Lines Matching refs:CI

2 …verify-machineinstrs -mattr=+load-store-opt < %s | FileCheck -enable-var-scope --check-prefix=CI %s
10 ; CI-LABEL: simple_write2_one_val_f32:
11 ; CI: ; %bb.0:
12 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x2
13 ; CI-NEXT: s_mov_b32 s3, 0xf000
14 ; CI-NEXT: s_mov_b32 s2, 0
15 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
16 ; CI-NEXT: v_mov_b32_e32 v1, 0
17 ; CI-NEXT: s_waitcnt lgkmcnt(0)
18 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[0:3], 0 addr64
19 ; CI-NEXT: s_mov_b32 m0, -1
20 ; CI-NEXT: s_waitcnt vmcnt(0)
21 ; CI-NEXT: ds_write2_b32 v0, v1, v1 offset1:8
22 ; CI-NEXT: s_endpgm
45 ; CI-LABEL: simple_write2_two_val_f32:
46 ; CI: ; %bb.0:
47 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x2
48 ; CI-NEXT: s_mov_b32 s3, 0xf000
49 ; CI-NEXT: s_mov_b32 s2, 0
50 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
51 ; CI-NEXT: v_mov_b32_e32 v1, 0
52 ; CI-NEXT: s_waitcnt lgkmcnt(0)
53 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[0:3], 0 addr64 glc
54 ; CI-NEXT: s_waitcnt vmcnt(0)
55 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[0:3], 0 addr64 offset:4 glc
56 ; CI-NEXT: s_waitcnt vmcnt(0)
57 ; CI-NEXT: s_mov_b32 m0, -1
58 ; CI-NEXT: ds_write2_b32 v0, v2, v1 offset1:8
59 ; CI-NEXT: s_endpgm
86 ; CI-LABEL: simple_write2_two_val_f32_volatile_0:
87 ; CI: ; %bb.0:
88 ; CI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x2
89 ; CI-NEXT: s_mov_b32 s7, 0xf000
90 ; CI-NEXT: s_mov_b32 s6, 0
91 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
92 ; CI-NEXT: v_mov_b32_e32 v1, 0
93 ; CI-NEXT: s_waitcnt lgkmcnt(0)
94 ; CI-NEXT: s_mov_b64 s[4:5], s[0:1]
95 ; CI-NEXT: s_mov_b64 s[0:1], s[2:3]
96 ; CI-NEXT: s_mov_b64 s[2:3], s[6:7]
97 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[4:7], 0 addr64 glc
98 ; CI-NEXT: s_waitcnt vmcnt(0)
99 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[0:3], 0 addr64 glc
100 ; CI-NEXT: s_waitcnt vmcnt(0)
101 ; CI-NEXT: s_mov_b32 m0, -1
102 ; CI-NEXT: ds_write_b32 v0, v2
103 ; CI-NEXT: ds_write_b32 v0, v1 offset:32
104 ; CI-NEXT: s_endpgm
132 ; CI-LABEL: simple_write2_two_val_f32_volatile_1:
133 ; CI: ; %bb.0:
134 ; CI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x2
135 ; CI-NEXT: s_mov_b32 s7, 0xf000
136 ; CI-NEXT: s_mov_b32 s6, 0
137 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
138 ; CI-NEXT: v_mov_b32_e32 v1, 0
139 ; CI-NEXT: s_waitcnt lgkmcnt(0)
140 ; CI-NEXT: s_mov_b64 s[4:5], s[0:1]
141 ; CI-NEXT: s_mov_b64 s[0:1], s[2:3]
