // CHECK:   [[TMP8:%.*]] = bitcast <2 x float> [[TMP7]] to <8 x i8>
8392 // CHECK:   [[TMP10:%.*]] = bitcast <2 x float> [[TMP9]] to <8 x i8>
8396 // CHECK:   [[TMP12:%.*]] = bitcast <2 x float> [[TMP11]] to <8 x i8>
8397 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP6]] to <2 x float>
8398 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP8]] to <2 x float>
8399 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP10]] to <2 x float>
8400 // CHECK:   [[TMP16:%.*]] = bitcast <8 x i8> [[TMP12]] to <2 x float>
8402 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <2 x float>, <2 x float>, <2 x float>, <2 x float> }*
8404 // CHECK:   [[TMP18:%.*]] = bitcast %struct.float32x2x4_t* %agg.result to i8*
8405 // CHECK:   [[TMP19:%.*]] = bitcast %struct.float32x2x4_t* [[__RET]] to i8*
8417 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
8419 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly8x8x4_t* [[__S1]] to i8*
8420 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x8x4_t* [[B]] to i8*
8422 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x8x4_t* [[__RET]] to i8*
8436 // CHECK:   [[TMP8:%.*]] = bitcast i8* [[TMP3]] to { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> }*
8438 // CHECK:   [[TMP9:%.*]] = bitcast %struct.poly8x8x4_t* %agg.result to i8*
8439 // CHECK:   [[TMP10:%.*]] = bitcast %struct.poly8x8x4_t* [[__RET]] to i8*
8451 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
8453 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x4x4_t* [[__S1]] to i8*
8454 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x4x4_t* [[B]] to i8*
8456 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly16x4x4_t* [[__RET]] to i8*
8457 // CHECK:   [[TMP4:%.*]] = bitcast i16* %a to i8*
8461 // CHECK:   [[TMP6:%.*]] = bitcast <4 x i16> [[TMP5]] to <8 x i8>
8465 // CHECK:   [[TMP8:%.*]] = bitcast <4 x i16> [[TMP7]] to <8 x i8>
8469 // CHECK:   [[TMP10:%.*]] = bitcast <4 x i16> [[TMP9]] to <8 x i8>
8473 // CHECK:   [[TMP12:%.*]] = bitcast <4 x i16> [[TMP11]] to <8 x i8>
8474 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP6]] to <4 x i16>
8475 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP8]] to <4 x i16>
8476 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP10]] to <4 x i16>
8477 // CHECK:   [[TMP16:%.*]] = bitcast <8 x i8> [[TMP12]] to <4 x i16>
8479 // CHECK:   [[TMP17:%.*]] = bitcast i8* [[TMP3]] to { <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16> }*
8481 // CHECK:   [[TMP18:%.*]] = bitcast %struct.poly16x4x4_t* %agg.result to i8*
8482 // CHECK:   [[TMP19:%.*]] = bitcast %struct.poly16x4x4_t* [[__RET]] to i8*
8498 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
8499 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
8500 // CHECK:   [[VMAX_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
8501 // CHECK:   [[VMAX_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
8503 // CHECK:   [[VMAX_V3_I:%.*]] = bitcast <4 x i16> [[VMAX_V2_I]] to <8 x i8>
8504 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VMAX_V3_I]] to <4 x i16>
8511 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
8512 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
8513 // CHECK:   [[VMAX_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
8514 // CHECK:   [[VMAX_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
8516 // CHECK:   [[VMAX_V3_I:%.*]] = bitcast <2 x i32> [[VMAX_V2_I]] to <8 x i8>
8517 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VMAX_V3_I]] to <2 x i32>
8531 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
8532 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
8533 // CHECK:   [[VMAX_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
8534 // CHECK:   [[VMAX_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
8536 // CHECK:   [[VMAX_V3_I:%.*]] = bitcast <4 x i16> [[VMAX_V2_I]] to <8 x i8>
8537 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VMAX_V3_I]] to <4 x i16>
8544 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
8545 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
8546 // CHECK:   [[VMAX_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
8547 // CHECK:   [[VMAX_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
8549 // CHECK:   [[VMAX_V3_I:%.*]] = bitcast <2 x i32> [[VMAX_V2_I]] to <8 x i8>
8550 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VMAX_V3_I]] to <2 x i32>
8557 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
8558 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
8559 // CHECK:   [[VMAX_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
8560 // CHECK:   [[VMAX_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
8562 // CHECK:   [[VMAX_V3_I:%.*]] = bitcast <2 x float> [[VMAX_V2_I]] to <8 x i8>
8563 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VMAX_V3_I]] to <2 x float>
8577 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
8578 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
8579 // CHECK:   [[VMAXQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
8580 // CHECK:   [[VMAXQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
8582 // CHECK:   [[VMAXQ_V3_I:%.*]] = bitcast <8 x i16> [[VMAXQ_V2_I]] to <16 x i8>
8583 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VMAXQ_V3_I]] to <8 x i16>
8590 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
8591 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
8592 // CHECK:   [[VMAXQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
8593 // CHECK:   [[VMAXQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
8595 // CHECK:   [[VMAXQ_V3_I:%.*]] = bitcast <4 x i32> [[VMAXQ_V2_I]] to <16 x i8>
8596 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VMAXQ_V3_I]] to <4 x i32>
8610 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
8611 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
8612 // CHECK:   [[VMAXQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
8613 // CHECK:   [[VMAXQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
8615 // CHECK:   [[VMAXQ_V3_I:%.*]] = bitcast <8 x i16> [[VMAXQ_V2_I]] to <16 x i8>
8616 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VMAXQ_V3_I]] to <8 x i16>
8623 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
8624 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
8625 // CHECK:   [[VMAXQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
8626 // CHECK:   [[VMAXQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
8628 // CHECK:   [[VMAXQ_V3_I:%.*]] = bitcast <4 x i32> [[VMAXQ_V2_I]] to <16 x i8>
8629 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VMAXQ_V3_I]] to <4 x i32>
8636 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
8637 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
8638 // CHECK:   [[VMAXQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
8639 // CHECK:   [[VMAXQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
8641 // CHECK:   [[VMAXQ_V3_I:%.*]] = bitcast <4 x float> [[VMAXQ_V2_I]] to <16 x i8>
8642 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VMAXQ_V3_I]] to <4 x float>
8657 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
8658 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
8659 // CHECK:   [[VMIN_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
8660 // CHECK:   [[VMIN_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
8662 // CHECK:   [[VMIN_V3_I:%.*]] = bitcast <4 x i16> [[VMIN_V2_I]] to <8 x i8>
8663 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VMIN_V3_I]] to <4 x i16>
8670 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
8671 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
8672 // CHECK:   [[VMIN_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
8673 // CHECK:   [[VMIN_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
8675 // CHECK:   [[VMIN_V3_I:%.*]] = bitcast <2 x i32> [[VMIN_V2_I]] to <8 x i8>
8676 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VMIN_V3_I]] to <2 x i32>
8690 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
8691 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
8692 // CHECK:   [[VMIN_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
8693 // CHECK:   [[VMIN_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
8695 // CHECK:   [[VMIN_V3_I:%.*]] = bitcast <4 x i16> [[VMIN_V2_I]] to <8 x i8>
8696 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VMIN_V3_I]] to <4 x i16>
8703 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
8704 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
8705 // CHECK:   [[VMIN_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
8706 // CHECK:   [[VMIN_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
8708 // CHECK:   [[VMIN_V3_I:%.*]] = bitcast <2 x i32> [[VMIN_V2_I]] to <8 x i8>
8709 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VMIN_V3_I]] to <2 x i32>
8716 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
8717 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
8718 // CHECK:   [[VMIN_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
8719 // CHECK:   [[VMIN_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
8721 // CHECK:   [[VMIN_V3_I:%.*]] = bitcast <2 x float> [[VMIN_V2_I]] to <8 x i8>
8722 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VMIN_V3_I]] to <2 x float>
8736 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
8737 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
8738 // CHECK:   [[VMINQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
8739 // CHECK:   [[VMINQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
8741 // CHECK:   [[VMINQ_V3_I:%.*]] = bitcast <8 x i16> [[VMINQ_V2_I]] to <16 x i8>
8742 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VMINQ_V3_I]] to <8 x i16>
8749 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
8750 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
8751 // CHECK:   [[VMINQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
8752 // CHECK:   [[VMINQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
8754 // CHECK:   [[VMINQ_V3_I:%.*]] = bitcast <4 x i32> [[VMINQ_V2_I]] to <16 x i8>
8755 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VMINQ_V3_I]] to <4 x i32>
8769 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
8770 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
8771 // CHECK:   [[VMINQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
8772 // CHECK:   [[VMINQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
8774 // CHECK:   [[VMINQ_V3_I:%.*]] = bitcast <8 x i16> [[VMINQ_V2_I]] to <16 x i8>
8775 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VMINQ_V3_I]] to <8 x i16>
8782 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
8783 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
8784 // CHECK:   [[VMINQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
8785 // CHECK:   [[VMINQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
8787 // CHECK:   [[VMINQ_V3_I:%.*]] = bitcast <4 x i32> [[VMINQ_V2_I]] to <16 x i8>
8788 // CHECK:   [[TMP2:%.*]] = bitcast
8795 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
8796 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
8797 // CHECK:   [[VMINQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
8798 // CHECK:   [[VMINQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
8800 // CHECK:   [[VMINQ_V3_I:%.*]] = bitcast <4 x float> [[VMINQ_V2_I]] to <16 x i8>
8801 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VMINQ_V3_I]] to <4 x float>
8930 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
8931 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %c to <8 x i8>
8932 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
8933 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
8942 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
8943 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %c to <8 x i8>
8944 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
8945 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
8962 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
8963 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %c to <8 x i8>
8964 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
8965 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
8974 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
8975 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %c to <8 x i8>
8976 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
8977 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
8988 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
8989 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[SHUFFLE]] to <8 x i8>
8990 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
8991 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
9001 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
9002 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
9003 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
9004 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
9014 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
9015 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[SHUFFLE]] to <8 x i8>
9016 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
9017 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
9027 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
9028 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
9029 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
9030 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
9044 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
9045 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[VECINIT3_I]] to <8 x i8>
9046 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
9047 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
9058 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
9059 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[VECINIT1_I]] to <8 x i8>
9060 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
9061 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
9074 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
9075 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[VECINIT3_I]] to <8 x i8>
9076 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
9077 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
9088 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
9089 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[VECINIT1_I]] to <8 x i8>
9090 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
9091 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
9436 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
9437 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %c to <8 x i8>
9438 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
9439 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
9448 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
9449 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %c to <8 x i8>
9450 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
9451 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
9468 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
9469 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %c to <8 x i8>
9470 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
9471 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
9480 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
9481 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %c to <8 x i8>
9482 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
9483 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
9494 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
9495 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[SHUFFLE]] to <8 x i8>
9496 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
9497 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
9507 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
9508 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
9509 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
9510 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
9520 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
9521 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[SHUFFLE]] to <8 x i8>
9522 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
9523 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
9533 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
9534 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
9535 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
9536 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
9550 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
9551 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[VECINIT3_I]] to <8 x i8>
9552 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
9553 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
9564 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
9565 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[VECINIT1_I]] to <8 x i8>
9566 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
9567 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
9580 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
9581 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[VECINIT3_I]] to <8 x i8>
9582 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
9583 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
9594 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
9595 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[VECINIT1_I]] to <8 x i8>
9596 // CHECK:   [[VMULL_I_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
9597 // CHECK:   [[VMULL1_I_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
9828 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
9829 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
9837 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
9838 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
9853 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
9854 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
9862 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
9863 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
9872 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
9873 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
9881 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
9882 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
9890 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
9891 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
9899 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
9900 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
9908 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
9909 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
9917 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
9918 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
10328 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
10329 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
10330 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
10331 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
10339 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
10340 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
10341 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
10342 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
10357 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
10358 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
10359 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
10360 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
10368 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
10369 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
10370 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
10371 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
10388 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
10389 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[SHUFFLE]] to <8 x i8>
10390 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
10391 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
10400 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
10401 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
10402 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
10403 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
10412 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
10413 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[SHUFFLE]] to <8 x i8>
10414 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
10415 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
10424 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
10425 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
10426 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
10427 // CHECK:   [[VMULL1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
10436 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
10441 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[VECINIT3_I]] to <8 x i8>
10442 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
10443 // CHECK:   [[VMULL4_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
10451 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
10454 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[VECINIT1_I]] to <8 x i8>
10455 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
10456 // CHECK:   [[VMULL2_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
10464 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
10469 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[VECINIT3_I]] to <8 x i8>
10470 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
10471 // CHECK:   [[VMULL4_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
10479 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
10482 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[VECINIT1_I]] to <8 x i8>
10483 // CHECK:   [[VMULL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
10484 // CHECK:   [[VMULL2_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11100 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
11101 // CHECK:   [[VPADAL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
11109 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
11110 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11111 // CHECK:   [[VPADAL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
11112 // CHECK:   [[VPADAL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11120 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
11121 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11122 // CHECK:   [[VPADAL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
11123 // CHECK:   [[VPADAL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11131 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
11132 // CHECK:   [[VPADAL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
11140 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
11141 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11142 // CHECK:   [[VPADAL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
11143 // CHECK:   [[VPADAL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11151 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
11152 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11153 // CHECK:   [[VPADAL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
11154 // CHECK:   [[VPADAL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11162 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
11163 // CHECK:   [[VPADALQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
11171 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
11172 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
11173 // CHECK:   [[VPADALQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
11174 // CHECK:   [[VPADALQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
11182 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
11183 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
11184 // CHECK:   [[VPADALQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
11185 // CHECK:   [[VPADALQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
11193 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
11194 // CHECK:   [[VPADALQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
11202 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
11203 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
11204 // CHECK:   [[VPADALQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
11205 // CHECK:   [[VPADALQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
11213 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
11214 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
11215 // CHECK:   [[VPADALQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
11216 // CHECK:   [[VPADALQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
11232 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
11233 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11234 // CHECK:   [[VPADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
11235 // CHECK:   [[VPADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11237 // CHECK:   [[VPADD_V3_I:%.*]] = bitcast <4 x i16> [[VPADD_V2_I]] to <8 x i8>
11238 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPADD_V3_I]] to <4 x i16>
11245 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
11246 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11247 // CHECK:   [[VPADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
11248 // CHECK:   [[VPADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11250 // CHECK:   [[VPADD_V3_I:%.*]] = bitcast <2 x i32> [[VPADD_V2_I]] to <8 x i8>
11251 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPADD_V3_I]] to <2 x i32>
11265 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
11266 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11267 // CHECK:   [[VPADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
11268 // CHECK:   [[VPADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11270 // CHECK:   [[VPADD_V3_I:%.*]] = bitcast <4 x i16> [[VPADD_V2_I]] to <8 x i8>
11271 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPADD_V3_I]] to <4 x i16>
11278 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
11279 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11280 // CHECK:   [[VPADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
11281 // CHECK:   [[VPADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11283 // CHECK:   [[VPADD_V3_I:%.*]] = bitcast <2 x i32> [[VPADD_V2_I]] to <8 x i8>
11284 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPADD_V3_I]] to <2 x i32>
11291 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
11292 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
11293 // CHECK:   [[VPADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
11294 // CHECK:   [[VPADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
11296 // CHECK:   [[VPADD_V3_I:%.*]] = bitcast <2 x float> [[VPADD_V2_I]] to <8 x i8>
11297 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPADD_V3_I]] to <2 x float>
11312 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
11313 // CHECK:   [[VPADDL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
11321 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
11322 // CHECK:   [[VPADDL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
11337 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
11338 // CHECK:   [[VPADDL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
11346 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
11347 // CHECK:   [[VPADDL_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
11362 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
11363 // CHECK:   [[VPADDL_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
11371 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
11372 // CHECK:   [[VPADDL_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
11387 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
11388 // CHECK:   [[VPADDL_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
11396 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
11397 // CHECK:   [[VPADDL_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
11413 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
11414 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11415 // CHECK:   [[VPMAX_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
11416 bitcast <8 x i8> [[TMP1]] to <4 x i16>
11418 // CHECK:   [[VPMAX_V3_I:%.*]] = bitcast <4 x i16> [[VPMAX_V2_I]] to <8 x i8>
11419 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPMAX_V3_I]] to <4 x i16>
11426 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
11427 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11428 // CHECK:   [[VPMAX_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
11429 // CHECK:   [[VPMAX_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11431 // CHECK:   [[VPMAX_V3_I:%.*]] = bitcast <2 x i32> [[VPMAX_V2_I]] to <8 x i8>
11432 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPMAX_V3_I]] to <2 x i32>
11446 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
11447 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11448 // CHECK:   [[VPMAX_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
11449 // CHECK:   [[VPMAX_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11451 // CHECK:   [[VPMAX_V3_I:%.*]] = bitcast <4 x i16> [[VPMAX_V2_I]] to <8 x i8>
11452 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPMAX_V3_I]] to <4 x i16>
11459 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
11460 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11461 // CHECK:   [[VPMAX_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
11462 // CHECK:   [[VPMAX_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11464 // CHECK:   [[VPMAX_V3_I:%.*]] = bitcast <2 x i32> [[VPMAX_V2_I]] to <8 x i8>
11465 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPMAX_V3_I]] to <2 x i32>
11472 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
11473 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
11474 // CHECK:   [[VPMAX_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
11475 // CHECK:   [[VPMAX_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
11477 // CHECK:   [[VPMAX_V3_I:%.*]] = bitcast <2 x float> [[VPMAX_V2_I]] to <8 x i8>
11478 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPMAX_V3_I]] to <2 x float>
11493 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
11494 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11495 // CHECK:   [[VPMIN_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
11496 // CHECK:   [[VPMIN_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11498 // CHECK:   [[VPMIN_V3_I:%.*]] = bitcast <4 x i16> [[VPMIN_V2_I]] to <8 x i8>
11499 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPMIN_V3_I]] to <4 x i16>
11506 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
11507 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11508 // CHECK:   [[VPMIN_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
11509 // CHECK:   [[VPMIN_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11511 // CHECK:   [[VPMIN_V3_I:%.*]] = bitcast <2 x i32> [[VPMIN_V2_I]] to <8 x i8>
11512 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPMIN_V3_I]] to <2 x i32>
11526 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
11527 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11528 // CHECK:   [[VPMIN_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
11529 // CHECK:   [[VPMIN_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11531 // CHECK:   [[VPMIN_V3_I:%.*]] = bitcast <4 x i16> [[VPMIN_V2_I]] to <8 x i8>
11532 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPMIN_V3_I]] to <4 x i16>
11539 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
11540 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11541 // CHECK:   [[VPMIN_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
11542 // CHECK:   [[VPMIN_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11544 // CHECK:   [[VPMIN_V3_I:%.*]] = bitcast <2 x i32> [[VPMIN_V2_I]] to <8 x i8>
11545 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPMIN_V3_I]] to <2 x i32>
11552 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
11553 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
11554 // CHECK:   [[VPMIN_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
11555 // CHECK:   [[VPMIN_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
11557 // CHECK:   [[VPMIN_V3_I:%.*]] = bitcast <2 x float> [[VPMIN_V2_I]] to <8 x i8>
11558 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VPMIN_V3_I]] to <2 x float>
11573 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
11574 // CHECK:   [[VQABS_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
11576 // CHECK:   [[VQABS_V2_I:%.*]] = bitcast <4 x i16> [[VQABS_V1_I]] to <8 x i8>
11577 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VQABS_V2_I]] to <4 x i16>
11584 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
11585 // CHECK:   [[VQABS_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
11587 // CHECK:   [[VQABS_V2_I:%.*]] = bitcast <2 x i32> [[VQABS_V1_I]] to <8 x i8>
11588 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VQABS_V2_I]] to <2 x i32>
11602 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
11603 // CHECK:   [[VQABSQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
11605 // CHECK:   [[VQABSQ_V2_I:%.*]] = bitcast <8 x i16> [[VQABSQ_V1_I]] to <16 x i8>
11606 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[VQABSQ_V2_I]] to <8 x i16>
11613 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
11614 // CHECK:   [[VQABSQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
11616 // CHECK:   [[VQABSQ_V2_I:%.*]] = bitcast <4 x i32> [[VQABSQ_V1_I]] to <16 x i8>
11617 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[VQABSQ_V2_I]] to <4 x i32>
11632 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
11633 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11634 // CHECK:   [[VQADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
11635 // CHECK:   [[VQADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11637 // CHECK:   [[VQADD_V3_I:%.*]] = bitcast <4 x i16> [[VQADD_V2_I]] to <8 x i8>
11638 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQADD_V3_I]] to <4 x i16>
11645 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
11646 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11647 // CHECK:   [[VQADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
11648 // CHECK:   [[VQADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11650 // CHECK:   [[VQADD_V3_I:%.*]] = bitcast <2 x i32> [[VQADD_V2_I]] to <8 x i8>
11651 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQADD_V3_I]] to <2 x i32>
11658 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
11659 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
11660 // CHECK:   [[VQADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
11661 // CHECK:   [[VQADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
11663 // CHECK:   [[VQADD_V3_I:%.*]] = bitcast <1 x i64> [[VQADD_V2_I]] to <8 x i8>
11664 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQADD_V3_I]] to <1 x i64>
11678 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
11679 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11680 // CHECK:   [[VQADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
11681 // CHECK:   [[VQADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11683 // CHECK:   [[VQADD_V3_I:%.*]] = bitcast <4 x i16> [[VQADD_V2_I]] to <8 x i8>
11684 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQADD_V3_I]] to <4 x i16>
11691 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
11692 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11693 // CHECK:   [[VQADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
11694 // CHECK:   [[VQADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11696 // CHECK:   [[VQADD_V3_I:%.*]] = bitcast <2 x i32> [[VQADD_V2_I]] to <8 x i8>
11697 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQADD_V3_I]] to <2 x i32>
11704 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
11705 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
11706 // CHECK:   [[VQADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
11707 // CHECK:   [[VQADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
11709 // CHECK:   [[VQADD_V3_I:%.*]] = bitcast <1 x i64> [[VQADD_V2_I]] to <8 x i8>
11710 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQADD_V3_I]] to <1 x i64>
11724 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
11725 bitcast <8 x i16> %b to <16 x i8>
11726 // CHECK:   [[VQADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
11727 // CHECK:   [[VQADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
11729 // CHECK:   [[VQADDQ_V3_I:%.*]] = bitcast <8 x i16> [[VQADDQ_V2_I]] to <16 x i8>
11730 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQADDQ_V3_I]] to <8 x i16>
11737 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
11738 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
11739 // CHECK:   [[VQADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
11740 // CHECK:   [[VQADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
11742 // CHECK:   [[VQADDQ_V3_I:%.*]] = bitcast <4 x i32> [[VQADDQ_V2_I]] to <16 x i8>
11743 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQADDQ_V3_I]] to <4 x i32>
11750 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
11751 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
11752 // CHECK:   [[VQADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
11753 // CHECK:   [[VQADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
11755 // CHECK:   [[VQADDQ_V3_I:%.*]] = bitcast <2 x i64> [[VQADDQ_V2_I]] to <16 x i8>
11756 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQADDQ_V3_I]] to <2 x i64>
11770 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
11771 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
11772 // CHECK:   [[VQADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
11773 // CHECK:   [[VQADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
11775 // CHECK:   [[VQADDQ_V3_I:%.*]] = bitcast <8 x i16> [[VQADDQ_V2_I]] to <16 x i8>
11776 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQADDQ_V3_I]] to <8 x i16>
11783 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
11784 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
11785 // CHECK:   [[VQADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
11786 // CHECK:   [[VQADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
11788 // CHECK:   [[VQADDQ_V3_I:%.*]] = bitcast <4 x i32> [[VQADDQ_V2_I]] to <16 x i8>
11789 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQADDQ_V3_I]] to <4 x i32>
11796 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
11797 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
11798 // CHECK:   [[VQADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
11799 // CHECK:   [[VQADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
11801 // CHECK:   [[VQADDQ_V3_I:%.*]] = bitcast <2 x i64> [[VQADDQ_V2_I]] to <16 x i8>
11802 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQADDQ_V3_I]] to <2 x i64>
11810 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
11811 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11812 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %c to <8 x i8>
11813 // CHECK:   [[VQDMLAL_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11814 // CHECK:   [[VQDMLAL1_I:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
11816 // CHECK:   [[VQDMLAL_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
11824 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
11825 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11826 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %c to <8 x i8>
11827 // CHECK:   [[VQDMLAL_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11828 // CHECK:   [[VQDMLAL1_I:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
11830 // CHECK:   [[VQDMLAL_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
11840 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
11841 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11842 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> [[SHUFFLE]] to <8 x i8>
11843 // CHECK:   [[VQDMLAL_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11844 // CHECK:   [[VQDMLAL1_I:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
11846 // CHECK:   [[VQDMLAL_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
11855 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
11856 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11857 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
11858 // CHECK:   [[VQDMLAL_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11859 // CHECK:   [[VQDMLAL1_I:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
11861 // CHECK:   [[VQDMLAL_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
11870 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
11871 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11876 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> [[VECINIT3_I]] to <8 x i8>
11877 // CHECK:   [[VQDMLAL_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11878 // CHECK:   [[VQDMLAL4_I:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
11880 // CHECK:   [[VQDMLAL_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
11888 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
11889 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11892 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> [[VECINIT1_I]] to <8 x i8>
11893 // CHECK:   [[VQDMLAL_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11894 // CHECK:   [[VQDMLAL2_I:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
11896 // CHECK:   [[VQDMLAL_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
11905 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
11906 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11907 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %c to <8 x i8>
11908 // CHECK:   [[VQDMLAL_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11909 // CHECK:   [[VQDMLAL1_I:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
11911 // CHECK:   [[VQDMLSL_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
11919 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
11920 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11921 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %c to <8 x i8>
11922 // CHECK:   [[VQDMLAL_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11923 // CHECK:   [[VQDMLAL1_I:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
11925 // CHECK:   [[VQDMLSL_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
11935 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
11936 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11937 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> [[SHUFFLE]] to <8 x i8>
11938 // CHECK:   [[VQDMLAL_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11939 // CHECK:   [[VQDMLAL1_I:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
11941 // CHECK:   [[VQDMLSL_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
11950 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
11951 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11952 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
11953 // CHECK:   [[VQDMLAL_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11954 // CHECK:   [[VQDMLAL1_I:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
11956 // CHECK:   [[VQDMLSL_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
11965 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
11966 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
11971 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> [[VECINIT3_I]] to <8 x i8>
11972 // CHECK:   [[VQDMLAL_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
11973 // CHECK:   [[VQDMLAL4_I:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
11975 // CHECK:   [[VQDMLSL_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
11983 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
11984 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
11987 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> [[VECINIT1_I]] to <8 x i8>
11988 // CHECK:   [[VQDMLAL_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
11989 // CHECK:   [[VQDMLAL2_I:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
11991 // CHECK:   [[VQDMLSL_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
12000 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12001 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
12002 // CHECK:   [[VQDMULH_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12003 // CHECK:   [[VQDMULH_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
12005 // CHECK:   [[VQDMULH_V3_I:%.*]] = bitcast <4 x i16> [[VQDMULH_V2_I]] to <8 x i8>
12006 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQDMULH_V3_I]] to <4 x i16>
12013 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12014 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
12015 // CHECK:   [[VQDMULH_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12016 // CHECK:   [[VQDMULH_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
12018 // CHECK:   [[VQDMULH_V3_I:%.*]] = bitcast <2 x i32> [[VQDMULH_V2_I]] to <8 x i8>
12019 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQDMULH_V3_I]] to <2 x i32>
12026 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12027 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
12028 // CHECK:   [[VQDMULHQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12029 // CHECK:   [[VQDMULHQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
12031 // CHECK:   [[VQDMULHQ_V3_I:%.*]] = bitcast <8 x i16> [[VQDMULHQ_V2_I]] to <16 x i8>
12032 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQDMULHQ_V3_I]] to <8 x i16>
12039 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12040 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
12041 // CHECK:   [[VQDMULHQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12042 // CHECK:   [[VQDMULHQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
12044 // CHECK:   [[VQDMULHQ_V3_I:%.*]] = bitcast <4 x i32> [[VQDMULHQ_V2_I]] to <16 x i8>
12045 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQDMULHQ_V3_I]] to <4 x i32>
12054 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12055 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[SHUFFLE]] to <8 x i8>
12056 // CHECK:   [[VQDMULH_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12057 // CHECK:   [[VQDMULH_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
12059 // CHECK:   [[VQDMULH_V3_I:%.*]] = bitcast <4 x i16> [[VQDMULH_V2_I]] to <8 x i8>
12060 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQDMULH_V3_I]] to <4 x i16>
12068 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12069 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
12070 // CHECK:   [[VQDMULH_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12071 // CHECK:   [[VQDMULH_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
12073 // CHECK:   [[VQDMULH_V3_I:%.*]] = bitcast <2 x i32> [[VQDMULH_V2_I]] to <8 x i8>
12074 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQDMULH_V3_I]] to <2 x i32>
12082 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12083 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> [[SHUFFLE]] to <16 x i8>
12084 // CHECK:   [[VQDMULHQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12085 // CHECK:   [[VQDMULHQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
12087 // CHECK:   [[VQDMULHQ_V3_I:%.*]] = bitcast <8 x i16> [[VQDMULHQ_V2_I]] to <16 x i8>
12088 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQDMULHQ_V3_I]] to <8 x i16>
12096 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12097 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> [[SHUFFLE]] to <16 x i8>
12098 // CHECK:   [[VQDMULHQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12099 // CHECK:   [[VQDMULHQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
12101 // CHECK:   [[VQDMULHQ_V3_I:%.*]] = bitcast <4 x i32> [[VQDMULHQ_V2_I]] to <16 x i8>
12102 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQDMULHQ_V3_I]] to <4 x i32>
12110 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12115 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[VECINIT3_I]] to <8 x i8>
12116 // CHECK:   [[VQDMULH_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12117 // CHECK:   [[VQDMULH_V4_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
12119 // CHECK:   [[VQDMULH_V6_I:%.*]] = bitcast <4 x i16> [[VQDMULH_V5_I]] to <8 x i8>
12120 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQDMULH_V6_I]] to <4 x i16>
12127 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12130 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[VECINIT1_I]] to <8 x i8>
12131 // CHECK:   [[VQDMULH_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12132 // CHECK:   [[VQDMULH_V2_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
12134 // CHECK:   [[VQDMULH_V4_I:%.*]] = bitcast <2 x i32> [[VQDMULH_V3_I]] to <8 x i8>
12135 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQDMULH_V4_I]] to <2 x i32>
12142 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12151 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> [[VECINIT7_I]] to <16 x i8>
12152 // CHECK:   [[VQDMULHQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12153 // CHECK:   [[VQDMULHQ_V8_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
12155 // CHECK:   [[VQDMULHQ_V10_I:%.*]] = bitcast <8 x i16> [[VQDMULHQ_V9_I]] to <16 x i8>
12156 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQDMULHQ_V10_I]] to <8 x i16>
12163 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12168 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> [[VECINIT3_I]] to <16 x i8>
12169 // CHECK:   [[VQDMULHQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12170 // CHECK:   [[VQDMULHQ_V4_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
12172 // CHECK:   [[VQDMULHQ_V6_I:%.*]] = bitcast <4 x i32> [[VQDMULHQ_V5_I]] to <16 x i8>
12173 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQDMULHQ_V6_I]] to <4 x i32>
12181 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12182 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
12183 // CHECK:   [[VQDMULL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12184 // CHECK:   [[VQDMULL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
12186 // CHECK:   [[VQDMULL_V3_I:%.*]] = bitcast <4 x i32> [[VQDMULL_V2_I]] to <16 x i8>
12187 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQDMULL_V3_I]] to <4 x i32>
12194 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12195 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
12196 // CHECK:   [[VQDMULL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12197 // CHECK:   [[VQDMULL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
12199 // CHECK:   [[VQDMULL_V3_I:%.*]] = bitcast <2 x i64> [[VQDMULL_V2_I]] to <16 x i8>
12200 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQDMULL_V3_I]] to <2 x i64>
12209 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12210 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[SHUFFLE]] to <8 x i8>
12211 // CHECK:   [[VQDMULL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12212 // CHECK:   [[VQDMULL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
12214 // CHECK:   [[VQDMULL_V3_I:%.*]] = bitcast <4 x i32> [[VQDMULL_V2_I]] to <16 x i8>
12215 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQDMULL_V3_I]] to <4 x i32>
12223 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12224 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
12225 // CHECK:   [[VQDMULL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12226 // CHECK:   [[VQDMULL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
12228 // CHECK:   [[VQDMULL_V3_I:%.*]] = bitcast <2 x i64> [[VQDMULL_V2_I]] to <16 x i8>
12229 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQDMULL_V3_I]] to <2 x i64>
12237 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12242 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[VECINIT3_I]] to <8 x i8>
12243 // CHECK:   [[VQDMULL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12244 // CHECK:   [[VQDMULL_V4_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
12246 // CHECK:   [[VQDMULL_V6_I:%.*]] = bitcast <4 x i32> [[VQDMULL_V5_I]] to <16 x i8>
12247 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQDMULL_V6_I]] to <4 x i32>
12254 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12257 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[VECINIT1_I]] to <8 x i8>
12258 // CHECK:   [[VQDMULL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12259 // CHECK:   [[VQDMULL_V2_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
12261 // CHECK:   [[VQDMULL_V4_I:%.*]] = bitcast <2 x i64> [[VQDMULL_V3_I]] to <16 x i8>
12262 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQDMULL_V4_I]] to <2 x i64>
12270 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12271 // CHECK:   [[VQMOVN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12279 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12280 // CHECK:   [[VQMOVN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12282 // CHECK:   [[VQMOVN_V2_I:%.*]] = bitcast <4 x i16> [[VQMOVN_V1_I]] to <8 x i8>
12283 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VQMOVN_V2_I]] to <4 x i16>
12290 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
12291 // CHECK:   [[VQMOVN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
12293 // CHECK:   [[VQMOVN_V2_I:%.*]] = bitcast <2 x i32> [[VQMOVN_V1_I]] to <8 x i8>
12294 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VQMOVN_V2_I]] to <2 x i32>
12301 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12302 // CHECK:   [[VQMOVN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12310 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12311 // CHECK:   [[VQMOVN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12313 // CHECK:   [[VQMOVN_V2_I:%.*]] = bitcast <4 x i16> [[VQMOVN_V1_I]] to <8 x i8>
12314 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VQMOVN_V2_I]] to <4 x i16>
12321 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
12322 // CHECK:   [[VQMOVN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
12324 // CHECK:   [[VQMOVN_V2_I:%.*]] = bitcast <2 x i32> [[VQMOVN_V1_I]] to <8 x i8>
12325 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VQMOVN_V2_I]] to <2 x i32>
12333 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12334 // CHECK:   [[VQMOVUN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12342 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12343 // CHECK:   [[VQMOVUN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12345 // CHECK:   [[VQMOVUN_V2_I:%.*]] = bitcast <4 x i16> [[VQMOVUN_V1_I]] to <8 x i8>
12346 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VQMOVUN_V2_I]] to <4 x i16>
12353 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
12354 // CHECK:   [[VQMOVUN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
12356 // CHECK:   [[VQMOVUN_V2_I:%.*]] = bitcast <2 x i32> [[VQMOVUN_V1_I]] to <8 x i8>
12357 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VQMOVUN_V2_I]] to <2 x i32>
12372 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12373 // CHECK:   [[VQNEG_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12375 // CHECK:   [[VQNEG_V2_I:%.*]] = bitcast <4 x i16> [[VQNEG_V1_I]] to <8 x i8>
12376 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VQNEG_V2_I]] to <4 x i16>
12383 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12384 // CHECK:   [[VQNEG_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12386 // CHECK:   [[VQNEG_V2_I:%.*]] = bitcast <2 x i32> [[VQNEG_V1_I]] to <8 x i8>
12387 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[VQNEG_V2_I]] to <2 x i32>
12401 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12402 // CHECK:   [[VQNEGQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12404 // CHECK:   [[VQNEGQ_V2_I:%.*]] = bitcast <8 x i16> [[VQNEGQ_V1_I]] to <16 x i8>
12405 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[VQNEGQ_V2_I]] to <8 x i16>
12412 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12413 // CHECK:   [[VQNEGQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12415 // CHECK:   [[VQNEGQ_V2_I:%.*]] = bitcast <4 x i32> [[VQNEGQ_V1_I]] to <16 x i8>
12416 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[VQNEGQ_V2_I]] to <4 x i32>
12424 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12425 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
12426 // CHECK:   [[VQRDMULH_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12427 // CHECK:   [[VQRDMULH_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
12429 // CHECK:   [[VQRDMULH_V3_I:%.*]] = bitcast <4 x i16> [[VQRDMULH_V2_I]] to <8 x i8>
12430 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQRDMULH_V3_I]] to <4 x i16>
12437 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12438 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
12439 // CHECK:   [[VQRDMULH_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12440 // CHECK:   [[VQRDMULH_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
12442 // CHECK:   [[VQRDMULH_V3_I:%.*]] = bitcast <2 x i32> [[VQRDMULH_V2_I]] to <8 x i8>
12443 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQRDMULH_V3_I]] to <2 x i32>
12450 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12451 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
12452 // CHECK:   [[VQRDMULHQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12453 // CHECK:   [[VQRDMULHQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
12455 // CHECK:   [[VQRDMULHQ_V3_I:%.*]] = bitcast <8 x i16> [[VQRDMULHQ_V2_I]] to <16 x i8>
12456 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQRDMULHQ_V3_I]] to <8 x i16>
12463 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12464 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
12465 // CHECK:   [[VQRDMULHQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12466 // CHECK:   [[VQRDMULHQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
12468 // CHECK:   [[VQRDMULHQ_V3_I:%.*]] = bitcast <4 x i32> [[VQRDMULHQ_V2_I]] to <16 x i8>
12469 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQRDMULHQ_V3_I]] to <4 x i32>
12478 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12479 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[SHUFFLE]] to <8 x i8>
12480 // CHECK:   [[VQRDMULH_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12481 // CHECK:   [[VQRDMULH_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
12483 // CHECK:   [[VQRDMULH_V3_I:%.*]] = bitcast <4 x i16> [[VQRDMULH_V2_I]] to <8 x i8>
12484 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQRDMULH_V3_I]] to <4 x i16>
12492 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12493 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
12494 // CHECK:   [[VQRDMULH_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12495 // CHECK:   [[VQRDMULH_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
12497 // CHECK:   [[VQRDMULH_V3_I:%.*]] = bitcast <2 x i32> [[VQRDMULH_V2_I]] to <8 x i8>
12498 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQRDMULH_V3_I]] to <2 x i32>
12506 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12507 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> [[SHUFFLE]] to <16 x i8>
12508 // CHECK:   [[VQRDMULHQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12509 // CHECK:   [[VQRDMULHQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
12511 // CHECK:   [[VQRDMULHQ_V3_I:%.*]] = bitcast <8 x i16> [[VQRDMULHQ_V2_I]] to <16 x i8>
12512 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQRDMULHQ_V3_I]] to <8 x i16>
12520 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12521 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> [[SHUFFLE]] to <16 x i8>
12522 // CHECK:   [[VQRDMULHQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12523 // CHECK:   [[VQRDMULHQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
12525 // CHECK:   [[VQRDMULHQ_V3_I:%.*]] = bitcast <4 x i32> [[VQRDMULHQ_V2_I]] to <16 x i8>
12526 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQRDMULHQ_V3_I]] to <4 x i32>
12534 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12539 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> [[VECINIT3_I]] to <8 x i8>
12540 // CHECK:   [[VQRDMULH_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12541 // CHECK:   [[VQRDMULH_V4_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
12543 // CHECK:   [[VQRDMULH_V6_I:%.*]] = bitcast <4 x i16> [[VQRDMULH_V5_I]] to <8 x i8>
12544 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQRDMULH_V6_I]] to <4 x i16>
12551 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12554 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> [[VECINIT1_I]] to <8 x i8>
12555 // CHECK:   [[VQRDMULH_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12556 // CHECK:   [[VQRDMULH_V2_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
12558 // CHECK:   [[VQRDMULH_V4_I:%.*]] = bitcast <2 x i32> [[VQRDMULH_V3_I]] to <8 x i8>
12559 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQRDMULH_V4_I]] to <2 x i32>
12566 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12575 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> [[VECINIT7_I]] to <16 x i8>
12576 // CHECK:   [[VQRDMULHQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12577 // CHECK:   [[VQRDMULHQ_V8_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
12579 // CHECK:   [[VQRDMULHQ_V10_I:%.*]] = bitcast <8 x i16> [[VQRDMULHQ_V9_I]] to <16 x i8>
12580 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQRDMULHQ_V10_I]] to <8 x i16>
12587 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12592 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> [[VECINIT3_I]] to <16 x i8>
12593 // CHECK:   [[VQRDMULHQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12594 // CHECK:   [[VQRDMULHQ_V4_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
12596 // CHECK:   [[VQRDMULHQ_V6_I:%.*]] = bitcast <4 x i32> [[VQRDMULHQ_V5_I]] to <16 x i8>
12597 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQRDMULHQ_V6_I]] to <4 x i32>
12612 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12613 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
12614 // CHECK:   [[VQRSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12615 // CHECK:   [[VQRSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
12617 // CHECK:   [[VQRSHL_V3_I:%.*]] = bitcast <4 x i16> [[VQRSHL_V2_I]] to <8 x i8>
12618 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQRSHL_V3_I]] to <4 x i16>
12625 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12626 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
12627 // CHECK:   [[VQRSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12628 // CHECK:   [[VQRSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
12630 // CHECK:   [[VQRSHL_V3_I:%.*]] = bitcast <2 x i32> [[VQRSHL_V2_I]] to <8 x i8>
12631 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQRSHL_V3_I]] to <2 x i32>
12638 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
12639 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
12640 // CHECK:   [[VQRSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
12641 // CHECK:   [[VQRSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
12643 // CHECK:   [[VQRSHL_V3_I:%.*]] = bitcast <1 x i64> [[VQRSHL_V2_I]] to <8 x i8>
12644 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQRSHL_V3_I]] to <1 x i64>
12658 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12659 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
12660 // CHECK:   [[VQRSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12661 // CHECK:   [[VQRSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
12663 // CHECK:   [[VQRSHL_V3_I:%.*]] = bitcast <4 x i16> [[VQRSHL_V2_I]] to <8 x i8>
12664 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQRSHL_V3_I]] to <4 x i16>
12671 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12672 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
12673 // CHECK:   [[VQRSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12674 // CHECK:   [[VQRSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
12676 // CHECK:   [[VQRSHL_V3_I:%.*]] = bitcast <2 x i32> [[VQRSHL_V2_I]] to <8 x i8>
12677 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQRSHL_V3_I]] to <2 x i32>
12684 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
12685 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
12686 // CHECK:   [[VQRSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
12687 // CHECK:   [[VQRSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
12689 // CHECK:   [[VQRSHL_V3_I:%.*]] = bitcast <1 x i64> [[VQRSHL_V2_I]] to <8 x i8>
12690 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQRSHL_V3_I]] to <1 x i64>
12704 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12705 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
12706 // CHECK:   [[VQRSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12707 // CHECK:   [[VQRSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
12709 // CHECK:   [[VQRSHLQ_V3_I:%.*]] = bitcast <8 x i16> [[VQRSHLQ_V2_I]] to <16 x i8>