142 ; CI-NEXT: s_mov_b64 s[2:3], s[6:7]
143 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[4:7], 0 addr64 glc
144 ; CI-NEXT: s_waitcnt vmcnt(0)
145 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[0:3], 0 addr64 glc
146 ; CI-NEXT: s_waitcnt vmcnt(0)
147 ; CI-NEXT: s_mov_b32 m0, -1
148 ; CI-NEXT: ds_write_b32 v0, v2
149 ; CI-NEXT: ds_write_b32 v0, v1 offset:32
150 ; CI-NEXT: s_endpgm
183 ; CI-LABEL: simple_write2_two_val_subreg2_mixed_f32:
184 ; CI: ; %bb.0:
185 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x2
186 ; CI-NEXT: s_mov_b32 s3, 0xf000
187 ; CI-NEXT: s_mov_b32 s2, 0
188 ; CI-NEXT: v_lshlrev_b32_e32 v1, 3, v0
189 ; CI-NEXT: v_mov_b32_e32 v2, 0
190 ; CI-NEXT: s_waitcnt lgkmcnt(0)
191 ; CI-NEXT: buffer_load_dwordx2 v[3:4], v[1:2], s[0:3], 0 addr64 glc
192 ; CI-NEXT: s_waitcnt vmcnt(0)
193 ; CI-NEXT: buffer_load_dwordx2 v[1:2], v[1:2], s[0:3], 0 addr64 offset:8 glc
194 ; CI-NEXT: s_waitcnt vmcnt(0)
195 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
196 ; CI-NEXT: s_mov_b32 m0, -1
197 ; CI-NEXT: ds_write2_b32 v0, v3, v2 offset1:8
198 ; CI-NEXT: s_endpgm
230 ; CI-LABEL: simple_write2_two_val_subreg2_f32:
231 ; CI: ; %bb.0:
232 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x2
233 ; CI-NEXT: s_mov_b32 s3, 0xf000
234 ; CI-NEXT: s_mov_b32 s2, 0
235 ; CI-NEXT: v_lshlrev_b32_e32 v1, 3, v0
236 ; CI-NEXT: v_mov_b32_e32 v2, 0
237 ; CI-NEXT: s_waitcnt lgkmcnt(0)
238 ; CI-NEXT: buffer_load_dwordx2 v[1:2], v[1:2], s[0:3], 0 addr64
239 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
240 ; CI-NEXT: s_mov_b32 m0, -1
241 ; CI-NEXT: s_waitcnt vmcnt(0)
242 ; CI-NEXT: ds_write2_b32 v0, v1, v2 offset1:8
243 ; CI-NEXT: s_endpgm
269 ; CI-LABEL: simple_write2_two_val_subreg4_f32:
270 ; CI: ; %bb.0:
271 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x2
272 ; CI-NEXT: s_mov_b32 s3, 0xf000
273 ; CI-NEXT: s_mov_b32 s2, 0
274 ; CI-NEXT: v_lshlrev_b32_e32 v1, 4, v0
275 ; CI-NEXT: v_mov_b32_e32 v2, 0
276 ; CI-NEXT: s_waitcnt lgkmcnt(0)
277 ; CI-NEXT: buffer_load_dwordx4 v[1:4], v[1:2], s[0:3], 0 addr64
278 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
279 ; CI-NEXT: s_mov_b32 m0, -1
280 ; CI-NEXT: s_waitcnt vmcnt(0)
281 ; CI-NEXT: ds_write2_b32 v0, v1, v4 offset1:8
282 ; CI-NEXT: s_endpgm
308 ; CI-LABEL: simple_write2_two_val_max_offset_f32:
309 ; CI: ; %bb.0:
310 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x2
311 ; CI-NEXT: s_mov_b32 s3, 0xf000
312 ; CI-NEXT: s_mov_b32 s2, 0