12710 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQRSHLQ_V3_I]] to <8 x i16>
12717 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12718 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
12719 // CHECK:   [[VQRSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12720 // CHECK:   [[VQRSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
12722 // CHECK:   [[VQRSHLQ_V3_I:%.*]] = bitcast <4 x i32> [[VQRSHLQ_V2_I]] to <16 x i8>
12723 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQRSHLQ_V3_I]] to <4 x i32>
12730 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
12731 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
12732 // CHECK:   [[VQRSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
12733 // CHECK:   [[VQRSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
12735 // CHECK:   [[VQRSHLQ_V3_I:%.*]] = bitcast <2 x i64> [[VQRSHLQ_V2_I]] to <16 x i8>
12736 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQRSHLQ_V3_I]] to <2 x i64>
12750 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12751 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
12752 // CHECK:   [[VQRSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12753 // CHECK:   [[VQRSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
12755 // CHECK:   [[VQRSHLQ_V3_I:%.*]] = bitcast <8 x i16> [[VQRSHLQ_V2_I]] to <16 x i8>
12756 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQRSHLQ_V3_I]] to <8 x i16>
12763 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12764 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
12765 // CHECK:   [[VQRSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12766 // CHECK:   [[VQRSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
12768 // CHECK:   [[VQRSHLQ_V3_I:%.*]] = bitcast <4 x i32> [[VQRSHLQ_V2_I]] to <16 x i8>
12769 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQRSHLQ_V3_I]] to <4 x i32>
12776 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
12777 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
12778 // CHECK:   [[VQRSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
12779 // CHECK:   [[VQRSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
12781 // CHECK:   [[VQRSHLQ_V3_I:%.*]] = bitcast <2 x i64> [[VQRSHLQ_V2_I]] to <16 x i8>
12782 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQRSHLQ_V3_I]] to <2 x i64>
12790 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12791 // CHECK:   [[VQRSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12799 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12800 // CHECK:   [[VQRSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12808 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
12809 // CHECK:   [[VQRSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
12817 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12818 // CHECK:   [[VQRSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12826 // CHECK:   [[TMP0:%.*]] = bitcast
12827 // CHECK:   [[VQRSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12835 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
12836 // CHECK:   [[VQRSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
12845 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12846 // CHECK:   [[VQRSHRUN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12854 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12855 // CHECK:   [[VQRSHRUN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12863 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
12864 // CHECK:   [[VQRSHRUN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
12880 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12881 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
12882 // CHECK:   [[VQSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12883 // CHECK:   [[VQSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
12885 // CHECK:   [[VQSHL_V3_I:%.*]] = bitcast <4 x i16> [[VQSHL_V2_I]] to <8 x i8>
12886 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQSHL_V3_I]] to <4 x i16>
12893 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12894 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
12895 // CHECK:   [[VQSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12896 // CHECK:   [[VQSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
12898 // CHECK:   [[VQSHL_V3_I:%.*]] = bitcast <2 x i32> [[VQSHL_V2_I]] to <8 x i8>
12899 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQSHL_V3_I]] to <2 x i32>
12906 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
12907 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
12908 // CHECK:   [[VQSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
12909 // CHECK:   [[VQSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
12911 // CHECK:   [[VQSHL_V3_I:%.*]] = bitcast <1 x i64> [[VQSHL_V2_I]] to <8 x i8>
12912 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQSHL_V3_I]] to <1 x i64>
12926 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
12927 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
12928 // CHECK:   [[VQSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
12929 // CHECK:   [[VQSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
12931 // CHECK:   [[VQSHL_V3_I:%.*]] = bitcast <4 x i16> [[VQSHL_V2_I]] to <8 x i8>
12932 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQSHL_V3_I]] to <4 x i16>
12939 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
12940 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
12941 // CHECK:   [[VQSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
12942 // CHECK:   [[VQSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
12944 // CHECK:   [[VQSHL_V3_I:%.*]] = bitcast <2 x i32> [[VQSHL_V2_I]] to <8 x i8>
12945 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQSHL_V3_I]] to <2 x i32>
12952 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
12953 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
12954 // CHECK:   [[VQSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
12955 // CHECK:   [[VQSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
12957 // CHECK:   [[VQSHL_V3_I:%.*]] = bitcast <1 x i64> [[VQSHL_V2_I]] to <8 x i8>
12958 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQSHL_V3_I]] to <1 x i64>
12972 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
12973 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
12974 // CHECK:   [[VQSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
12975 // CHECK:   [[VQSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
12977 // CHECK:   [[VQSHLQ_V3_I:%.*]] = bitcast <8 x i16> [[VQSHLQ_V2_I]] to <16 x i8>
12978 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQSHLQ_V3_I]] to <8 x i16>
12985 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
12986 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
12987 // CHECK:   [[VQSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
12988 // CHECK:   [[VQSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
12990 // CHECK:   [[VQSHLQ_V3_I:%.*]] = bitcast <4 x i32> [[VQSHLQ_V2_I]] to <16 x i8>
12991 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQSHLQ_V3_I]] to <4 x i32>
12998 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
12999 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
13000 // CHECK:   [[VQSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
13001 // CHECK:   [[VQSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
13003 // CHECK:   [[VQSHLQ_V3_I:%.*]] = bitcast <2 x i64> [[VQSHLQ_V2_I]] to <16 x i8>
13004 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQSHLQ_V3_I]] to <2 x i64>
13018 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
13019 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
13020 // CHECK:   [[VQSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
13021 // CHECK:   [[VQSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
13023 // CHECK:   [[VQSHLQ_V3_I:%.*]] = bitcast <8 x i16> [[VQSHLQ_V2_I]] to <16 x i8>
13024 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQSHLQ_V3_I]] to <8 x i16>
13031 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
13032 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
13033 // CHECK:   [[VQSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
13034 // CHECK:   [[VQSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
13036 // CHECK:   [[VQSHLQ_V3_I:%.*]] = bitcast <4 x i32> [[VQSHLQ_V2_I]] to <16 x i8>
13037 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQSHLQ_V3_I]] to <4 x i32>
13044 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
13045 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
13046 // CHECK:   [[VQSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
13047 // CHECK:   [[VQSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
13049 // CHECK:   [[VQSHLQ_V3_I:%.*]] = bitcast <2 x i64> [[VQSHLQ_V2_I]] to <16 x i8>
13050 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQSHLQ_V3_I]] to <2 x i64>
13065 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
13066 // CHECK:   [[VQSHLU_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
13074 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
13075 // CHECK:   [[VQSHLU_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
13083 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
13084 // CHECK:   [[VQSHLU_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
13099 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
13100 // CHECK:   [[VQSHLU_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
13108 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
13109 // CHECK:   [[VQSHLU_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
13117 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
13118 // CHECK:   [[VQSHLU_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
13134 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
13135 // CHECK:   [[VQSHL_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
13143 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
13144 // CHECK:   [[VQSHL_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
13152 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
13153 // CHECK:   [[VQSHL_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
13168 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
13169 // CHECK:   [[VQSHL_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
13177 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
13178 // CHECK:   [[VQSHL_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
13186 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
13187 // CHECK:   [[VQSHL_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
13202 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
13203 // CHECK:   [[VQSHL_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
13211 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
13212 // CHECK:   [[VQSHL_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
13220 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
13221 // CHECK:   [[VQSHL_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
13236 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
13237 // CHECK:   [[VQSHL_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
13245 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
13246 // CHECK:   [[VQSHL_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
13254 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
13255 // CHECK:   [[VQSHL_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
13264 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
13265 // CHECK:   [[VQSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
13273 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
13274 // CHECK:   [[VQSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
13282 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
13283 // CHECK:   [[VQSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
13291 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
13292 // CHECK:   [[VQSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
13300 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
13301 // CHECK:   [[VQSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
13309 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
13310 // CHECK:   [[VQSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
13319 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
13320 // CHECK:   [[VQSHRUN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
13328 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
13329 // CHECK:   [[VQSHRUN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
13337 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
13338 // CHECK:   [[VQSHRUN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
13354 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
13355 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
13356 // CHECK:   [[VQSUB_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
13357 // CHECK:   [[VQSUB_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
13359 // CHECK:   [[VQSUB_V3_I:%.*]] = bitcast <4 x i16> [[VQSUB_V2_I]] to <8 x i8>
13360 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQSUB_V3_I]] to <4 x i16>
13367 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
13368 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
13369 // CHECK:   [[VQSUB_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
13370 // CHECK:   [[VQSUB_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
13372 // CHECK:   [[VQSUB_V3_I:%.*]] = bitcast <2 x i32> [[VQSUB_V2_I]] to <8 x i8>
13373 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQSUB_V3_I]] to <2 x i32>
13380 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
13381 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
13382 // CHECK:   [[VQSUB_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
13383 // CHECK:   [[VQSUB_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
13385 // CHECK:   [[VQSUB_V3_I:%.*]] = bitcast <1 x i64> [[VQSUB_V2_I]] to <8 x i8>
13386 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQSUB_V3_I]] to <1 x i64>
13400 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
13401 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
13402 // CHECK:   [[VQSUB_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
13403 // CHECK:   [[VQSUB_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
13405 // CHECK:   [[VQSUB_V3_I:%.*]] = bitcast <4 x i16> [[VQSUB_V2_I]] to <8 x i8>
13406 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQSUB_V3_I]] to <4 x i16>
13413 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
13414 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
13415 // CHECK:   [[VQSUB_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
13416 // CHECK:   [[VQSUB_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
13418 // CHECK:   [[VQSUB_V3_I:%.*]] = bitcast <2 x i32> [[VQSUB_V2_I]] to <8 x i8>
13419 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQSUB_V3_I]] to <2 x i32>
13426 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
13427 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
13428 // CHECK:   [[VQSUB_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
13429 // CHECK:   [[VQSUB_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
13431 // CHECK:   [[VQSUB_V3_I:%.*]] = bitcast <1 x i64> [[VQSUB_V2_I]] to <8 x i8>
13432 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VQSUB_V3_I]] to <1 x i64>
13446 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
13447 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
13448 // CHECK:   [[VQSUBQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
13449 // CHECK:   [[VQSUBQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
13451 // CHECK:   [[VQSUBQ_V3_I:%.*]] = bitcast <8 x i16> [[VQSUBQ_V2_I]] to <16 x i8>
13452 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQSUBQ_V3_I]] to <8 x i16>
13459 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
13460 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
13461 // CHECK:   [[VQSUBQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
13462 // CHECK:   [[VQSUBQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
13464 // CHECK:   [[VQSUBQ_V3_I:%.*]] = bitcast <4 x i32> [[VQSUBQ_V2_I]] to <16 x i8>
13465 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQSUBQ_V3_I]] to <4 x i32>
13472 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
13473 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
13474 // CHECK:   [[VQSUBQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
13475 // CHECK:   [[VQSUBQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
13477 // CHECK:   [[VQSUBQ_V3_I:%.*]] = bitcast <2 x i64> [[VQSUBQ_V2_I]] to <16 x i8>
13478 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQSUBQ_V3_I]] to <2 x i64>
13492 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
13493 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
13494 // CHECK:   [[VQSUBQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
13495 // CHECK:   [[VQSUBQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
13497 // CHECK:   [[VQSUBQ_V3_I:%.*]] = bitcast <8 x i16> [[VQSUBQ_V2_I]] to <16 x i8>
13498 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQSUBQ_V3_I]] to <8 x i16>
13505 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
13506 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
13507 // CHECK:   [[VQSUBQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
13508 // CHECK:   [[VQSUBQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
13510 // CHECK:   [[VQSUBQ_V3_I:%.*]] = bitcast <4 x i32> [[VQSUBQ_V2_I]] to <16 x i8>
13511 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQSUBQ_V3_I]] to <4 x i32>
13518 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
13519 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
13520 // CHECK:   [[VQSUBQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
13521 // CHECK:   [[VQSUBQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
13523 // CHECK:   [[VQSUBQ_V3_I:%.*]] = bitcast <2 x i64> [[VQSUBQ_V2_I]] to <16 x i8>
13524 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VQSUBQ_V3_I]] to <2 x i64>
13532 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
13533 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
13534 // CHECK:   [[VRADDHN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
13535 // CHECK:   [[VRADDHN_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
13543 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
13544 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
13545 // CHECK:   [[VRADDHN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
13546 // CHECK:   [[VRADDHN_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
13548 // CHECK:   [[VRADDHN_V3_I:%.*]] = bitcast <4 x i16> [[VRADDHN_V2_I]] to <8 x i8>
13549 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRADDHN_V3_I]] to <4 x i16>
13556 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
13557 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
13558 // CHECK:   [[VRADDHN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
13559 // CHECK:   [[VRADDHN_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
13561 // CHECK:   [[VRADDHN_V3_I:%.*]] = bitcast <2 x i32> [[VRADDHN_V2_I]] to <8 x i8>
13562 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRADDHN_V3_I]] to <2 x i32>
13569 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
13570 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
13571 // CHECK:   [[VRADDHN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
13572 // CHECK:   [[VRADDHN_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
13580 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
13581 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
13582 // CHECK:   [[VRADDHN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
13583 // CHECK:   [[VRADDHN_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
13585 // CHECK:   [[VRADDHN_V3_I:%.*]] = bitcast <4 x i16> [[VRADDHN_V2_I]] to <8 x i8>
13586 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRADDHN_V3_I]] to <4 x i16>
13593 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
13594 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
13595 // CHECK:   [[VRADDHN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
13596 // CHECK:   [[VRADDHN_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
13598 // CHECK:   [[VRADDHN_V3_I:%.*]] = bitcast <2 x i32> [[VRADDHN_V2_I]] to <8 x i8>
13599 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRADDHN_V3_I]] to <2 x i32>
13607 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
13608 // CHECK:   [[VRECPE_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
13616 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
13617 // CHECK:   [[VRECPE_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
13625 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
13626 // CHECK:   [[VRECPEQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
13634 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
13635 // CHECK:   [[VRECPEQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
13644 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
13645 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
13646 // CHECK:   [[VRECPS_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
13647 // CHECK:   [[VRECPS_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
13649 // CHECK:   [[VRECPS_V3_I:%.*]] = bitcast <2 x float> [[VRECPS_V2_I]] to <8 x i8>
13650 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRECPS_V3_I]] to <2 x float>
13657 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
13658 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
13659 // CHECK:   [[VRECPSQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
13660 // CHECK:   [[VRECPSQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
13662 // CHECK:   [[VRECPSQ_V3_I:%.*]] = bitcast <4 x float> [[VRECPSQ_V2_I]] to <16 x i8>
13663 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VRECPSQ_V3_I]] to <4 x float>
13671 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
13678 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
13685 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
13698 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
13705 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
13712 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
13719 // CHECK:   [[TMP0:%.*]] = bitcast <4 x half> %a to <8 x i8>
13726 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
13739 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
13746 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <4 x i16>
13753 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <4 x i16>
13760 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <4 x i16>
13767 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <4 x i16>
13780 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <4 x i16>
13787 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <4 x i16>
13794 // CHECK:   [[TMP0:%.*]] = bitcast <4 x half> %a to <4 x i16>
13801 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <4 x i16>
13808 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <4 x i16>
13821 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <2 x i32>
13828 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <2 x i32>
13835 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <2 x i32>
13842 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <2 x i32>
13849 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <2 x i32>
13862 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <2 x i32>
13869 // CHECK:   [[TMP0:%.*]] = bitcast <4 x half> %a to <2 x i32>
13876 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <2 x i32>
13883 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <2 x i32>
13890 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <2 x i32>
13897 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <1 x i64>
13904 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <1 x i64>
13911 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <1 x i64>
13918 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <1 x i64>
13925 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <1 x i64>
13932 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <1 x i64>
13945 // CHECK:   [[TMP0:%.*]] = bitcast <4 x half> %a to <1 x i64>
13952 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <1 x i64>
13959 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <1 x i64>
13966 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <1 x i64>
13979 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
13986 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
13993 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
14000 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
14007 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
14014 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
14021 // CHECK:   [[TMP0:%.*]] = bitcast <4 x half> %a to <8 x i8>
14028 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
14041 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
14048 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <4 x i16>
14061 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <4 x i16>
14068 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <4 x i16>
14075 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <4 x i16>
14082 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <4 x i16>
14089 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <4 x i16>
14096 // CHECK:   [[TMP0:%.*]] = bitcast <4 x half> %a to <4 x i16>
14103 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <4 x i16>
14110 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <4 x i16>
14123 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <2 x i32>
14130 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <2 x i32>
14143 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <2 x i32>
14150 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <2 x i32>
14157 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <2 x i32>
14164 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <2 x i32>
14171 // CHECK:   [[TMP0:%.*]] = bitcast <4 x half> %a to <2 x i32>
14178 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <2 x i32>
14185 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <2 x i32>
14192 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <2 x i32>
14199 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <1 x i64>
14206 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <1 x i64>
14213 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <1 x i64>
14226 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <1 x i64>
14233 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <1 x i64>
14240 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <1 x i64>
14247 // CHECK:   [[TMP0:%.*]] = bitcast <4 x half> %a to <1 x i64>
14254 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <1 x i64>
14261 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <1 x i64>
14268 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <1 x i64>
14275 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <4 x half>
14282 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <4 x half>
14289 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <4 x half>
14296 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <4 x half>
14303 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <4 x half>
14310 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <4 x half>
14317 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <4 x half>
14324 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <4 x half>
14331 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <4 x half>
14338 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <4 x half>
14345 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <4 x half>
14352 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <2 x float>
14359 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <2 x float>
14366 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <2 x float>
14373 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <2 x float>
14380 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <2 x float>
14387 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <2 x float>
14394 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <2 x float>
14401 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <2 x float>
14408 // CHECK:   [[TMP0:%.*]] = bitcast <4 x half> %a to <2 x float>
14415 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <2 x float>
14422 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <2 x float>
14435 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
14442 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
14449 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
14462 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
14469 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
14476 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
14483 // CHECK:   [[TMP0:%.*]] = bitcast <4 x half> %a to <8 x i8>
14490 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
14497 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
14504 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <4 x i16>
14517 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <4 x i16>
14524 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <4 x i16>
14531 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <4 x i16>
14544 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <4 x i16>
14551 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <4 x i16>
14558 // CHECK:   [[TMP0:%.*]] = bitcast <4 x half> %a to <4 x i16>
14565 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <4 x i16>
14572 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i8> %a to <4 x i16>
14579 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
14586 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
14593 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
14606 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
14613 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
14620 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
14627 // CHECK:   [[TMP0:%.*]] = bitcast <8 x half> %a to <16 x i8>
14634 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
14647 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
14654 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <8 x i16>
14661 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <8 x i16>
14668 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <8 x i16>
14675 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <8 x i16>
14688 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <8 x i16>
14695 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <8 x i16>
14702 // CHECK:   [[TMP0:%.*]] = bitcast <8 x half> %a to <8 x i16>
14709 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <8 x i16>
14716 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <8 x i16>
14729 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <4 x i32>
14736 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <4 x i32>
14743 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <4 x i32>
14750 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <4 x i32>
14757 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <4 x i32>
14770 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <4 x i32>
14777 // CHECK:   [[TMP0:%.*]] = bitcast <8 x half> %a to <4 x i32>
14784 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <4 x i32>
14791 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <4 x i32>
14798 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <4 x i32>
14805 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <2 x i64>
14812 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <2 x i64>
14819 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <2 x i64>
14826 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <2 x i64>
14833 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <2 x i64>
14840 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <2 x i64>
14853 // CHECK:   [[TMP0:%.*]] = bitcast <8 x half> %a to <2 x i64>
14860 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <2 x i64>
14867 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <2 x i64>
14874 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <2 x i64>
14887 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
14894 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
14901 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
14908 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
14915 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
14922 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
14929 // CHECK:   [[TMP0:%.*]] = bitcast <8 x half> %a to <16 x i8>
14936 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
14949 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
14956 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <8 x i16>
14969 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <8 x i16>
14976 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <8 x i16>
14983 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <8 x i16>
14990 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <8 x i16>
14997 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <8 x i16>
15004 // CHECK:   [[TMP0:%.*]] = bitcast <8 x half> %a to <8 x i16>
15011 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <8 x i16>
15018 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <8 x i16>
15031 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <4 x i32>
15038 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <4 x i32>
15051 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <4 x i32>
15058 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <4 x i32>
15065 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <4 x i32>
15072 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <4 x i32>
15079 // CHECK:   [[TMP0:%.*]] = bitcast <8 x half> %a to <4 x i32>
15086 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <4 x i32>
15093 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <4 x i32>
15100 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <4 x i32>
15107 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <2 x i64>
15114 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <2 x i64>
15121 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <2 x i64>
15134 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <2 x i64>
15141 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <2 x i64>
15148 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <2 x i64>
15155 // CHECK:   [[TMP0:%.*]] = bitcast <8 x half> %a to <2 x i64>
15162 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <2 x i64>
15169 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <2 x i64>
15176 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <2 x i64>
15183 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <8 x half>
15190 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <8 x half>
15197 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <8 x half>
15204 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <8 x half>
15211 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <8 x half>
15218 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <8 x half>
15225 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <8 x half>
15232 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <8 x half>
15239 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <8 x half>
15246 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <8 x half>
15253 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <8 x half>
15260 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <4 x float>
15267 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <4 x float>
15274 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <4 x float>
15281 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <4 x float>
15288 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <4 x float>
15295 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <4 x float>
15302 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <4 x float>
15309 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <4 x float>
15316 // CHECK:   [[TMP0:%.*]] = bitcast <8 x half> %a to <4 x float>
15323 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <4 x float>
15330 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <4 x float>
15343 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
15350 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
15357 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
15370 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
15377 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
15384 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
15391 // CHECK:   [[TMP0:%.*]] = bitcast <8 x half> %a to <16 x i8>
15398 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
15405 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
15412 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <8 x i16>
15425 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <8 x i16>
15432 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <8 x i16>
15439 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <8 x i16>
15452 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <8 x i16>
15459 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <8 x i16>
15466 // CHECK:   [[TMP0:%.*]] = bitcast <8 x half> %a to <8 x i16>
15473 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <8 x i16>
15480 // CHECK:   [[TMP0:%.*]] = bitcast <16 x i8> %a to <8 x i16>
15750 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
15751 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
15752 // CHECK:   [[VRHADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
15753 // CHECK:   [[VRHADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
15755 // CHECK:   [[VRHADD_V3_I:%.*]] = bitcast <4 x i16> [[VRHADD_V2_I]] to <8 x i8>
15756 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRHADD_V3_I]] to <4 x i16>
15763 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
15764 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
15765 // CHECK:   [[VRHADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
15766 // CHECK:   [[VRHADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
15768 // CHECK:   [[VRHADD_V3_I:%.*]] = bitcast <2 x i32> [[VRHADD_V2_I]] to <8 x i8>
15769 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRHADD_V3_I]] to <2 x i32>
15783 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
15784 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
15785 // CHECK:   [[VRHADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
15786 // CHECK:   [[VRHADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
15788 // CHECK:   [[VRHADD_V3_I:%.*]] = bitcast <4 x i16> [[VRHADD_V2_I]] to <8 x i8>
15789 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRHADD_V3_I]] to <4 x i16>
15796 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
15797 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
15798 // CHECK:   [[VRHADD_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
15799 // CHECK:   [[VRHADD_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
15801 // CHECK:   [[VRHADD_V3_I:%.*]] = bitcast <2 x i32> [[VRHADD_V2_I]] to <8 x i8>
15802 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRHADD_V3_I]] to <2 x i32>
15816 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
15817 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
15818 // CHECK:   [[VRHADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
15819 // CHECK:   [[VRHADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
15821 // CHECK:   [[VRHADDQ_V3_I:%.*]] = bitcast <8 x i16> [[VRHADDQ_V2_I]] to <16 x i8>
15822 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VRHADDQ_V3_I]] to <8 x i16>
15829 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
15830 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
15831 // CHECK:   [[VRHADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
15832 // CHECK:   [[VRHADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
15834 // CHECK:   [[VRHADDQ_V3_I:%.*]] = bitcast <4 x i32> [[VRHADDQ_V2_I]] to <16 x i8>
15835 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VRHADDQ_V3_I]] to <4 x i32>
15849 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
15850 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
15851 // CHECK:   [[VRHADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
15852 // CHECK:   [[VRHADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
15854 // CHECK:   [[VRHADDQ_V3_I:%.*]] = bitcast <8 x i16> [[VRHADDQ_V2_I]] to <16 x i8>
15855 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VRHADDQ_V3_I]] to <8 x i16>
15862 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
15863 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
15864 // CHECK:   [[VRHADDQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
15865 // CHECK:   [[VRHADDQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
15867 // CHECK:   [[VRHADDQ_V3_I:%.*]] = bitcast <4 x i32> [[VRHADDQ_V2_I]] to <16 x i8>
15868 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VRHADDQ_V3_I]] to <4 x i32>
15883 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
15884 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
15885 // CHECK:   [[VRSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
15886 // CHECK:   [[VRSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
15888 // CHECK:   [[VRSHL_V3_I:%.*]] = bitcast <4 x i16> [[VRSHL_V2_I]] to <8 x i8>
15889 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRSHL_V3_I]] to <4 x i16>
15896 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
15897 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
15898 // CHECK:   [[VRSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
15899 // CHECK:   [[VRSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
15901 // CHECK:   [[VRSHL_V3_I:%.*]] = bitcast <2 x i32> [[VRSHL_V2_I]] to <8 x i8>
15902 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRSHL_V3_I]] to <2 x i32>
15909 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
15910 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
15911 // CHECK:   [[VRSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
15912 // CHECK:   [[VRSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
15914 // CHECK:   [[VRSHL_V3_I:%.*]] = bitcast <1 x i64> [[VRSHL_V2_I]] to <8 x i8>
15915 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRSHL_V3_I]] to <1 x i64>
15929 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
15930 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
15931 // CHECK:   [[VRSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
15932 // CHECK:   [[VRSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
15934 // CHECK:   [[VRSHL_V3_I:%.*]] = bitcast <4 x i16> [[VRSHL_V2_I]] to <8 x i8>
15935 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRSHL_V3_I]] to <4 x i16>
15942 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
15943 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
15944 // CHECK:   [[VRSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
15945 // CHECK:   [[VRSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
15947 // CHECK:   [[VRSHL_V3_I:%.*]] = bitcast <2 x i32> [[VRSHL_V2_I]] to <8 x i8>
15948 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRSHL_V3_I]] to <2 x i32>
15955 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
15956 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
15957 // CHECK:   [[VRSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
15958 // CHECK:   [[VRSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
15960 // CHECK:   [[VRSHL_V3_I:%.*]] = bitcast <1 x i64> [[VRSHL_V2_I]] to <8 x i8>
15961 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRSHL_V3_I]] to <1 x i64>
15975 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
15976 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
15977 // CHECK:   [[VRSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
15978 // CHECK:   [[VRSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
15980 // CHECK:   [[VRSHLQ_V3_I:%.*]] = bitcast <8 x i16> [[VRSHLQ_V2_I]] to <16 x i8>
15981 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VRSHLQ_V3_I]] to <8 x i16>
15988 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
15989 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
15990 // CHECK:   [[VRSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
15991 // CHECK:   [[VRSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
15993 // CHECK:   [[VRSHLQ_V3_I:%.*]] = bitcast <4 x i32> [[VRSHLQ_V2_I]] to <16 x i8>
15994 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VRSHLQ_V3_I]] to <4 x i32>
16001 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
16002 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
16003 // CHECK:   [[VRSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
16004 // CHECK:   [[VRSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
16006 // CHECK:   [[VRSHLQ_V3_I:%.*]] = bitcast <2 x i64> [[VRSHLQ_V2_I]] to <16 x i8>
16007 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VRSHLQ_V3_I]] to <2 x i64>
16021 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
16022 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
16023 // CHECK:   [[VRSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16024 // CHECK:   [[VRSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
16026 // CHECK:   [[VRSHLQ_V3_I:%.*]] = bitcast <8 x i16> [[VRSHLQ_V2_I]] to <16 x i8>
16027 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VRSHLQ_V3_I]] to <8 x i16>
16034 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
16035 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
16036 // CHECK:   [[VRSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16037 // CHECK:   [[VRSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
16039 // CHECK:   [[VRSHLQ_V3_I:%.*]] = bitcast <4 x i32> [[VRSHLQ_V2_I]] to <16 x i8>
16040 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VRSHLQ_V3_I]] to <4 x i32>
16047 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
16048 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
16049 // CHECK:   [[VRSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
16050 // CHECK:   [[VRSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
16052 // CHECK:   [[VRSHLQ_V3_I:%.*]] = bitcast <2 x i64> [[VRSHLQ_V2_I]] to <16 x i8>
16053 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VRSHLQ_V3_I]] to <2 x i64>
16061 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
16062 // CHECK:   [[VRSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16070 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
16071 // CHECK:   [[VRSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16079 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
16080 // CHECK:   [[VRSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
16088 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
16089 // CHECK:   [[VRSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16097 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
16098 // CHECK:   [[VRSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16106 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
16107 // CHECK:   [[VRSHRN_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
16123 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
16124 // CHECK:   [[VRSHR_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
16132 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
16133 // CHECK:   [[VRSHR_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
16141 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
16142 // CHECK:   [[VRSHR_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
16157 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
16158 // CHECK:   [[VRSHR_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
16166 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
16167 bitcast <8 x i8> [[TMP0]] to <2 x i32>
16175 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
16176 // CHECK:   [[VRSHR_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
16191 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
16192 // CHECK:   [[VRSHR_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16200 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
16201 // CHECK:   [[VRSHR_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16209 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
16210 // CHECK:   [[VRSHR_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
16225 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
16226 // CHECK:   [[VRSHR_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16234 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
16235 // CHECK:   [[VRSHR_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16243 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
16244 // CHECK:   [[VRSHR_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
16253 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
16254 // CHECK:   [[VRSQRTE_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
16262 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
16263 // CHECK:   [[VRSQRTE_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
16271 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
16272 // CHECK:   [[VRSQRTEQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
16280 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
16281 // CHECK:   [[VRSQRTEQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16290 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %a to <8 x i8>
16291 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
16292 // CHECK:   [[VRSQRTS_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
16293 // CHECK:   [[VRSQRTS_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
16295 // CHECK:   [[VRSQRTS_V3_I:%.*]] = bitcast <2 x float> [[VRSQRTS_V2_I]] to <8 x i8>
16296 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRSQRTS_V3_I]] to <2 x float>
16303 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %a to <16 x i8>
16304 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
16305 // CHECK:   [[VRSQRTSQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
16306 // CHECK:   [[VRSQRTSQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
16308 // CHECK:   [[VRSQRTSQ_V3_I:%.*]] = bitcast <4 x float> [[VRSQRTSQ_V2_I]] to <16 x i8>
16309 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VRSQRTSQ_V3_I]] to <4 x float>
16325 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
16326 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
16327 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
16328 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
16337 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
16338 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
16339 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
16340 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
16349 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
16350 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
16351 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
16352 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
16369 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
16370 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
16371 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
16372 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
16381 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
16382 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
16383 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
16384 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
16393 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
16394 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
16395 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
16396 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
16413 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
16414 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
16415 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16416 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
16425 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
16426 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
16427 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16428 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
16437 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
16438 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
16439 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
16440 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
16457 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
16458 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
16459 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16460 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
16469 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
16470 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
16471 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16472 // CHECK:   [[TMP3:%.*]] = bitcast
16481 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
16482 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
16483 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
16484 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
16494 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
16495 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
16496 // CHECK:   [[VRSUBHN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16497 // CHECK:   [[VRSUBHN_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
16505 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
16506 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
16507 // CHECK:   [[VRSUBHN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16508 // CHECK:   [[VRSUBHN_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
16510 // CHECK:   [[VRSUBHN_V3_I:%.*]] = bitcast <4 x i16> [[VRSUBHN_V2_I]] to <8 x i8>
16511 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRSUBHN_V3_I]] to <4 x i16>
16518 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
16519 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
16520 // CHECK:   [[VRSUBHN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
16521 // CHECK:   [[VRSUBHN_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
16523 // CHECK:   [[VRSUBHN_V3_I:%.*]] = bitcast <2 x i32> [[VRSUBHN_V2_I]] to <8 x i8>
16524 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRSUBHN_V3_I]] to <2 x i32>
16531 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
16532 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
16533 // CHECK:   [[VRSUBHN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16534 // CHECK:   [[VRSUBHN_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
16542 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
16543 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
16544 // CHECK:   [[VRSUBHN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16545 // CHECK:   [[VRSUBHN_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
16547 // CHECK:   [[VRSUBHN_V3_I:%.*]] = bitcast <4 x i16> [[VRSUBHN_V2_I]] to <8 x i8>
16548 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRSUBHN_V3_I]] to <4 x i16>
16555 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
16556 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
16557 // CHECK:   [[VRSUBHN_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
16558 // CHECK:   [[VRSUBHN_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
16560 // CHECK:   [[VRSUBHN_V3_I:%.*]] = bitcast <2 x i32> [[VRSUBHN_V2_I]] to <8 x i8>
16561 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VRSUBHN_V3_I]] to <2 x i32>
16576 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
16577 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
16585 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
16586 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
16601 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
16602 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
16610 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
16611 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
16626 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
16627 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
16635 // CHECK:   [[TMP0:%.*]] = bitcast <2 x float> %b to <8 x i8>
16636 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x float>
16650 // CHECK:   [[TMP1:%.*]] = bitcast half* [[__REINT_246]] to i16*
16652 // CHECK:   [[TMP3:%.*]] = bitcast <4 x half>* [[__REINT1_246]] to <4 x i16>*
16654 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
16655 // CHECK:   [[TMP6:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
16658 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16>* [[__REINT2_246]] to <4 x half>*
16673 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %b to <16 x i8>
16674 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16682 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %b to <16 x i8>
16683 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16698 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %b to <16 x i8>
16699 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16707 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %b to <16 x i8>
16708 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16723 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %b to <16 x i8>
16724 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16732 // CHECK:   [[TMP0:%.*]] = bitcast <4 x float> %b to <16 x i8>
16733 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x float>
16747 // CHECK:   [[TMP1:%.*]] = bitcast half* [[__REINT_248]] to i16*
16749 // CHECK:   [[TMP3:%.*]] = bitcast <8 x half>* [[__REINT1_248]] to <8 x i16>*
16751 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
16752 // CHECK:   [[TMP6:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
16755 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16>* [[__REINT2_248]] to <8 x half>*
16764 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %b to <8 x i8>
16765 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
16774 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %b to <8 x i8>
16775 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
16783 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %b to <16 x i8>
16784 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
16792 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %b to <16 x i8>
16793 bitcast <16 x i8> [[TMP0]] to <2 x i64>
16809 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
16810 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
16811 // CHECK:   [[VSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
16812 // CHECK:   [[VSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
16814 // CHECK:   [[VSHL_V3_I:%.*]] = bitcast <4 x i16> [[VSHL_V2_I]] to <8 x i8>
16815 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VSHL_V3_I]] to <4 x i16>
16822 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
16823 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
16824 // CHECK:   [[VSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
16825 // CHECK:   [[VSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
16827 // CHECK:   [[VSHL_V3_I:%.*]] = bitcast <2 x i32> [[VSHL_V2_I]] to <8 x i8>
16828 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VSHL_V3_I]] to <2 x i32>
16835 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
16836 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
16837 // CHECK:   [[VSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
16838 // CHECK:   [[VSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
16840 // CHECK:   [[VSHL_V3_I:%.*]] = bitcast <1 x i64> [[VSHL_V2_I]] to <8 x i8>
16841 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VSHL_V3_I]] to <1 x i64>
16855 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
16856 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
16857 // CHECK:   [[VSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
16858 // CHECK:   [[VSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
16860 // CHECK:   [[VSHL_V3_I:%.*]] = bitcast <4 x i16> [[VSHL_V2_I]] to <8 x i8>
16861 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VSHL_V3_I]] to <4 x i16>
16868 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
16869 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
16870 // CHECK:   [[VSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
16871 // CHECK:   [[VSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
16873 // CHECK:   [[VSHL_V3_I:%.*]] = bitcast <2 x i32> [[VSHL_V2_I]] to <8 x i8>
16874 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VSHL_V3_I]] to <2 x i32>
16881 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
16882 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
16883 // CHECK:   [[VSHL_V_I:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
16884 // CHECK:   [[VSHL_V1_I:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
16886 // CHECK:   [[VSHL_V3_I:%.*]] = bitcast <1 x i64> [[VSHL_V2_I]] to <8 x i8>
16887 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[VSHL_V3_I]] to <1 x i64>
16901 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
16902 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
16903 // CHECK:   [[VSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16904 // CHECK:   [[VSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
16906 // CHECK:   [[VSHLQ_V3_I:%.*]] = bitcast <8 x i16> [[VSHLQ_V2_I]] to <16 x i8>
16907 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VSHLQ_V3_I]] to <8 x i16>
16914 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
16915 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
16916 // CHECK:   [[VSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16917 // CHECK:   [[VSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
16919 // CHECK:   [[VSHLQ_V3_I:%.*]] = bitcast <4 x i32> [[VSHLQ_V2_I]] to <16 x i8>
16920 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VSHLQ_V3_I]] to <4 x i32>
16927 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
16928 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
16929 // CHECK:   [[VSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
16930 // CHECK:   [[VSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
16932 // CHECK:   [[VSHLQ_V3_I:%.*]] = bitcast <2 x i64> [[VSHLQ_V2_I]] to <16 x i8>
16933 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VSHLQ_V3_I]] to <2 x i64>
16947 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
16948 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
16949 // CHECK:   [[VSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
16950 // CHECK:   [[VSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
16952 // CHECK:   [[VSHLQ_V3_I:%.*]] = bitcast <8 x i16> [[VSHLQ_V2_I]] to <16 x i8>
16953 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VSHLQ_V3_I]] to <8 x i16>
16960 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
16961 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
16962 // CHECK:   [[VSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
16963 // CHECK:   [[VSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
16965 // CHECK:   [[VSHLQ_V3_I:%.*]] = bitcast <4 x i32> [[VSHLQ_V2_I]] to <16 x i8>
16966 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VSHLQ_V3_I]] to <4 x i32>
16973 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
16974 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
16975 // CHECK:   [[VSHLQ_V_I:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
16976 // CHECK:   [[VSHLQ_V1_I:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
16978 // CHECK:   [[VSHLQ_V3_I:%.*]] = bitcast <2 x i64> [[VSHLQ_V2_I]] to <16 x i8>
16979 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[VSHLQ_V3_I]] to <2 x i64>
16995 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
16996 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17005 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
17006 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
17023 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
17024 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17033 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
17034 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
17051 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
17052 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17060 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
17061 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
17069 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
17070 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
17085 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
17086 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17094 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
17095 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
17103 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
17104 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