313 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
314 ; CI-NEXT: v_mov_b32_e32 v1, 0
315 ; CI-NEXT: s_waitcnt lgkmcnt(0)
316 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[0:3], 0 addr64 glc
317 ; CI-NEXT: s_waitcnt vmcnt(0)
318 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[0:3], 0 addr64 offset:4 glc
319 ; CI-NEXT: s_waitcnt vmcnt(0)
320 ; CI-NEXT: s_mov_b32 m0, -1
321 ; CI-NEXT: ds_write2_b32 v0, v2, v1 offset1:255
322 ; CI-NEXT: s_endpgm
349 ; CI-LABEL: simple_write2_two_val_too_far_f32:
350 ; CI: ; %bb.0:
351 ; CI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x2
352 ; CI-NEXT: s_mov_b32 s7, 0xf000
353 ; CI-NEXT: s_mov_b32 s6, 0
354 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
355 ; CI-NEXT: v_mov_b32_e32 v1, 0
356 ; CI-NEXT: s_waitcnt lgkmcnt(0)
357 ; CI-NEXT: s_mov_b64 s[4:5], s[0:1]
358 ; CI-NEXT: s_mov_b64 s[0:1], s[2:3]
359 ; CI-NEXT: s_mov_b64 s[2:3], s[6:7]
360 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[4:7], 0 addr64
361 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[0:3], 0 addr64
362 ; CI-NEXT: s_mov_b32 m0, -1
363 ; CI-NEXT: s_waitcnt vmcnt(1)
364 ; CI-NEXT: ds_write_b32 v0, v2
365 ; CI-NEXT: s_waitcnt vmcnt(0)
366 ; CI-NEXT: ds_write_b32 v0, v1 offset:1028
367 ; CI-NEXT: s_endpgm
395 ; CI-LABEL: simple_write2_two_val_f32_x2:
396 ; CI: ; %bb.0:
397 ; CI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x2
398 ; CI-NEXT: s_mov_b32 s7, 0xf000
399 ; CI-NEXT: s_mov_b32 s6, 0
400 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
401 ; CI-NEXT: v_mov_b32_e32 v1, 0
402 ; CI-NEXT: s_waitcnt lgkmcnt(0)
403 ; CI-NEXT: s_mov_b64 s[4:5], s[0:1]
404 ; CI-NEXT: s_mov_b64 s[0:1], s[2:3]
405 ; CI-NEXT: s_mov_b64 s[2:3], s[6:7]
406 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[4:7], 0 addr64
407 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[0:3], 0 addr64
408 ; CI-NEXT: s_mov_b32 m0, -1
409 ; CI-NEXT: s_waitcnt vmcnt(0)
410 ; CI-NEXT: ds_write2_b32 v0, v2, v1 offset1:8
411 ; CI-NEXT: ds_write2_b32 v0, v2, v1 offset0:11 offset1:27
412 ; CI-NEXT: s_endpgm
451 ; CI-LABEL: simple_write2_two_val_f32_x2_nonzero_base:
452 ; CI: ; %bb.0:
453 ; CI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x2
454 ; CI-NEXT: s_mov_b32 s7, 0xf000
455 ; CI-NEXT: s_mov_b32 s6, 0
456 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
457 ; CI-NEXT: v_mov_b32_e32 v1, 0
458 ; CI-NEXT: s_waitcnt lgkmcnt(0)
459 ; CI-NEXT: s_mov_b64 s[4:5], s[0:1]
460 ; CI-NEXT: s_mov_b64 s[0:1], s[2:3]