17119 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17120 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17128 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
17129 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
17137 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
17138 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
17153 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17154 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17162 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
17163 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
17171 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
17172 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
17181 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17182 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17191 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
17192 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
17201 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
17202 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
17211 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17212 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17221 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
17222 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
17231 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
17232 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
17249 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
17250 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17258 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
17259 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
17267 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
17268 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
17283 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
17284 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17292 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
17293 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
17301 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
17302 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
17317 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17318 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17326 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
17327 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
17335 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
17336 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
17351 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17352 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17360 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
17361 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
17369 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
17370 // CHECK:   [[TMP1:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
17386 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
17387 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
17388 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17389 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
17397 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
17398 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
17399 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
17400 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
17408 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
17409 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
17410 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
17411 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
17426 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
17427 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
17428 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17429 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
17437 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
17438 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
17439 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
17440 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
17448 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
17449 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
17450 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
17451 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
17466 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
17467 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
17468 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17469 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
17484 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17485 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
17486 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17487 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
17495 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
17496 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
17497 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
17498 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
17506 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
17507 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
17508 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
17509 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
17524 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17525 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
17526 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17527 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
17535 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
17536 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
17537 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
17538 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
17546 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
17547 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
17548 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
17549 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
17564 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17565 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
17566 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17567 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
17584 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
17585 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
17586 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17587 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
17596 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
17597 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
17598 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
17599 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
17608 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
17609 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
17610 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
17611 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
17628 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
17629 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
17630 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17631 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
17640 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
17641 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
17642 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
17643 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
17652 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
17653 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
17654 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
17655 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
17672 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17673 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
17674 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17675 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
17684 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
17685 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
17686 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
17687 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
17696 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
17697 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
17698 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
17699 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
17716 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17717 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
17718 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17719 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
17728 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
17729 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
17730 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
17731 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
17740 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
17741 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
17742 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
17743 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
17760 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
17761 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
17762 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17763 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
17771 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
17772 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
17773 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
17774 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
17782 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
17783 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
17784 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
17785 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
17800 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
17801 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
17802 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17803 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
17811 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
17812 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
17813 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
17814 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
17822 // CHECK:   [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
17823 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
17824 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
17825 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
17840 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
17841 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
17842 // CHECK:   [[VSLI_N:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
17843 // CHECK:   [[VSLI_N1:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
17858 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17859 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
17860 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17861 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
17869 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
17870 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
17871 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
17872 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
17880 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
17881 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
17882 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
17883 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
17898 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17899 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
17900 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17901 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
17909 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
17910 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
17911 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
17912 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
17920 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
17921 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
17922 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
17923 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
17938 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
17939 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
17940 // CHECK:   [[VSLI_N:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
17941 // CHECK:   [[VSLI_N1:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
17957 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
17958 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
17959 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
17967 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
17968 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
17969 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
17977 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
17978 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
17979 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
17994 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
17995 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
17996 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
18004 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
18005 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
18006 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
18014 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
18015 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
18016 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
18024 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
18025 // CHECK:   [[TMP1:%.*]] = bitcast <8 x half> %b to <16 x i8>
18026 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
18034 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
18035 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
18036 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
18051 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
18052 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
18053 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
18068 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
18069 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
18070 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
18078 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
18079 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
18080 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
18088 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
18089 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
18090 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
18105 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
18106 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
18107 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
18115 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
18116 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
18117 // CHECK:   [[TMP2:%.*]] = bitcast
18125 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
18126 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
18127 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
18135 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
18136 // CHECK:   [[TMP1:%.*]] = bitcast <4 x half> %b to <8 x i8>
18137 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
18145 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
18146 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
18147 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
18162 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
18163 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
18164 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
18181 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
18182 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
18183 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
18185 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i16*
18193 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
18194 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
18195 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
18197 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i32*
18205 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
18206 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
18207 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
18224 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
18225 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
18226 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
18228 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i16*
18236 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
18237 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
18238 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
18240 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i32*
18248 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
18249 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
18250 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
18259 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
18260 // CHECK:   [[TMP1:%.*]] = bitcast <8 x half> %b to <16 x i8>
18261 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
18263 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i16*
18271 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
18272 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %b to <16 x i8>
18273 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
18275 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to float*
18291 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
18292 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
18293 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
18295 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i16*
18311 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
18312 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
18313 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
18315 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i16*
18323 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
18324 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
18325 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
18327 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i32*
18335 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
18336 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
18337 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
18339 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i64*
18355 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
18356 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
18357 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
18359 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i16*
18367 // CHECK:   [[TMP0:%.*]] = bitcast i32* %a to i8*
18368 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
18369 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
18371 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i32*
18379 // CHECK:   [[TMP0:%.*]] = bitcast i64* %a to i8*
18380 // CHECK:   [[TMP1:%.*]] = bitcast <1 x i64> %b to <8 x i8>
18381 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x i64>
18383 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i64*
18391 // CHECK:   [[TMP0:%.*]] = bitcast half* %a to i8*
18392 // CHECK:   [[TMP1:%.*]] = bitcast <4 x half> %b to <8 x i8>
18393 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
18395 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i16*
18403 // CHECK:   [[TMP0:%.*]] = bitcast float* %a to i8*
18404 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %b to <8 x i8>
18405 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
18407 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to float*
18423 // CHECK:   [[TMP0:%.*]] = bitcast i16* %a to i8*
18424 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
18425 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
18427 // CHECK:   [[TMP4:%.*]] = bitcast i8* [[TMP0]] to i16*
18439 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <16 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
18441 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint8x16x2_t* [[__S1]] to i8*
18442 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x16x2_t* [[B]] to i8*
18460 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
18462 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x8x2_t* [[__S1]] to i8*
18463 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x8x2_t* [[B]] to i8*
18465 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
18469 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
18473 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
18474 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
18475 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
18486 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i32>]* [[COERCE_DIVE]] to [4 x i64]*
18488 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x4x2_t* [[__S1]] to i8*
18489 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x4x2_t* [[B]] to i8*
18491 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
18495 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i32> [[TMP4]] to <16 x i8>
18499 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i32> [[TMP6]] to <16 x i8>
18500 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x i32>
18501 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x i32>
18512 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <16 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
18514 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int8x16x2_t* [[__S1]] to i8*
18515 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x16x2_t* [[B]] to i8*
18533 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
18535 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x8x2_t* [[__S1]] to i8*
18536 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x8x2_t* [[B]] to i8*
18538 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
18542 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
18546 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
18547 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
18548 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
18559 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i32>]* [[COERCE_DIVE]] to [4 x i64]*
18561 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x4x2_t* [[__S1]] to i8*
18562 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x4x2_t* [[B]] to i8*
18564 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
18568 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i32> [[TMP4]] to <16 x i8>
18572 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i32> [[TMP6]] to <16 x i8>
18573 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x i32>
18574 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x i32>
18585 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x half>]* [[COERCE_DIVE]] to [4 x i64]*
18587 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x8x2_t* [[__S1]] to i8*
18588 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x8x2_t* [[B]] to i8*
18590 // CHECK:   [[TMP3:%.*]] = bitcast half* %a to i8*
18594 // CHECK:   [[TMP5:%.*]] = bitcast <8 x half> [[TMP4]] to <16 x i8>
18598 // CHECK:   [[TMP7:%.*]] = bitcast <8 x half> [[TMP6]] to <16 x i8>
18599 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
18600 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
18611 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x float>]* [[COERCE_DIVE]] to [4 x i64]*
18613 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x4x2_t* [[__S1]] to i8*
18614 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x4x2_t* [[B]] to i8*
18616 // CHECK:   [[TMP3:%.*]] = bitcast float* %a to i8*
18620 // CHECK:   [[TMP5:%.*]] = bitcast <4 x float> [[TMP4]] to <16 x i8>
18624 // CHECK:   [[TMP7:%.*]] = bitcast <4 x float> [[TMP6]] to <16 x i8>
18625 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x float>
18626 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x float>
18637 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <16 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
18639 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly8x16x2_t* [[__S1]] to i8*
18640 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x16x2_t* [[B]] to i8*
18658 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
18660 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x8x2_t* [[__S1]] to i8*
18661 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x8x2_t* [[B]] to i8*
18663 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
18667 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
18671 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
18672 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
18673 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
18684 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
18686 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint8x8x2_t* [[__S1]] to i8*
18687 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x8x2_t* [[B]] to i8*
18705 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i16>]* [[COERCE_DIVE]] to [2 x i64]*
18707 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x4x2_t* [[__S1]] to i8*
18708 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x4x2_t* [[B]] to i8*
18710 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
18714 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
18718 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
18719 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
18720 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
18731 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <2 x i32>]* [[COERCE_DIVE]] to [2 x i64]*
18733 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x2x2_t* [[__S1]] to i8*
18734 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x2x2_t* [[B]] to i8*
18736 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
18740 // CHECK:   [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <8 x i8>
18744 // CHECK:   [[TMP7:%.*]] = bitcast <2 x i32> [[TMP6]] to <8 x i8>
18745 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x i32>
18746 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x i32>
18757 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <1 x i64>]* [[COERCE_DIVE]] to [2 x i64]*
18759 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint64x1x2_t* [[__S1]] to i8*
18760 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint64x1x2_t* [[B]] to i8*
18762 // CHECK:   [[TMP3:%.*]] = bitcast i64* %a to i8*
18766 // CHECK:   [[TMP5:%.*]] = bitcast <1 x i64> [[TMP4]] to <8 x i8>
18770 // CHECK:   [[TMP7:%.*]] = bitcast <1 x i64> [[TMP6]] to <8 x i8>
18771 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <1 x i64>
18772 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <1 x i64>
18783 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
18785 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int8x8x2_t* [[__S1]] to i8*
18786 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x8x2_t* [[B]] to i8*
18804 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i16>]* [[COERCE_DIVE]] to [2 x i64]*
18806 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x4x2_t* [[__S1]] to i8*
18807 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x4x2_t* [[B]] to i8*
18809 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
18813 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
18817 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
18818 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
18819 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
18830 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <2 x i32>]* [[COERCE_DIVE]] to [2 x i64]*
18832 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x2x2_t* [[__S1]] to i8*
18833 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x2x2_t* [[B]] to i8*
18835 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
18839 // CHECK:   [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <8 x i8>
18843 // CHECK:   [[TMP7:%.*]] = bitcast <2 x i32> [[TMP6]] to <8 x i8>
18844 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x i32>
18845 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x i32>
18856 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <1 x i64>]* [[COERCE_DIVE]] to [2 x i64]*
18858 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int64x1x2_t* [[__S1]] to i8*
18859 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int64x1x2_t* [[B]] to i8*
18861 // CHECK:   [[TMP3:%.*]] = bitcast i64* %a to i8*
18865 // CHECK:   [[TMP5:%.*]] = bitcast <1 x i64> [[TMP4]] to <8 x i8>
18869 // CHECK:   [[TMP7:%.*]] = bitcast <1 x i64> [[TMP6]] to <8 x i8>
18870 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <1 x i64>
18871 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <1 x i64>
18882 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x half>]* [[COERCE_DIVE]] to [2 x i64]*
18884 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x4x2_t* [[__S1]] to i8*
18885 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x4x2_t* [[B]] to i8*
18887 // CHECK:   [[TMP3:%.*]] = bitcast half* %a to i8*
18891 // CHECK:   [[TMP5:%.*]] = bitcast <4 x half> [[TMP4]] to <8 x i8>
18895 // CHECK:   [[TMP7:%.*]] = bitcast <4 x half> [[TMP6]] to <8 x i8>
18896 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
18897 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
18908 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <2 x float>]* [[COERCE_DIVE]] to [2 x i64]*
18910 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x2x2_t* [[__S1]] to i8*
18911 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x2x2_t* [[B]] to i8*
18913 // CHECK:   [[TMP3:%.*]] = bitcast float* %a to i8*
18917 // CHECK:   [[TMP5:%.*]] = bitcast <2 x float> [[TMP4]] to <8 x i8>
18921 // CHECK:   [[TMP7:%.*]] = bitcast <2 x float> [[TMP6]] to <8 x i8>
18922 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x float>
18923 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x float>
18934 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
18936 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly8x8x2_t* [[__S1]] to i8*
18937 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x8x2_t* [[B]] to i8*
18955 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i16>]* [[COERCE_DIVE]] to [2 x i64]*
18957 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x4x2_t* [[__S1]] to i8*
18958 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x4x2_t* [[B]] to i8*
18960 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
18964 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
18968 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
18969 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
18970 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
18982 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
18984 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x8x2_t* [[__S1]] to i8*
18985 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x8x2_t* [[B]] to i8*
18987 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
18991 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
18995 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
18996 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
18997 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
19008 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i32>]* [[COERCE_DIVE]] to [4 x i64]*
19010 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x4x2_t* [[__S1]] to i8*
19011 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x4x2_t* [[B]] to i8*
19013 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
19017 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i32> [[TMP4]] to <16 x i8>
19021 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i32> [[TMP6]] to <16 x i8>
19022 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x i32>
19023 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x i32>
19034 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
19036 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x8x2_t* [[__S1]] to i8*
19037 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x8x2_t* [[B]] to i8*
19039 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
19043 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
19047 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
19048 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
19049 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