461 ; CI-NEXT: s_mov_b64 s[2:3], s[6:7]
462 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[4:7], 0 addr64
463 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[0:3], 0 addr64
464 ; CI-NEXT: s_mov_b32 m0, -1
465 ; CI-NEXT: s_waitcnt vmcnt(0)
466 ; CI-NEXT: ds_write2_b32 v0, v2, v1 offset0:3 offset1:8
467 ; CI-NEXT: ds_write2_b32 v0, v2, v1 offset0:11 offset1:27
468 ; CI-NEXT: s_endpgm
507 ; CI-LABEL: write2_ptr_subreg_arg_two_val_f32:
508 ; CI: ; %bb.0:
509 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x2
510 ; CI-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x6
511 ; CI-NEXT: s_mov_b32 s3, 0xf000
512 ; CI-NEXT: s_mov_b32 s2, 0
513 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
514 ; CI-NEXT: s_waitcnt lgkmcnt(0)
515 ; CI-NEXT: s_mov_b64 s[0:1], s[4:5]
516 ; CI-NEXT: v_mov_b32_e32 v1, 0
517 ; CI-NEXT: s_mov_b64 s[4:5], s[6:7]
518 ; CI-NEXT: s_mov_b64 s[6:7], s[2:3]
519 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[0:3], 0 addr64
520 ; CI-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64
521 ; CI-NEXT: v_mov_b32_e32 v1, s8
522 ; CI-NEXT: s_mov_b32 m0, -1
523 ; CI-NEXT: v_mov_b32_e32 v3, s9
524 ; CI-NEXT: s_waitcnt vmcnt(1)
525 ; CI-NEXT: ds_write_b32 v1, v2 offset:32
526 ; CI-NEXT: s_waitcnt vmcnt(0)
527 ; CI-NEXT: ds_write_b32 v3, v0 offset:32
528 ; CI-NEXT: s_endpgm
567 ; CI-LABEL: simple_write2_one_val_f64:
568 ; CI: ; %bb.0:
569 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x2
570 ; CI-NEXT: s_mov_b32 s3, 0xf000
571 ; CI-NEXT: s_mov_b32 s2, 0
572 ; CI-NEXT: v_lshlrev_b32_e32 v0, 3, v0
573 ; CI-NEXT: v_mov_b32_e32 v1, 0
574 ; CI-NEXT: s_waitcnt lgkmcnt(0)
575 ; CI-NEXT: buffer_load_dwordx2 v[1:2], v[0:1], s[0:3], 0 addr64
576 ; CI-NEXT: s_mov_b32 m0, -1
577 ; CI-NEXT: s_waitcnt vmcnt(0)
578 ; CI-NEXT: ds_write2_b64 v0, v[1:2], v[1:2] offset1:8
579 ; CI-NEXT: s_endpgm
602 ; CI-LABEL: misaligned_simple_write2_one_val_f64:
603 ; CI: ; %bb.0:
604 ; CI-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x2
605 ; CI-NEXT: s_load_dword s0, s[0:1], 0x4
606 ; CI-NEXT: s_mov_b32 s7, 0xf000
607 ; CI-NEXT: s_mov_b32 s6, 0
608 ; CI-NEXT: v_lshlrev_b32_e32 v0, 3, v0
609 ; CI-NEXT: v_mov_b32_e32 v1, 0
610 ; CI-NEXT: s_waitcnt lgkmcnt(0)
611 ; CI-NEXT: buffer_load_dwordx2 v[1:2], v[0:1], s[4:7], 0 addr64
612 ; CI-NEXT: v_add_i32_e32 v0, vcc, s0, v0
613 ; CI-NEXT: s_mov_b32 m0, -1
614 ; CI-NEXT: s_waitcnt vmcnt(0)
615 ; CI-NEXT: ds_write2_b32 v0, v1, v2 offset1:1
616 ; CI-NEXT: ds_write2_b32 v0, v1, v2 offset0:14 offset1:15
617 ; CI-NEXT: s_endpgm
643 ; CI-LABEL: unaligned_offset_simple_write2_one_val_f64:
644 ; CI: ; %bb.0:
645 ; CI-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x2
646 ; CI-NEXT: s_load_dword s0, s[0:1], 0x4
647 ; CI-NEXT: s_mov_b32 s7, 0xf000
648 ; CI-NEXT: s_mov_b32 s6, 0
649 ; CI-NEXT: v_lshlrev_b32_e32 v0, 3, v0
650 ; CI-NEXT: v_mov_b32_e32 v1, 0
651 ; CI-NEXT: s_waitcnt lgkmcnt(0)
652 ; CI-NEXT: buffer_load_dwordx2 v[1:2], v[0:1], s[4:7], 0 addr64
653 ; CI-NEXT: v_add_i32_e32 v0, vcc, s0, v0
654 ; CI-NEXT: s_mov_b32 m0, -1
655 ; CI-NEXT: s_waitcnt vmcnt(0)
656 ; CI-NEXT: v_lshrrev_b32_e32 v3, 24, v1
657 ; CI-NEXT: ds_write_b8 v0, v1 offset:5
658 ; CI-NEXT: v_lshrrev_b32_e32 v4, 16, v1
659 ; CI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
660 ; CI-NEXT: ds_write_b8 v0, v2 offset:13
661 ; CI-NEXT: ds_write_b8 v0, v1 offset:9
662 ; CI-NEXT: v_lshrrev_b32_e32 v1, 24, v2
663 ; CI-NEXT: v_lshrrev_b32_e32 v6, 16, v2
664 ; CI-NEXT: v_lshrrev_b32_e32 v2, 8, v2
665 ; CI-NEXT: ds_write_b8 v0, v3 offset:8
666 ; CI-NEXT: ds_write_b8 v0, v4 offset:7
667 ; CI-NEXT: ds_write_b8 v0, v5 offset:6
668 ; CI-NEXT: ds_write_b8 v0, v1 offset:16
669 ; CI-NEXT: ds_write_b8 v0, v6 offset:15
670 ; CI-NEXT: ds_write_b8 v0, v2 offset:14
671 ; CI-NEXT: ds_write_b8 v0, v3 offset:12
672 ; CI-NEXT: ds_write_b8 v0, v4 offset:11
673 ; CI-NEXT: ds_write_b8 v0, v5 offset:10
674 ; CI-NEXT: s_endpgm
730 ; CI-LABEL: simple_write2_two_val_f64:
731 ; CI: ; %bb.0:
732 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x2
733 ; CI-NEXT: s_mov_b32 s3, 0xf000
734 ; CI-NEXT: s_mov_b32 s2, 0
735 ; CI-NEXT: v_lshlrev_b32_e32 v0, 3, v0
736 ; CI-NEXT: v_mov_b32_e32 v1, 0
737 ; CI-NEXT: s_waitcnt lgkmcnt(0)
738 ; CI-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64 glc
739 ; CI-NEXT: s_waitcnt vmcnt(0)
740 ; CI-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[0:3], 0 addr64 offset:8 glc
741 ; CI-NEXT: s_waitcnt vmcnt(0)
742 ; CI-NEXT: s_mov_b32 m0, -1
743 ; CI-NEXT: ds_write2_b64 v0, v[2:3], v[4:5] offset1:8
744 ; CI-NEXT: s_endpgm
773 ; CI-LABEL: store_constant_adjacent_offsets:
774 ; CI: ; %bb.0:
775 ; CI-NEXT: v_mov_b32_e32 v0, 0x7b
776 ; CI-NEXT: v_mov_b32_e32 v1, v0
777 ; CI-NEXT: v_mov_b32_e32 v2, 0
778 ; CI-NEXT: s_mov_b32 m0, -1
779 ; CI-NEXT: ds_write_b64 v2, v[0:1]
780 ; CI-NEXT: s_endpgm
795 ; CI-LABEL: store_constant_disjoint_offsets:
796 ; CI: ; %bb.0:
797 ; CI-NEXT: v_mov_b32_e32 v0, 0x7b
798 ; CI-NEXT: v_mov_b32_e32 v1, 0