19060 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i32>]* [[COERCE_DIVE]] to [4 x i64]*
19062 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x4x2_t* [[__S1]] to i8*
19063 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x4x2_t* [[B]] to i8*
19065 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
19069 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i32> [[TMP4]] to <16 x i8>
19073 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i32> [[TMP6]] to <16 x i8>
19074 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x i32>
19075 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x i32>
19086 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x half>]* [[COERCE_DIVE]] to [4 x i64]*
19088 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x8x2_t* [[__S1]] to i8*
19089 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x8x2_t* [[B]] to i8*
19091 // CHECK:   [[TMP3:%.*]] = bitcast half* %a to i8*
19095 // CHECK:   [[TMP5:%.*]] = bitcast <8 x half> [[TMP4]] to <16 x i8>
19099 // CHECK:   [[TMP7:%.*]] = bitcast <8 x half> [[TMP6]] to <16 x i8>
19100 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
19101 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
19112 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x float>]* [[COERCE_DIVE]] to [4 x i64]*
19114 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x4x2_t* [[__S1]] to i8*
19115 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x4x2_t* [[B]] to i8*
19117 // CHECK:   [[TMP3:%.*]] = bitcast float* %a to i8*
19121 // CHECK:   [[TMP5:%.*]] = bitcast <4 x float> [[TMP4]] to <16 x i8>
19125 // CHECK:   [[TMP7:%.*]] = bitcast <4 x float> [[TMP6]] to <16 x i8>
19126 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x float>
19127 // CHECK:   [[TMP9:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x float>
19138 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
19140 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x8x2_t* [[__S1]] to i8*
19141 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x8x2_t* [[B]] to i8*
19143 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
19147 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
19151 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
19152 // CHECK:   [[TMP8:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
19153 bitcast <16 x i8> [[TMP7]] to <8 x i16>
19164 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
19166 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint8x8x2_t* [[__S1]] to i8*
19167 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x8x2_t* [[B]] to i8*
19185 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i16>]* [[COERCE_DIVE]] to [2 x i64]*
19187 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x4x2_t* [[__S1]] to i8*
19188 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x4x2_t* [[B]] to i8*
19190 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
19194 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
19198 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
19199 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
19200 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
19211 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <2 x i32>]* [[COERCE_DIVE]] to [2 x i64]*
19213 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x2x2_t* [[__S1]] to i8*
19214 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x2x2_t* [[B]] to i8*
19216 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
19220 // CHECK:   [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <8 x i8>
19224 // CHECK:   [[TMP7:%.*]] = bitcast <2 x i32> [[TMP6]] to <8 x i8>
19225 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x i32>
19226 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x i32>
19237 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
19239 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int8x8x2_t* [[__S1]] to i8*
19240 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x8x2_t* [[B]] to i8*
19258 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i16>]* [[COERCE_DIVE]] to [2 x i64]*
19260 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x4x2_t* [[__S1]] to i8*
19261 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x4x2_t* [[B]] to i8*
19263 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
19267 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
19271 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
19272 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
19273 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
19284 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <2 x i32>]* [[COERCE_DIVE]] to [2 x i64]*
19286 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x2x2_t* [[__S1]] to i8*
19287 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x2x2_t* [[B]] to i8*
19289 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
19293 // CHECK:   [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <8 x i8>
19297 // CHECK:   [[TMP7:%.*]] = bitcast <2 x i32> [[TMP6]] to <8 x i8>
19298 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x i32>
19299 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x i32>
19310 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x half>]* [[COERCE_DIVE]] to [2 x i64]*
19312 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x4x2_t* [[__S1]] to i8*
19313 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x4x2_t* [[B]] to i8*
19315 // CHECK:   [[TMP3:%.*]] = bitcast half* %a to i8*
19319 // CHECK:   [[TMP5:%.*]] = bitcast <4 x half> [[TMP4]] to <8 x i8>
19323 // CHECK:   [[TMP7:%.*]] = bitcast <4 x half> [[TMP6]] to <8 x i8>
19324 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
19325 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
19336 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <2 x float>]* [[COERCE_DIVE]] to [2 x i64]*
19338 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x2x2_t* [[__S1]] to i8*
19339 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x2x2_t* [[B]] to i8*
19341 // CHECK:   [[TMP3:%.*]] = bitcast float* %a to i8*
19345 // CHECK:   [[TMP5:%.*]] = bitcast <2 x float> [[TMP4]] to <8 x i8>
19349 // CHECK:   [[TMP7:%.*]] = bitcast <2 x float> [[TMP6]] to <8 x i8>
19350 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x float>
19351 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x float>
19362 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
19364 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly8x8x2_t* [[__S1]] to i8*
19365 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x8x2_t* [[B]] to i8*
19383 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <4 x i16>]* [[COERCE_DIVE]] to [2 x i64]*
19385 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x4x2_t* [[__S1]] to i8*
19386 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x4x2_t* [[B]] to i8*
19388 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
19392 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
19396 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
19397 // CHECK:   [[TMP8:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
19398 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
19410 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <16 x i8>]* [[COERCE_DIVE]] to [6 x i64]*
19412 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint8x16x3_t* [[__S1]] to i8*
19413 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x16x3_t* [[B]] to i8*
19434 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i16>]* [[COERCE_DIVE]] to [6 x i64]*
19436 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x8x3_t* [[__S1]] to i8*
19437 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x8x3_t* [[B]] to i8*
19439 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
19443 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
19447 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
19451 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i16> [[TMP8]] to <16 x i8>
19452 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
19453 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
19454 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
19465 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i32>]* [[COERCE_DIVE]] to [6 x i64]*
19467 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x4x3_t* [[__S1]] to i8*
19468 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x4x3_t* [[B]] to i8*
19470 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
19474 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i32> [[TMP4]] to <16 x i8>
19478 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i32> [[TMP6]] to <16 x i8>
19482 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i32> [[TMP8]] to <16 x i8>
19483 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x i32>
19484 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x i32>
19485 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <4 x i32>
19496 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <16 x i8>]* [[COERCE_DIVE]] to [6 x i64]*
19498 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int8x16x3_t* [[__S1]] to i8*
19499 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x16x3_t* [[B]] to i8*
19520 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i16>]* [[COERCE_DIVE]] to [6 x i64]*
19522 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x8x3_t* [[__S1]] to i8*
19523 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x8x3_t* [[B]] to i8*
19525 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
19529 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
19533 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
19537 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i16> [[TMP8]] to <16 x i8>
19538 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
19539 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
19540 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
19551 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i32>]* [[COERCE_DIVE]] to [6 x i64]*
19553 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x4x3_t* [[__S1]] to i8*
19554 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x4x3_t* [[B]] to i8*
19556 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
19560 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i32> [[TMP4]] to <16 x i8>
19564 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i32> [[TMP6]] to <16 x i8>
19568 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i32> [[TMP8]] to <16 x i8>
19569 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x i32>
19570 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x i32>
19571 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <4 x i32>
19582 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x half>]* [[COERCE_DIVE]] to [6 x i64]*
19584 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x8x3_t* [[__S1]] to i8*
19585 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x8x3_t* [[B]] to i8*
19587 // CHECK:   [[TMP3:%.*]] = bitcast half* %a to i8*
19591 // CHECK:   [[TMP5:%.*]] = bitcast <8 x half> [[TMP4]] to <16 x i8>
19595 // CHECK:   [[TMP7:%.*]] = bitcast <8 x half> [[TMP6]] to <16 x i8>
19599 // CHECK:   [[TMP9:%.*]] = bitcast <8 x half> [[TMP8]] to <16 x i8>
19600 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
19601 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
19602 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
19613 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x float>]* [[COERCE_DIVE]] to [6 x i64]*
19615 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x4x3_t* [[__S1]] to i8*
19616 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x4x3_t* [[B]] to i8*
19618 // CHECK:   [[TMP3:%.*]] = bitcast float* %a to i8*
19622 // CHECK:   [[TMP5:%.*]] = bitcast <4 x float> [[TMP4]] to <16 x i8>
19626 // CHECK:   [[TMP7:%.*]] = bitcast <4 x float> [[TMP6]] to <16 x i8>
19630 // CHECK:   [[TMP9:%.*]] = bitcast <4 x float> [[TMP8]] to <16 x i8>
19631 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x float>
19632 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x float>
19633 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <4 x float>
19644 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <16 x i8>]* [[COERCE_DIVE]] to [6 x i64]*
19646 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly8x16x3_t* [[__S1]] to i8*
19647 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x16x3_t* [[B]] to i8*
19668 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i16>]* [[COERCE_DIVE]] to [6 x i64]*
19670 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x8x3_t* [[__S1]] to i8*
19671 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x8x3_t* [[B]] to i8*
19673 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
19677 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
19681 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
19685 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i16> [[TMP8]] to <16 x i8>
19686 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
19687 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
19688 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
19699 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
19701 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint8x8x3_t* [[__S1]] to i8*
19702 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x8x3_t* [[B]] to i8*
19723 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i16>]* [[COERCE_DIVE]] to [3 x i64]*
19725 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x4x3_t* [[__S1]] to i8*
19726 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x4x3_t* [[B]] to i8*
19728 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
19732 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
19736 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
19740 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i16> [[TMP8]] to <8 x i8>
19741 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
19742 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
19743 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
19754 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <2 x i32>]* [[COERCE_DIVE]] to [3 x i64]*
19756 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x2x3_t* [[__S1]] to i8*
19757 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x2x3_t* [[B]] to i8*
19759 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
19763 // CHECK:   [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <8 x i8>
19767 // CHECK:   [[TMP7:%.*]] = bitcast <2 x i32> [[TMP6]] to <8 x i8>
19771 // CHECK:   [[TMP9:%.*]] = bitcast <2 x i32> [[TMP8]] to <8 x i8>
19772 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x i32>
19773 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x i32>
19774 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <2 x i32>
19785 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <1 x i64>]* [[COERCE_DIVE]] to [3 x i64]*
19787 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint64x1x3_t* [[__S1]] to i8*
19788 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint64x1x3_t* [[B]] to i8*
19790 // CHECK:   [[TMP3:%.*]] = bitcast i64* %a to i8*
19794 // CHECK:   [[TMP5:%.*]] = bitcast <1 x i64> [[TMP4]] to <8 x i8>
19798 // CHECK:   [[TMP7:%.*]] = bitcast <1 x i64> [[TMP6]] to <8 x i8>
19802 // CHECK:   [[TMP9:%.*]] = bitcast <1 x i64> [[TMP8]] to <8 x i8>
19803 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <1 x i64>
19804 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <1 x i64>
19805 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <1 x i64>
19816 // CHECK:   [[TMP0:%.*]] = bitcast
19818 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int8x8x3_t* [[__S1]] to i8*
19819 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x8x3_t* [[B]] to i8*
19840 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i16>]* [[COERCE_DIVE]] to [3 x i64]*
19842 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x4x3_t* [[__S1]] to i8*
19843 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x4x3_t* [[B]] to i8*
19845 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
19849 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
19853 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
19857 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i16> [[TMP8]] to <8 x i8>
19858 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
19859 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
19860 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
19871 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <2 x i32>]* [[COERCE_DIVE]] to [3 x i64]*
19873 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x2x3_t* [[__S1]] to i8*
19874 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x2x3_t* [[B]] to i8*
19876 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
19880 // CHECK:   [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <8 x i8>
19884 // CHECK:   [[TMP7:%.*]] = bitcast <2 x i32> [[TMP6]] to <8 x i8>
19888 // CHECK:   [[TMP9:%.*]] = bitcast <2 x i32> [[TMP8]] to <8 x i8>
19889 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x i32>
19890 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x i32>
19891 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <2 x i32>
19902 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <1 x i64>]* [[COERCE_DIVE]] to [3 x i64]*
19904 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int64x1x3_t* [[__S1]] to i8*
19905 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int64x1x3_t* [[B]] to i8*
19907 // CHECK:   [[TMP3:%.*]] = bitcast i64* %a to i8*
19911 // CHECK:   [[TMP5:%.*]] = bitcast <1 x i64> [[TMP4]] to <8 x i8>
19915 // CHECK:   [[TMP7:%.*]] = bitcast <1 x i64> [[TMP6]] to <8 x i8>
19919 // CHECK:   [[TMP9:%.*]] = bitcast <1 x i64> [[TMP8]] to <8 x i8>
19920 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <1 x i64>
19921 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <1 x i64>
19922 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <1 x i64>
19933 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x half>]* [[COERCE_DIVE]] to [3 x i64]*
19935 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x4x3_t* [[__S1]] to i8*
19936 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x4x3_t* [[B]] to i8*
19938 // CHECK:   [[TMP3:%.*]] = bitcast half* %a to i8*
19942 // CHECK:   [[TMP5:%.*]] = bitcast <4 x half> [[TMP4]] to <8 x i8>
19946 // CHECK:   [[TMP7:%.*]] = bitcast <4 x half> [[TMP6]] to <8 x i8>
19950 // CHECK:   [[TMP9:%.*]] = bitcast <4 x half> [[TMP8]] to <8 x i8>
19951 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
19952 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
19953 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
19964 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <2 x float>]* [[COERCE_DIVE]] to [3 x i64]*
19966 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x2x3_t* [[__S1]] to i8*
19967 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x2x3_t* [[B]] to i8*
19969 // CHECK:   [[TMP3:%.*]] = bitcast float* %a to i8*
19973 // CHECK:   [[TMP5:%.*]] = bitcast <2 x float> [[TMP4]] to <8 x i8>
19977 // CHECK:   [[TMP7:%.*]] = bitcast <2 x float> [[TMP6]] to <8 x i8>
19981 // CHECK:   [[TMP9:%.*]] = bitcast <2 x float> [[TMP8]] to <8 x i8>
19982 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x float>
19983 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x float>
19984 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <2 x float>
19995 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
19997 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly8x8x3_t* [[__S1]] to i8*
19998 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x8x3_t* [[B]] to i8*
20019 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i16>]* [[COERCE_DIVE]] to [3 x i64]*
20021 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x4x3_t* [[__S1]] to i8*
20022 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x4x3_t* [[B]] to i8*
20024 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
20028 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
20032 // CHECK:   [[TMP7:%.*]] = bitcast
20036 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i16> [[TMP8]] to <8 x i8>
20037 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
20038 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
20039 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
20051 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i16>]* [[COERCE_DIVE]] to [6 x i64]*
20053 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x8x3_t* [[__S1]] to i8*
20054 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x8x3_t* [[B]] to i8*
20056 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
20060 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
20064 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
20068 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i16> [[TMP8]] to <16 x i8>
20069 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
20070 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
20071 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
20082 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i32>]* [[COERCE_DIVE]] to [6 x i64]*
20084 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x4x3_t* [[__S1]] to i8*
20085 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x4x3_t* [[B]] to i8*
20087 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
20091 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i32> [[TMP4]] to <16 x i8>
20095 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i32> [[TMP6]] to <16 x i8>
20099 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i32> [[TMP8]] to <16 x i8>
20100 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x i32>
20101 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x i32>
20102 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <4 x i32>
20113 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i16>]* [[COERCE_DIVE]] to [6 x i64]*
20115 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x8x3_t* [[__S1]] to i8*
20116 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x8x3_t* [[B]] to i8*
20118 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
20122 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
20126 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
20130 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i16> [[TMP8]] to <16 x i8>
20131 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
20132 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
20133 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
20144 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i32>]* [[COERCE_DIVE]] to [6 x i64]*
20146 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x4x3_t* [[__S1]] to i8*
20147 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x4x3_t* [[B]] to i8*
20149 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
20153 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i32> [[TMP4]] to <16 x i8>
20157 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i32> [[TMP6]] to <16 x i8>
20161 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i32> [[TMP8]] to <16 x i8>
20162 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x i32>
20163 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x i32>
20164 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <4 x i32>
20175 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x half>]* [[COERCE_DIVE]] to [6 x i64]*
20177 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x8x3_t* [[__S1]] to i8*
20178 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x8x3_t* [[B]] to i8*
20180 // CHECK:   [[TMP3:%.*]] = bitcast half* %a to i8*
20184 // CHECK:   [[TMP5:%.*]] = bitcast <8 x half> [[TMP4]] to <16 x i8>
20188 // CHECK:   [[TMP7:%.*]] = bitcast <8 x half> [[TMP6]] to <16 x i8>
20192 // CHECK:   [[TMP9:%.*]] = bitcast <8 x half> [[TMP8]] to <16 x i8>
20193 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
20194 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
20195 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
20206 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x float>]* [[COERCE_DIVE]] to [6 x i64]*
20208 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x4x3_t* [[__S1]] to i8*
20209 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x4x3_t* [[B]] to i8*
20211 // CHECK:   [[TMP3:%.*]] = bitcast float* %a to i8*
20215 // CHECK:   [[TMP5:%.*]] = bitcast <4 x float> [[TMP4]] to <16 x i8>
20219 // CHECK:   [[TMP7:%.*]] = bitcast <4 x float> [[TMP6]] to <16 x i8>
20223 // CHECK:   [[TMP9:%.*]] = bitcast <4 x float> [[TMP8]] to <16 x i8>
20224 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x float>
20225 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x float>
20226 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <4 x float>
20237 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i16>]* [[COERCE_DIVE]] to [6 x i64]*
20239 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x8x3_t* [[__S1]] to i8*
20240 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x8x3_t* [[B]] to i8*
20242 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
20246 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
20250 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
20254 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i16> [[TMP8]] to <16 x i8>
20255 // CHECK:   [[TMP10:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
20256 // CHECK:   [[TMP11:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
20257 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
20268 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
20270 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint8x8x3_t* [[__S1]] to i8*
20271 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x8x3_t* [[B]] to i8*
20292 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i16>]* [[COERCE_DIVE]] to [3 x i64]*
20294 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x4x3_t* [[__S1]] to i8*
20295 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x4x3_t* [[B]] to i8*
20297 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
20301 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
20305 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
20309 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i16> [[TMP8]] to <8 x i8>
20310 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
20311 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
20312 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
20323 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <2 x i32>]* [[COERCE_DIVE]] to [3 x i64]*
20325 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x2x3_t* [[__S1]] to i8*
20326 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x2x3_t* [[B]] to i8*
20328 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
20332 // CHECK:   [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <8 x i8>
20336 // CHECK:   [[TMP7:%.*]] = bitcast <2 x i32> [[TMP6]] to <8 x i8>
20340 // CHECK:   [[TMP9:%.*]] = bitcast <2 x i32> [[TMP8]] to <8 x i8>
20341 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x i32>
20342 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x i32>
20343 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <2 x i32>
20354 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
20356 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int8x8x3_t* [[__S1]] to i8*
20357 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x8x3_t* [[B]] to i8*
20378 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i16>]* [[COERCE_DIVE]] to [3 x i64]*
20380 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x4x3_t* [[__S1]] to i8*
20381 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x4x3_t* [[B]] to i8*
20383 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
20387 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
20391 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
20395 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i16> [[TMP8]] to <8 x i8>
20396 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
20397 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
20398 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
20409 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <2 x i32>]* [[COERCE_DIVE]] to [3 x i64]*
20411 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x2x3_t* [[__S1]] to i8*
20412 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x2x3_t* [[B]] to i8*
20414 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
20418 // CHECK:   [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <8 x i8>
20422 // CHECK:   [[TMP7:%.*]] = bitcast <2 x i32> [[TMP6]] to <8 x i8>
20426 // CHECK:   [[TMP9:%.*]] = bitcast <2 x i32> [[TMP8]] to <8 x i8>
20427 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x i32>
20428 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x i32>
20429 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <2 x i32>
20440 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x half>]* [[COERCE_DIVE]] to [3 x i64]*
20442 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x4x3_t* [[__S1]] to i8*
20443 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x4x3_t* [[B]] to i8*
20445 // CHECK:   [[TMP3:%.*]] = bitcast half* %a to i8*
20449 // CHECK:   [[TMP5:%.*]] = bitcast <4 x half> [[TMP4]] to <8 x i8>
20453 // CHECK:   [[TMP7:%.*]] = bitcast <4 x half> [[TMP6]] to <8 x i8>
20457 // CHECK:   [[TMP9:%.*]] = bitcast <4 x half> [[TMP8]] to <8 x i8>
20458 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
20459 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
20460 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
20471 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <2 x float>]* [[COERCE_DIVE]] to [3 x i64]*
20473 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x2x3_t* [[__S1]] to i8*
20474 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x2x3_t* [[B]] to i8*
20476 // CHECK:   [[TMP3:%.*]] = bitcast float* %a to i8*
20480 // CHECK:   [[TMP5:%.*]] = bitcast <2 x float> [[TMP4]] to <8 x i8>
20484 // CHECK:   [[TMP7:%.*]] = bitcast <2 x float> [[TMP6]] to <8 x i8>
20488 // CHECK:   [[TMP9:%.*]] = bitcast <2 x float> [[TMP8]] to <8 x i8>
20489 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x float>
20490 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x float>
20491 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <2 x float>
20502 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
20504 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly8x8x3_t* [[__S1]] to i8*
20505 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x8x3_t* [[B]] to i8*
20526 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <4 x i16>]* [[COERCE_DIVE]] to [3 x i64]*
20528 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x4x3_t* [[__S1]] to i8*
20529 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x4x3_t* [[B]] to i8*
20531 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
20535 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
20539 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
20543 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i16> [[TMP8]] to <8 x i8>
20544 // CHECK:   [[TMP10:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
20545 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
20546 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
20558 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <16 x i8>]* [[COERCE_DIVE]] to [8 x i64]*
20560 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint8x16x4_t* [[__S1]] to i8*
20561 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x16x4_t* [[B]] to i8*
20585 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i16>]* [[COERCE_DIVE]] to [8 x i64]*
20587 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x8x4_t* [[__S1]] to i8*
20588 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x8x4_t* [[B]] to i8*
20590 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
20594 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
20598 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
20602 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i16> [[TMP8]] to <16 x i8>
20606 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i16> [[TMP10]] to <16 x i8>
20607 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
20608 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
20609 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
20610 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <8 x i16>
20621 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i32>]* [[COERCE_DIVE]] to [8 x i64]*
20623 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x4x4_t* [[__S1]] to i8*
20624 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x4x4_t* [[B]] to i8*
20626 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
20630 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i32> [[TMP4]] to <16 x i8>
20634 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i32> [[TMP6]] to <16 x i8>
20638 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i32> [[TMP8]] to <16 x i8>
20642 // CHECK:   [[TMP11:%.*]] = bitcast <4 x i32> [[TMP10]] to <16 x i8>
20643 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x i32>
20644 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x i32>
20645 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <4 x i32>
20646 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <4 x i32>
20657 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <16 x i8>]* [[COERCE_DIVE]] to [8 x i64]*
20659 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int8x16x4_t* [[__S1]] to i8*
20660 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x16x4_t* [[B]] to i8*
20684 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i16>]* [[COERCE_DIVE]] to [8 x i64]*
20686 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x8x4_t* [[__S1]] to i8*
20687 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x8x4_t* [[B]] to i8*
20689 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
20693 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
20697 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
20701 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i16> [[TMP8]] to <16 x i8>
20705 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i16> [[TMP10]] to <16 x i8>
20706 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
20707 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
20708 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
20709 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <8 x i16>
20720 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i32>]* [[COERCE_DIVE]] to [8 x i64]*
20722 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x4x4_t* [[__S1]] to i8*
20723 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x4x4_t* [[B]] to i8*
20725 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
20729 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i32> [[TMP4]] to <16 x i8>
20733 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i32> [[TMP6]] to <16 x i8>
20737 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i32> [[TMP8]] to <16 x i8>
20741 // CHECK:   [[TMP11:%.*]] = bitcast <4 x i32> [[TMP10]] to <16 x i8>
20742 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x i32>
20743 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x i32>
20744 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <4 x i32>
20745 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <4 x i32>
20756 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x half>]* [[COERCE_DIVE]] to [8 x i64]*
20758 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x8x4_t* [[__S1]] to i8*
20759 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x8x4_t* [[B]] to i8*
20761 // CHECK:   [[TMP3:%.*]] = bitcast half* %a to i8*
20765 // CHECK:   [[TMP5:%.*]] = bitcast <8 x half> [[TMP4]] to <16 x i8>
20769 // CHECK:   [[TMP7:%.*]] = bitcast <8 x half> [[TMP6]] to <16 x i8>
20773 // CHECK:   [[TMP9:%.*]] = bitcast <8 x half> [[TMP8]] to <16 x i8>
20777 // CHECK:   [[TMP11:%.*]] = bitcast <8 x half> [[TMP10]] to <16 x i8>
20778 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
20779 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
20780 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
20781 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <8 x i16>
20792 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x float>]* [[COERCE_DIVE]] to [8 x i64]*
20794 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x4x4_t* [[__S1]] to i8*
20795 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x4x4_t* [[B]] to i8*
20797 // CHECK:   [[TMP3:%.*]] = bitcast float* %a to i8*
20801 // CHECK:   [[TMP5:%.*]] = bitcast <4 x float> [[TMP4]] to <16 x i8>
20805 // CHECK:   [[TMP7:%.*]] = bitcast <4 x float> [[TMP6]] to <16 x i8>
20809 // CHECK:   [[TMP9:%.*]] = bitcast <4 x float> [[TMP8]] to <16 x i8>
20813 // CHECK:   [[TMP11:%.*]] = bitcast <4 x float> [[TMP10]] to <16 x i8>
20814 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x float>
20815 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x float>
20816 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <4 x float>
20817 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <4 x float>
20828 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <16 x i8>]* [[COERCE_DIVE]] to [8 x i64]*
20830 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly8x16x4_t* [[__S1]] to i8*
20831 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x16x4_t* [[B]] to i8*
20855 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i16>]* [[COERCE_DIVE]] to [8 x i64]*
20857 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x8x4_t* [[__S1]] to i8*
20858 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x8x4_t* [[B]] to i8*
20860 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
20864 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
20868 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
20872 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i16> [[TMP8]] to <16 x i8>
20876 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i16> [[TMP10]] to <16 x i8>
20877 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
20878 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
20879 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
20880 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <8 x i16>
20891 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
20893 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint8x8x4_t* [[__S1]] to i8*
20894 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x8x4_t* [[B]] to i8*
20918 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
20920 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x4x4_t* [[__S1]] to i8*
20921 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x4x4_t* [[B]] to i8*
20923 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
20927 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
20931 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
20935 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i16> [[TMP8]] to <8 x i8>
20939 // CHECK:   [[TMP11:%.*]] = bitcast <4 x i16> [[TMP10]] to <8 x i8>
20940 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
20941 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
20942 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
20943 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <4 x i16>
20954 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <2 x i32>]* [[COERCE_DIVE]] to [4 x i64]*
20956 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x2x4_t* [[__S1]] to i8*
20957 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x2x4_t* [[B]] to i8*
20959 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
20963 // CHECK:   [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <8 x i8>
20967 // CHECK:   [[TMP7:%.*]] = bitcast <2 x i32> [[TMP6]] to <8 x i8>
20971 // CHECK:   [[TMP9:%.*]] = bitcast <2 x i32> [[TMP8]] to <8 x i8>
20975 // CHECK:   [[TMP11:%.*]] = bitcast <2 x i32> [[TMP10]] to <8 x i8>
20976 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x i32>
20977 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x i32>
20978 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <2 x i32>
20979 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <2 x i32>
20990 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <1 x i64>]* [[COERCE_DIVE]] to [4 x i64]*
20992 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint64x1x4_t* [[__S1]] to i8*
20993 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint64x1x4_t* [[B]] to i8*
20995 // CHECK:   [[TMP3:%.*]] = bitcast i64* %a to i8*
20999 // CHECK:   [[TMP5:%.*]] = bitcast <1 x i64> [[TMP4]] to <8 x i8>
21003 // CHECK:   [[TMP7:%.*]] = bitcast <1 x i64> [[TMP6]] to <8 x i8>
21007 // CHECK:   [[TMP9:%.*]] = bitcast <1 x i64> [[TMP8]] to <8 x i8>
21011 // CHECK:   [[TMP11:%.*]] = bitcast <1 x i64> [[TMP10]] to <8 x i8>
21012 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <1 x i64>
21013 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <1 x i64>
21014 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <1 x i64>
21015 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <1 x i64>
21026 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
21028 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int8x8x4_t* [[__S1]] to i8*
21029 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x8x4_t* [[B]] to i8*
21053 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
21055 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x4x4_t* [[__S1]] to i8*
21056 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x4x4_t* [[B]] to i8*
21058 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
21062 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
21066 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
21070 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i16> [[TMP8]] to <8 x i8>
21074 // CHECK:   [[TMP11:%.*]] = bitcast <4 x i16> [[TMP10]] to <8 x i8>
21075 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
21076 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
21077 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
21078 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <4 x i16>
21089 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <2 x i32>]* [[COERCE_DIVE]] to [4 x i64]*
21091 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x2x4_t* [[__S1]] to i8*
21092 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x2x4_t* [[B]] to i8*
21094 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
21098 // CHECK:   [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <8 x i8>
21102 // CHECK:   [[TMP7:%.*]] = bitcast <2 x i32> [[TMP6]] to <8 x i8>
21106 // CHECK:   [[TMP9:%.*]] = bitcast <2 x i32> [[TMP8]] to <8 x i8>
21110 // CHECK:   [[TMP11:%.*]] = bitcast <2 x i32> [[TMP10]] to <8 x i8>
21111 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x i32>
21112 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x i32>
21113 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <2 x i32>
21114 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <2 x i32>
21125 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <1 x i64>]* [[COERCE_DIVE]] to [4 x i64]*
21127 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int64x1x4_t* [[__S1]] to i8*
21128 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int64x1x4_t* [[B]] to i8*
21130 // CHECK:   [[TMP3:%.*]] = bitcast i64* %a to i8*
21134 // CHECK:   [[TMP5:%.*]] = bitcast <1 x i64> [[TMP4]] to <8 x i8>
21138 // CHECK:   [[TMP7:%.*]] = bitcast <1 x i64> [[TMP6]] to <8 x i8>
21142 // CHECK:   [[TMP9:%.*]] = bitcast <1 x i64> [[TMP8]] to <8 x i8>
21146 // CHECK:   [[TMP11:%.*]] = bitcast <1 x i64> [[TMP10]] to <8 x i8>
21147 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <1 x i64>
21148 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <1 x i64>
21149 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <1 x i64>
21150 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <1 x i64>
21161 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x half>]* [[COERCE_DIVE]] to [4 x i64]*
21163 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x4x4_t* [[__S1]] to i8*
21164 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x4x4_t* [[B]] to i8*
21166 // CHECK:   [[TMP3:%.*]] = bitcast half* %a to i8*
21170 // CHECK:   [[TMP5:%.*]] = bitcast <4 x half> [[TMP4]] to <8 x i8>
21174 // CHECK:   [[TMP7:%.*]] = bitcast <4 x half> [[TMP6]] to <8 x i8>
21178 // CHECK:   [[TMP9:%.*]] = bitcast <4 x half> [[TMP8]] to <8 x i8>
21182 // CHECK:   [[TMP11:%.*]] = bitcast <4 x half> [[TMP10]] to <8 x i8>
21183 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
21184 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
21185 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
21186 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <4 x i16>
21197 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <2 x float>]* [[COERCE_DIVE]] to [4 x i64]*
21199 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x2x4_t* [[__S1]] to i8*
21200 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x2x4_t* [[B]] to i8*
21202 // CHECK:   [[TMP3:%.*]] = bitcast float* %a to i8*
21206 // CHECK:   [[TMP5:%.*]] = bitcast <2 x float> [[TMP4]] to <8 x i8>
21210 // CHECK:   [[TMP7:%.*]] = bitcast <2 x float> [[TMP6]] to <8 x i8>
21214 // CHECK:   [[TMP9:%.*]] = bitcast <2 x float> [[TMP8]] to <8 x i8>
21218 // CHECK:   [[TMP11:%.*]] = bitcast <2 x float> [[TMP10]] to <8 x i8>
21219 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x float>
21220 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x float>
21221 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <2 x float>
21222 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <2 x float>
21233 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
21235 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly8x8x4_t* [[__S1]] to i8*
21236 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x8x4_t* [[B]] to i8*
21260 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
21262 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x4x4_t* [[__S1]] to i8*
21263 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x4x4_t* [[B]] to i8*
21265 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
21269 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
21273 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
21277 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i16> [[TMP8]] to <8 x i8>
21281 // CHECK:   [[TMP11:%.*]] = bitcast <4 x i16> [[TMP10]] to <8 x i8>
21282 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
21283 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
21284 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
21285 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <4 x i16>
21297 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i16>]* [[COERCE_DIVE]] to [8 x i64]*
21299 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x8x4_t* [[__S1]] to i8*
21300 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x8x4_t* [[B]] to i8*
21302 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
21306 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
21310 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
21314 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i16> [[TMP8]] to <16 x i8>
21318 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i16> [[TMP10]] to <16 x i8>
21319 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
21320 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
21321 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
21322 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <8 x i16>
21333 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i32>]* [[COERCE_DIVE]] to [8 x i64]*
21335 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x4x4_t* [[__S1]] to i8*
21336 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x4x4_t* [[B]] to i8*
21338 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
21342 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i32> [[TMP4]] to <16 x i8>
21346 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i32> [[TMP6]] to <16 x i8>
21350 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i32> [[TMP8]] to <16 x i8>
21354 // CHECK:   [[TMP11:%.*]] = bitcast <4 x i32> [[TMP10]] to <16 x i8>
21355 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x i32>
21356 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x i32>
21357 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <4 x i32>
21358 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <4 x i32>
21369 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i16>]* [[COERCE_DIVE]] to [8 x i64]*
21371 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x8x4_t* [[__S1]] to i8*
21372 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x8x4_t* [[B]] to i8*
21374 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
21378 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
21382 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
21386 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i16> [[TMP8]] to <16 x i8>
21390 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i16> [[TMP10]] to <16 x i8>
21391 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
21392 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
21393 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
21394 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <8 x i16>
21405 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i32>]* [[COERCE_DIVE]] to [8 x i64]*
21407 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x4x4_t* [[__S1]] to i8*
21408 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x4x4_t* [[B]] to i8*
21410 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
21414 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i32> [[TMP4]] to <16 x i8>
21418 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i32> [[TMP6]] to <16 x i8>
21422 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i32> [[TMP8]] to <16 x i8>
21426 // CHECK:   [[TMP11:%.*]] = bitcast <4 x i32> [[TMP10]] to <16 x i8>
21427 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x i32>
21428 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x i32>
21429 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <4 x i32>
21430 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <4 x i32>
21441 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x half>]* [[COERCE_DIVE]] to [8 x i64]*
21443 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x8x4_t* [[__S1]] to i8*
21444 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x8x4_t* [[B]] to i8*
21446 // CHECK:   [[TMP3:%.*]] = bitcast half* %a to i8*
21450 // CHECK:   [[TMP5:%.*]] = bitcast <8 x half> [[TMP4]] to <16 x i8>
21454 // CHECK:   [[TMP7:%.*]] = bitcast <8 x half> [[TMP6]] to <16 x i8>
21458 // CHECK:   [[TMP9:%.*]] = bitcast <8 x half> [[TMP8]] to <16 x i8>
21462 // CHECK:   [[TMP11:%.*]] = bitcast <8 x half> [[TMP10]] to <16 x i8>
21463 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
21464 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
21465 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
21466 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <8 x i16>
21477 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x float>]* [[COERCE_DIVE]] to [8 x i64]*
21479 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x4x4_t* [[__S1]] to i8*
21480 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x4x4_t* [[B]] to i8*
21482 // CHECK:   [[TMP3:%.*]] = bitcast float* %a to i8*
21486 // CHECK:   [[TMP5:%.*]] = bitcast <4 x float> [[TMP4]] to <16 x i8>
21490 // CHECK:   [[TMP7:%.*]] = bitcast <4 x float> [[TMP6]] to <16 x i8>
21494 // CHECK:   [[TMP9:%.*]] = bitcast <4 x float> [[TMP8]] to <16 x i8>
21498 // CHECK:   [[TMP11:%.*]] = bitcast <4 x float> [[TMP10]] to <16 x i8>
21499 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <4 x float>
21500 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <4 x float>
21501 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <4 x float>
21502 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <4 x float>
21513 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i16>]* [[COERCE_DIVE]] to [8 x i64]*
21515 bitcast %struct.poly16x8x4_t* [[__S1]] to i8*
21516 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x8x4_t* [[B]] to i8*
21518 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
21522 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i16> [[TMP4]] to <16 x i8>
21526 // CHECK:   [[TMP7:%.*]] = bitcast <8 x i16> [[TMP6]] to <16 x i8>
21530 // CHECK:   [[TMP9:%.*]] = bitcast <8 x i16> [[TMP8]] to <16 x i8>
21534 // CHECK:   [[TMP11:%.*]] = bitcast <8 x i16> [[TMP10]] to <16 x i8>
21535 // CHECK:   [[TMP12:%.*]] = bitcast <16 x i8> [[TMP5]] to <8 x i16>
21536 // CHECK:   [[TMP13:%.*]] = bitcast <16 x i8> [[TMP7]] to <8 x i16>
21537 // CHECK:   [[TMP14:%.*]] = bitcast <16 x i8> [[TMP9]] to <8 x i16>
21538 // CHECK:   [[TMP15:%.*]] = bitcast <16 x i8> [[TMP11]] to <8 x i16>
21549 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
21551 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint8x8x4_t* [[__S1]] to i8*
21552 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint8x8x4_t* [[B]] to i8*
21576 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
21578 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint16x4x4_t* [[__S1]] to i8*
21579 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint16x4x4_t* [[B]] to i8*
21581 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
21585 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
21589 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
21593 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i16> [[TMP8]] to <8 x i8>
21597 // CHECK:   [[TMP11:%.*]] = bitcast <4 x i16> [[TMP10]] to <8 x i8>
21598 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
21599 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
21600 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
21601 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <4 x i16>
21612 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <2 x i32>]* [[COERCE_DIVE]] to [4 x i64]*
21614 // CHECK:   [[TMP1:%.*]] = bitcast %struct.uint32x2x4_t* [[__S1]] to i8*
21615 // CHECK:   [[TMP2:%.*]] = bitcast %struct.uint32x2x4_t* [[B]] to i8*
21617 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
21621 // CHECK:   [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <8 x i8>
21625 // CHECK:   [[TMP7:%.*]] = bitcast <2 x i32> [[TMP6]] to <8 x i8>
21629 // CHECK:   [[TMP9:%.*]] = bitcast <2 x i32> [[TMP8]] to <8 x i8>
21633 // CHECK:   [[TMP11:%.*]] = bitcast <2 x i32> [[TMP10]] to <8 x i8>
21634 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x i32>
21635 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x i32>
21636 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <2 x i32>
21637 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <2 x i32>
21648 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
21650 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int8x8x4_t* [[__S1]] to i8*
21651 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int8x8x4_t* [[B]] to i8*
21675 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
21677 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int16x4x4_t* [[__S1]] to i8*
21678 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int16x4x4_t* [[B]] to i8*
21680 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
21684 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
21688 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
21692 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i16> [[TMP8]] to <8 x i8>
21696 // CHECK:   [[TMP11:%.*]] = bitcast <4 x i16> [[TMP10]] to <8 x i8>
21697 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
21698 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
21699 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
21700 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <4 x i16>
21711 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <2 x i32>]* [[COERCE_DIVE]] to [4 x i64]*
21713 // CHECK:   [[TMP1:%.*]] = bitcast %struct.int32x2x4_t* [[__S1]] to i8*
21714 // CHECK:   [[TMP2:%.*]] = bitcast %struct.int32x2x4_t* [[B]] to i8*
21716 // CHECK:   [[TMP3:%.*]] = bitcast i32* %a to i8*
21720 // CHECK:   [[TMP5:%.*]] = bitcast <2 x i32> [[TMP4]] to <8 x i8>
21724 // CHECK:   [[TMP7:%.*]] = bitcast <2 x i32> [[TMP6]] to <8 x i8>
21728 // CHECK:   [[TMP9:%.*]] = bitcast <2 x i32> [[TMP8]] to <8 x i8>
21732 // CHECK:   [[TMP11:%.*]] = bitcast <2 x i32> [[TMP10]] to <8 x i8>
21733 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x i32>
21734 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x i32>
21735 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <2 x i32>
21736 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <2 x i32>
21747 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x half>]* [[COERCE_DIVE]] to [4 x i64]*
21749 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float16x4x4_t* [[__S1]] to i8*
21750 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float16x4x4_t* [[B]] to i8*
21752 // CHECK:   [[TMP3:%.*]] = bitcast half* %a to i8*
21756 // CHECK:   [[TMP5:%.*]] = bitcast <4 x half> [[TMP4]] to <8 x i8>
21760 // CHECK:   [[TMP7:%.*]] = bitcast <4 x half> [[TMP6]] to <8 x i8>
21764 // CHECK:   [[TMP9:%.*]] = bitcast <4 x half> [[TMP8]] to <8 x i8>
21768 // CHECK:   [[TMP11:%.*]] = bitcast <4 x half> [[TMP10]] to <8 x i8>
21769 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
21770 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
21771 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
21772 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <4 x i16>
21783 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <2 x float>]* [[COERCE_DIVE]] to [4 x i64]*
21785 // CHECK:   [[TMP1:%.*]] = bitcast %struct.float32x2x4_t* [[__S1]] to i8*
21786 // CHECK:   [[TMP2:%.*]] = bitcast %struct.float32x2x4_t* [[B]] to i8*
21788 // CHECK:   [[TMP3:%.*]] = bitcast float* %a to i8*
21792 // CHECK:   [[TMP5:%.*]] = bitcast <2 x float> [[TMP4]] to <8 x i8>
21796 // CHECK:   [[TMP7:%.*]] = bitcast <2 x float> [[TMP6]] to <8 x i8>
21800 // CHECK:   [[TMP9:%.*]] = bitcast <2 x float> [[TMP8]] to <8 x i8>
21804 // CHECK:   [[TMP11:%.*]] = bitcast <2 x float> [[TMP10]] to <8 x i8>
21805 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <2 x float>
21806 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <2 x float>
21807 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <2 x float>
21808 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <2 x float>
21819 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
21821 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly8x8x4_t* [[__S1]] to i8*
21822 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly8x8x4_t* [[B]] to i8*
21846 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <4 x i16>]* [[COERCE_DIVE]] to [4 x i64]*
21848 // CHECK:   [[TMP1:%.*]] = bitcast %struct.poly16x4x4_t* [[__S1]] to i8*
21849 // CHECK:   [[TMP2:%.*]] = bitcast %struct.poly16x4x4_t* [[B]] to i8*
21851 // CHECK:   [[TMP3:%.*]] = bitcast i16* %a to i8*
21855 // CHECK:   [[TMP5:%.*]] = bitcast <4 x i16> [[TMP4]] to <8 x i8>
21859 // CHECK:   [[TMP7:%.*]] = bitcast <4 x i16> [[TMP6]] to <8 x i8>
21863 // CHECK:   [[TMP9:%.*]] = bitcast <4 x i16> [[TMP8]] to <8 x i8>
21867 // CHECK:   [[TMP11:%.*]] = bitcast <4 x i16> [[TMP10]] to <8 x i8>
21868 // CHECK:   [[TMP12:%.*]] = bitcast <8 x i8> [[TMP5]] to <4 x i16>
21869 // CHECK:   [[TMP13:%.*]] = bitcast <8 x i8> [[TMP7]] to <4 x i16>
21870 // CHECK:   [[TMP14:%.*]] = bitcast <8 x i8> [[TMP9]] to <4 x i16>
21871 // CHECK:   [[TMP15:%.*]] = bitcast <8 x i8> [[TMP11]] to <4 x i16>
22007 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
22008 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
22009 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
22010 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
22020 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
22021 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
22022 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
22023 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
22033 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
22034 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
22035 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
22036 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
22046 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
22047 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
22048 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
22049 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
22059 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
22060 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
22061 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
22062 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
22072 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i64> %a to <16 x i8>
22073 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i64> %b to <16 x i8>
22074 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <2 x i64>
22075 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x i64>
22095 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
22096 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
22098 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %b to <8 x i8>
22099 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
22108 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
22109 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
22111 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %b to <8 x i8>
22112 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
22130 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
22131 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
22133 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %b to <8 x i8>
22134 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
22143 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
22144 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
22146 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %b to <8 x i8>
22147 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
22165 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
22166 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
22175 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
22176 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
22193 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %b to <8 x i8>
22194 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
22203 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %b to <8 x i8>
22204 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
22239 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
22242 // CHECK:   [[TMP1:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE1]] to [2 x i64]*
22245 // CHECK:   [[TMP3:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE_I]] to [2 x i64]*
22263 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
22266 // CHECK:   [[TMP1:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE1]] to [2 x i64]*
22269 // CHECK:   [[TMP3:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE_I]] to [2 x i64]*
22287 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
22290 // CHECK:   [[TMP1:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE1]] to [2 x i64]*
22293 // CHECK:   [[TMP3:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE_I]] to [2 x i64]*
22312 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
22315 // CHECK:   [[TMP1:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE1]] to [3 x i64]*
22318 // CHECK:   [[TMP3:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE_I]] to [3 x i64]*
22339 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
22342 // CHECK:   [[TMP1:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE1]] to [3 x i64]*
22345 // CHECK:   [[TMP3:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE_I]] to [3 x i64]*
22366 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
22369 // CHECK:   [[TMP1:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE1]] to [3 x i64]*
22372 // CHECK:   [[TMP3:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE_I]] to [3 x i64]*
22394 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
22397 // CHECK:   [[TMP1:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE1]] to [4 x i64]*
22400 // CHECK:   [[TMP3:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE_I]] to [4 x i64]*
22424 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
22427 // CHECK:   [[TMP1:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE1]] to [4 x i64]*
22430 // CHECK:   [[TMP3:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE_I]] to [4 x i64]*
22454 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
22457 // CHECK:   [[TMP1:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE1]] to [4 x i64]*
22460 // CHECK:   [[TMP3:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE_I]] to [4 x i64]*
22507 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
22510 bitcast [2 x <8 x i8>]* [[COERCE_DIVE1]] to [2 x i64]*
22513 // CHECK:   [[TMP3:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE_I]] to [2 x i64]*
22531 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
22534 // CHECK:   [[TMP1:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE1]] to [2 x i64]*
22537 // CHECK:   [[TMP3:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE_I]] to [2 x i64]*
22555 // CHECK:   [[TMP0:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE]] to [2 x i64]*
22558 // CHECK:   [[TMP1:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE1]] to [2 x i64]*
22561 // CHECK:   [[TMP3:%.*]] = bitcast [2 x <8 x i8>]* [[COERCE_DIVE_I]] to [2 x i64]*
22580 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
22583 // CHECK:   [[TMP1:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE1]] to [3 x i64]*
22586 // CHECK:   [[TMP3:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE_I]] to [3 x i64]*
22607 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
22610 // CHECK:   [[TMP1:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE1]] to [3 x i64]*
22613 // CHECK:   [[TMP3:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE_I]] to [3 x i64]*
22634 // CHECK:   [[TMP0:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE]] to [3 x i64]*
22637 // CHECK:   [[TMP1:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE1]] to [3 x i64]*
22640 // CHECK:   [[TMP3:%.*]] = bitcast [3 x <8 x i8>]* [[COERCE_DIVE_I]] to [3 x i64]*
22662 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
22665 // CHECK:   [[TMP1:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE1]] to [4 x i64]*
22668 // CHECK:   [[TMP3:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE_I]] to [4 x i64]*
22692 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
22695 // CHECK:   [[TMP1:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE1]] to [4 x i64]*
22698 // CHECK:   [[TMP3:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE_I]] to [4 x i64]*
22722 // CHECK:   [[TMP0:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE]] to [4 x i64]*
22725 // CHECK:   [[TMP1:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE1]] to [4 x i64]*
22728 // CHECK:   [[TMP3:%.*]] = bitcast [4 x <8 x i8>]* [[COERCE_DIVE_I]] to [4 x i64]*
22751 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x8x2_t* [[__RET_I]] to i8*
22752 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <8 x i8>*
22758 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x8x2_t* %agg.result to i8*
22759 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int8x8x2_t* [[__RET_I]] to i8*
22768 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x4x2_t* [[__RET_I]] to i8*
22769 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %a to <8 x i8>
22770 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %b to <8 x i8>
22771 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i16>*
22772 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
22773 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
22779 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int16x4x2_t* %agg.result to i8*
22780 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int16x4x2_t* [[__RET_I]] to i8*
22789 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x2x2_t* [[__RET_I]] to i8*
22790 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %a to <8 x i8>
22791 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %b to <8 x i8>
22792 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <2 x i32>*
22793 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
22794 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
22800 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int32x2x2_t* %agg.result to i8*
22801 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int32x2x2_t* [[__RET_I]] to i8*
22810 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x8x2_t* [[__RET_I]] to i8*
22811 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <8 x i8>*
22817 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x8x2_t* %agg.result to i8*
22818 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint8x8x2_t* [[__RET_I]] to i8*
22827 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x4x2_t* [[__RET_I]] to i8*
22828 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %a to <8 x i8>
22829 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %b to <8 x i8>
22830 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i16>*
22831 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
22832 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
22838 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint16x4x2_t* %agg.result to i8*
22839 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint16x4x2_t* [[__RET_I]] to i8*
22848 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x2x2_t* [[__RET_I]] to i8*
22849 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %a to <8 x i8>
22850 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %b to <8 x i8>
22851 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <2 x i32>*
22852 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
22853 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
22859 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint32x2x2_t* %agg.result to i8*
22860 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint32x2x2_t* [[__RET_I]] to i8*
22869 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x2x2_t* [[__RET_I]] to i8*
22870 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %a to <8 x i8>
22871 // CHECK:   [[TMP2:%.*]] = bitcast <2 x float> %b to <8 x i8>
22872 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <2 x float>*
22873 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
22874 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x float>
22880 // CHECK:   [[TMP7:%.*]] = bitcast %struct.float32x2x2_t* %agg.result to i8*
22881 // CHECK:   [[TMP8:%.*]] = bitcast %struct.float32x2x2_t* [[__RET_I]] to i8*
22890 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x8x2_t* [[__RET_I]] to i8*
22891 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <8 x i8>*
22897 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x8x2_t* %agg.result to i8*
22898 // CHECK:   [[TMP4:%.*]] = bitcast %struct.poly8x8x2_t* [[__RET_I]] to i8*
22907 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x4x2_t* [[__RET_I]] to i8*
22908 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %a to <8 x i8>
22909 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %b to <8 x i8>
22910 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i16>*
22911 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
22912 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
22918 // CHECK:   [[TMP7:%.*]] = bitcast %struct.poly16x4x2_t* %agg.result to i8*
22919 // CHECK:   [[TMP8:%.*]] = bitcast %struct.poly16x4x2_t* [[__RET_I]] to i8*
22928 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x16x2_t* [[__RET_I]] to i8*
22929 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <16 x i8>*
22935 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x16x2_t* %agg.result to i8*
22936 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int8x16x2_t* [[__RET_I]] to i8*
22945 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x8x2_t* [[__RET_I]] to i8*
22946 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %a to <16 x i8>
22947 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i16> %b to <16 x i8>
22948 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <8 x i16>*
22949 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
22950 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <8 x i16>
22956 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int16x8x2_t* %agg.result to i8*
22957 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int16x8x2_t* [[__RET_I]] to i8*
22966 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x4x2_t* [[__RET_I]] to i8*
22967 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %a to <16 x i8>
22968 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i32> %b to <16 x i8>
22969 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i32>*
22970 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
22971 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <4 x i32>
22977 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int32x4x2_t* %agg.result to i8*
22978 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int32x4x2_t* [[__RET_I]] to i8*
22987 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x16x2_t* [[__RET_I]] to i8*
22988 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <16 x i8>*
22994 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x16x2_t* %agg.result to i8*
22995 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint8x16x2_t* [[__RET_I]] to i8*
23004 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x8x2_t* [[__RET_I]] to i8*
23005 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %a to <16 x i8>
23006 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i16> %b to <16 x i8>
23007 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <8 x i16>*
23008 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
23009 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <8 x i16>
23015 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint16x8x2_t* %agg.result to i8*
23016 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint16x8x2_t* [[__RET_I]] to i8*
23025 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x4x2_t* [[__RET_I]] to i8*
23026 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %a to <16 x i8>
23027 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i32> %b to <16 x i8>
23028 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i32>*
23029 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
23030 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <4 x i32>
23036 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint32x4x2_t* %agg.result to i8*
23037 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint32x4x2_t* [[__RET_I]] to i8*
23046 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x4x2_t* [[__RET_I]] to i8*
23047 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %a to <16 x i8>
23048 // CHECK:   [[TMP2:%.*]] = bitcast <4 x float> %b to <16 x i8>
23049 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x float>*
23050 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
23051 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <4 x float>
23057 // CHECK:   [[TMP7:%.*]] = bitcast %struct.float32x4x2_t* %agg.result to i8*
23058 // CHECK:   [[TMP8:%.*]] = bitcast %struct.float32x4x2_t* [[__RET_I]] to i8*
23067 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x16x2_t* [[__RET_I]] to i8*
23068 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <16 x i8>*
23074 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x16x2_t* %agg.result to i8*
23075 // CHECK:   [[TMP4:%.*]] = bitcast %struct.poly8x16x2_t* [[__RET_I]] to i8*
23084 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x8x2_t* [[__RET_I]] to i8*
23085 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %a to <16 x i8>
23086 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i16> %b to <16 x i8>
23087 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <8 x i16>*
23088 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
23089 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <8 x i16>
23095 // CHECK:   [[TMP7:%.*]] = bitcast %struct.poly16x8x2_t* %agg.result to i8*
23096 // CHECK:   [[TMP8:%.*]] = bitcast %struct.poly16x8x2_t* [[__RET_I]] to i8*
23114 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
23115 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
23116 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
23117 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
23127 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
23128 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
23129 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
23130 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
23149 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
23150 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
23151 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
23152 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
23162 // CHECK:   [[TMP0:%.*]] = bitcast <2 x i32> %a to <8 x i8>
23163 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %b to <8 x i8>
23164 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <2 x i32>
23165 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
23184 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i16> %a to <8 x i8>
23185 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %b to <8 x i8>
23186 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i8> [[TMP0]] to <4 x i16>
23187 // CHECK:   [[TMP3:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
23206 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
23207 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
23208 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
23209 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
23219 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
23220 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
23221 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
23222 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
23241 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
23242 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
23243 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
23244 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
23254 // CHECK:   [[TMP0:%.*]] = bitcast <4 x i32> %a to <16 x i8>
23255 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %b to <16 x i8>
23256 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <4 x i32>
23257 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
23276 // CHECK:   [[TMP0:%.*]] = bitcast <8 x i16> %a to <16 x i8>
23277 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %b to <16 x i8>
23278 // CHECK:   [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to <8 x i16>
23279 // CHECK:   [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
23291 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x8x2_t* [[__RET_I]] to i8*
23292 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <8 x i8>*
23298 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x8x2_t* %agg.result to i8*
23299 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int8x8x2_t* [[__RET_I]] to i8*
23308 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x4x2_t* [[__RET_I]] to i8*
23309 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %a to <8 x i8>
23310 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %b to <8 x i8>
23311 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i16>*
23312 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
23313 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
23319 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int16x4x2_t* %agg.result to i8*
23320 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int16x4x2_t* [[__RET_I]] to i8*
23329 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x2x2_t* [[__RET_I]] to i8*
23330 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %a to <8 x i8>
23331 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %b to <8 x i8>
23332 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <2 x i32>*
23333 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
23334 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
23340 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int32x2x2_t* %agg.result to i8*
23341 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int32x2x2_t* [[__RET_I]] to i8*
23350 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x8x2_t* [[__RET_I]] to i8*
23351 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <8 x i8>*
23357 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x8x2_t* %agg.result to i8*
23358 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint8x8x2_t* [[__RET_I]] to i8*
23367 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x4x2_t* [[__RET_I]] to i8*
23368 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %a to <8 x i8>
23369 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %b to <8 x i8>
23370 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i16>*
23371 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
23372 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
23378 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint16x4x2_t* %agg.result to i8*
23379 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint16x4x2_t* [[__RET_I]] to i8*
23388 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x2x2_t* [[__RET_I]] to i8*
23389 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %a to <8 x i8>
23390 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %b to <8 x i8>
23391 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <2 x i32>*
23392 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
23393 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
23399 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint32x2x2_t* %agg.result to i8*
23400 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint32x2x2_t* [[__RET_I]] to i8*
23409 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x2x2_t* [[__RET_I]] to i8*
23410 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %a to <8 x i8>
23411 // CHECK:   [[TMP2:%.*]] = bitcast <2 x float> %b to <8 x i8>
23412 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <2 x float>*
23413 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
23414 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x float>
23420 // CHECK:   [[TMP7:%.*]] = bitcast %struct.float32x2x2_t* %agg.result to i8*
23421 // CHECK:   [[TMP8:%.*]] = bitcast %struct.float32x2x2_t* [[__RET_I]] to i8*
23430 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x8x2_t* [[__RET_I]] to i8*
23431 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <8 x i8>*
23437 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x8x2_t* %agg.result to i8*
23438 // CHECK:   [[TMP4:%.*]] = bitcast %struct.poly8x8x2_t* [[__RET_I]] to i8*
23447 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x4x2_t* [[__RET_I]] to i8*
23448 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %a to <8 x i8>
23449 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %b to <8 x i8>
23450 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i16>*
23451 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
23452 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
23458 // CHECK:   [[TMP7:%.*]] = bitcast %struct.poly16x4x2_t* %agg.result to i8*
23459 // CHECK:   [[TMP8:%.*]] = bitcast %struct.poly16x4x2_t* [[__RET_I]] to i8*
23468 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x16x2_t* [[__RET_I]] to i8*
23469 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <16 x i8>*
23475 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x16x2_t* %agg.result to i8*
23476 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int8x16x2_t* [[__RET_I]] to i8*
23485 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x8x2_t* [[__RET_I]] to i8*
23486 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %a to <16 x i8>
23487 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i16> %b to <16 x i8>
23488 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <8 x i16>*
23489 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
23490 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <8 x i16>
23496 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int16x8x2_t* %agg.result to i8*
23497 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int16x8x2_t* [[__RET_I]] to i8*
23506 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x4x2_t* [[__RET_I]] to i8*
23507 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %a to <16 x i8>
23508 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i32> %b to <16 x i8>
23509 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i32>*
23510 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
23511 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <4 x i32>
23517 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int32x4x2_t* %agg.result to i8*
23518 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int32x4x2_t* [[__RET_I]] to i8*
23527 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x16x2_t* [[__RET_I]] to i8*
23528 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <16 x i8>*
23534 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x16x2_t* %agg.result to i8*
23535 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint8x16x2_t* [[__RET_I]] to i8*
23544 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x8x2_t* [[__RET_I]] to i8*
23545 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %a to <16 x i8>
23546 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i16> %b to <16 x i8>
23547 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <8 x i16>*
23548 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
23549 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <8 x i16>
23555 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint16x8x2_t* %agg.result to i8*
23556 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint16x8x2_t* [[__RET_I]] to i8*
23565 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x4x2_t* [[__RET_I]] to i8*
23566 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %a to <16 x i8>
23567 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i32> %b to <16 x i8>
23568 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i32>*
23569 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
23570 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <4 x i32>
23576 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint32x4x2_t* %agg.result to i8*
23577 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint32x4x2_t* [[__RET_I]] to i8*
23586 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x4x2_t* [[__RET_I]] to i8*
23587 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %a to <16 x i8>
23588 // CHECK:   [[TMP2:%.*]] = bitcast <4 x float> %b to <16 x i8>
23589 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x float>*
23590 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
23591 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <4 x float>
23597 // CHECK:   [[TMP7:%.*]] = bitcast %struct.float32x4x2_t* %agg.result to i8*
23598 // CHECK:   [[TMP8:%.*]] = bitcast %struct.float32x4x2_t* [[__RET_I]] to i8*
23607 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x16x2_t* [[__RET_I]] to i8*
23608 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <16 x i8>*
23614 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x16x2_t* %agg.result to i8*
23615 // CHECK:   [[TMP4:%.*]] = bitcast %struct.poly8x16x2_t* [[__RET_I]] to i8*
23624 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x8x2_t* [[__RET_I]] to i8*
23625 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %a to <16 x i8>
23626 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i16> %b to <16 x i8>
23627 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <8 x i16>*
23628 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
23629 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <8 x i16>
23635 // CHECK:   [[TMP7:%.*]] = bitcast %struct.poly16x8x2_t* %agg.result to i8*
23636 // CHECK:   [[TMP8:%.*]] = bitcast %struct.poly16x8x2_t* [[__RET_I]] to i8*
23646 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x8x2_t* [[__RET_I]] to i8*
23647 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <8 x i8>*
23653 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x8x2_t* %agg.result to i8*
23654 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int8x8x2_t* [[__RET_I]] to i8*
23663 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x4x2_t* [[__RET_I]] to i8*
23664 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %a to <8 x i8>
23665 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %b to <8 x i8>
23666 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i16>*
23667 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
23668 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
23674 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int16x4x2_t* %agg.result to i8*
23675 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int16x4x2_t* [[__RET_I]] to i8*
23684 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x2x2_t* [[__RET_I]] to i8*
23685 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %a to <8 x i8>
23686 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %b to <8 x i8>
23687 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <2 x i32>*
23688 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
23689 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
23695 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int32x2x2_t* %agg.result to i8*
23696 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int32x2x2_t* [[__RET_I]] to i8*
23705 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x8x2_t* [[__RET_I]] to i8*
23706 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <8 x i8>*
23712 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x8x2_t* %agg.result to i8*
23713 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint8x8x2_t* [[__RET_I]] to i8*
23722 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x4x2_t* [[__RET_I]] to i8*
23723 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %a to <8 x i8>
23724 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %b to <8 x i8>
23725 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i16>*
23726 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
23727 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
23733 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint16x4x2_t* %agg.result to i8*
23734 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint16x4x2_t* [[__RET_I]] to i8*
23743 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x2x2_t* [[__RET_I]] to i8*
23744 // CHECK:   [[TMP1:%.*]] = bitcast <2 x i32> %a to <8 x i8>
23745 // CHECK:   [[TMP2:%.*]] = bitcast <2 x i32> %b to <8 x i8>
23746 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <2 x i32>*
23747 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x i32>
23748 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x i32>
23754 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint32x2x2_t* %agg.result to i8*
23755 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint32x2x2_t* [[__RET_I]] to i8*
23764 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x2x2_t* [[__RET_I]] to i8*
23765 // CHECK:   [[TMP1:%.*]] = bitcast <2 x float> %a to <8 x i8>
23766 // CHECK:   [[TMP2:%.*]] = bitcast <2 x float> %b to <8 x i8>
23767 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <2 x float>*
23768 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
23769 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <2 x float>
23775 // CHECK:   [[TMP7:%.*]] = bitcast %struct.float32x2x2_t* %agg.result to i8*
23776 // CHECK:   [[TMP8:%.*]] = bitcast %struct.float32x2x2_t* [[__RET_I]] to i8*
23785 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x8x2_t* [[__RET_I]] to i8*
23786 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <8 x i8>*
23792 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x8x2_t* %agg.result to i8*
23793 // CHECK:   [[TMP4:%.*]] = bitcast %struct.poly8x8x2_t* [[__RET_I]] to i8*
23802 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x4x2_t* [[__RET_I]] to i8*
23803 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i16> %a to <8 x i8>
23804 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i16> %b to <8 x i8>
23805 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i16>*
23806 // CHECK:   [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <4 x i16>
23807 // CHECK:   [[TMP5:%.*]] = bitcast <8 x i8> [[TMP2]] to <4 x i16>
23813 // CHECK:   [[TMP7:%.*]] = bitcast %struct.poly16x4x2_t* %agg.result to i8*
23814 // CHECK:   [[TMP8:%.*]] = bitcast %struct.poly16x4x2_t* [[__RET_I]] to i8*
23823 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int8x16x2_t* [[__RET_I]] to i8*
23824 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <16 x i8>*
23830 // CHECK:   [[TMP3:%.*]] = bitcast %struct.int8x16x2_t* %agg.result to i8*
23831 // CHECK:   [[TMP4:%.*]] = bitcast %struct.int8x16x2_t* [[__RET_I]] to i8*
23840 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int16x8x2_t* [[__RET_I]] to i8*
23841 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %a to <16 x i8>
23842 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i16> %b to <16 x i8>
23843 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <8 x i16>*
23844 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
23845 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <8 x i16>
23851 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int16x8x2_t* %agg.result to i8*
23852 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int16x8x2_t* [[__RET_I]] to i8*
23861 // CHECK:   [[TMP0:%.*]] = bitcast %struct.int32x4x2_t* [[__RET_I]] to i8*
23862 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %a to <16 x i8>
23863 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i32> %b to <16 x i8>
23864 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i32>*
23865 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
23866 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <4 x i32>
23872 // CHECK:   [[TMP7:%.*]] = bitcast %struct.int32x4x2_t* %agg.result to i8*
23873 // CHECK:   [[TMP8:%.*]] = bitcast %struct.int32x4x2_t* [[__RET_I]] to i8*
23882 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint8x16x2_t* [[__RET_I]] to i8*
23883 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <16 x i8>*
23889 // CHECK:   [[TMP3:%.*]] = bitcast %struct.uint8x16x2_t* %agg.result to i8*
23890 // CHECK:   [[TMP4:%.*]] = bitcast %struct.uint8x16x2_t* [[__RET_I]] to i8*
23899 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint16x8x2_t* [[__RET_I]] to i8*
23900 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %a to <16 x i8>
23901 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i16> %b to <16 x i8>
23902 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <8 x i16>*
23903 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
23904 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <8 x i16>
23910 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint16x8x2_t* %agg.result to i8*
23911 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint16x8x2_t* [[__RET_I]] to i8*
23920 // CHECK:   [[TMP0:%.*]] = bitcast %struct.uint32x4x2_t* [[__RET_I]] to i8*
23921 // CHECK:   [[TMP1:%.*]] = bitcast <4 x i32> %a to <16 x i8>
23922 // CHECK:   [[TMP2:%.*]] = bitcast <4 x i32> %b to <16 x i8>
23923 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x i32>*
23924 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x i32>
23925 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <4 x i32>
23931 // CHECK:   [[TMP7:%.*]] = bitcast %struct.uint32x4x2_t* %agg.result to i8*
23932 // CHECK:   [[TMP8:%.*]] = bitcast %struct.uint32x4x2_t* [[__RET_I]] to i8*
23941 // CHECK:   [[TMP0:%.*]] = bitcast %struct.float32x4x2_t* [[__RET_I]] to i8*
23942 // CHECK:   [[TMP1:%.*]] = bitcast <4 x float> %a to <16 x i8>
23943 // CHECK:   [[TMP2:%.*]] = bitcast <4 x float> %b to <16 x i8>
23944 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <4 x float>*
23945 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
23946 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <4 x float>
23952 // CHECK:   [[TMP7:%.*]] = bitcast %struct.float32x4x2_t* %agg.result to i8*
23953 // CHECK:   [[TMP8:%.*]] = bitcast %struct.float32x4x2_t* [[__RET_I]] to i8*
23962 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly8x16x2_t* [[__RET_I]] to i8*
23963 // CHECK:   [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <16 x i8>*
23969 // CHECK:   [[TMP3:%.*]] = bitcast %struct.poly8x16x2_t* %agg.result to i8*
23970 // CHECK:   [[TMP4:%.*]] = bitcast %struct.poly8x16x2_t* [[__RET_I]] to i8*
23979 // CHECK:   [[TMP0:%.*]] = bitcast %struct.poly16x8x2_t* [[__RET_I]] to i8*
23980 // CHECK:   [[TMP1:%.*]] = bitcast <8 x i16> %a to <16 x i8>
23981 // CHECK:   [[TMP2:%.*]] = bitcast <8 x i16> %b to <16 x i8>
23982 // CHECK:   [[TMP3:%.*]] = bitcast i8* [[TMP0]] to <8 x i16>*
23983 // CHECK:   [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <8 x i16>
23984 // CHECK:   [[TMP5:%.*]] = bitcast <16 x i8> [[TMP2]] to <8 x i16>
23990 // CHECK:   [[TMP7:%.*]] = bitcast %struct.poly16x8x2_t* %agg.result to i8*
23991 // CHECK:   [[TMP8:%.*]] = bitcast %struct.poly16x8x2_t* [[__RET_I]] to i8*
OpenGrok