799 ; CI-NEXT: s_mov_b32 m0, -1
800 ; CI-NEXT: ds_write2_b32 v1, v0, v0 offset1:2
801 ; CI-NEXT: s_endpgm
817 ; CI-LABEL: store_misaligned64_constant_offsets:
818 ; CI: ; %bb.0:
819 ; CI-NEXT: v_mov_b32_e32 v0, 0x7b
820 ; CI-NEXT: v_mov_b32_e32 v1, 0
821 ; CI-NEXT: v_mov_b32_e32 v2, v0
822 ; CI-NEXT: v_mov_b32_e32 v3, v1
823 ; CI-NEXT: s_mov_b32 m0, -1
824 ; CI-NEXT: ds_write_b128 v1, v[0:3]
825 ; CI-NEXT: s_endpgm
843 ; CI-LABEL: store_misaligned64_constant_large_offsets:
844 ; CI: ; %bb.0:
845 ; CI-NEXT: s_mov_b64 s[0:1], 0x7b
846 ; CI-NEXT: v_mov_b32_e32 v0, s0
847 ; CI-NEXT: v_mov_b32_e32 v2, 0
848 ; CI-NEXT: v_mov_b32_e32 v1, s1
849 ; CI-NEXT: s_mov_b32 m0, -1
850 ; CI-NEXT: ds_write_b64 v2, v[0:1] offset:16384
851 ; CI-NEXT: ds_write_b64 v2, v[0:1] offset:32760
852 ; CI-NEXT: s_endpgm
872 ; CI-LABEL: write2_sgemm_sequence:
873 ; CI: ; %bb.0:
874 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x4
875 ; CI-NEXT: s_mov_b32 m0, -1
876 ; CI-NEXT: s_waitcnt lgkmcnt(0)
877 ; CI-NEXT: s_load_dword s0, s[0:1], 0x0
878 ; CI-NEXT: s_lshl_b32 s1, s2, 2
879 ; CI-NEXT: s_add_i32 s2, s1, 0xc20
880 ; CI-NEXT: s_addk_i32 s1, 0xc60
881 ; CI-NEXT: v_mov_b32_e32 v0, s2
882 ; CI-NEXT: s_waitcnt lgkmcnt(0)
883 ; CI-NEXT: v_mov_b32_e32 v2, s0
884 ; CI-NEXT: v_mov_b32_e32 v3, s0
885 ; CI-NEXT: ds_write2_b32 v0, v2, v3 offset1:1
886 ; CI-NEXT: v_mov_b32_e32 v0, s1
887 ; CI-NEXT: ds_write2_b32 v0, v2, v3 offset1:1
888 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v1
889 ; CI-NEXT: ds_write2_b32 v0, v2, v3 offset1:1
890 ; CI-NEXT: ds_write2_b32 v0, v2, v3 offset0:32 offset1:33
891 ; CI-NEXT: ds_write2_b32 v0, v2, v3 offset0:64 offset1:65
892 ; CI-NEXT: s_endpgm
949 ; CI-LABEL: simple_write2_v4f32_superreg_align4:
950 ; CI: ; %bb.0:
951 ; CI-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x2
952 ; CI-NEXT: s_load_dword s4, s[0:1], 0x0
953 ; CI-NEXT: v_lshlrev_b32_e32 v0, 4, v0
954 ; CI-NEXT: s_mov_b32 m0, -1
955 ; CI-NEXT: s_waitcnt lgkmcnt(0)
956 ; CI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x0
957 ; CI-NEXT: v_add_i32_e32 v0, vcc, s4, v0
958 ; CI-NEXT: s_waitcnt lgkmcnt(0)
959 ; CI-NEXT: v_mov_b32_e32 v1, s0
960 ; CI-NEXT: v_mov_b32_e32 v2, s1
961 ; CI-NEXT: v_mov_b32_e32 v3, s2
962 ; CI-NEXT: ds_write2_b32 v0, v1, v2 offset1:1
963 ; CI-NEXT: v_mov_b32_e32 v1, s3
964 ; CI-NEXT: ds_write2_b32 v0, v3, v1 offset0:2 offset1:3
965 ; CI-NEXT: s_endpgm
1009 ; CI-LABEL: write2_v2i32_align1_odd_offset:
1010 ; CI: ; %bb.0: ; %entry
1011 ; CI-NEXT: v_mov_b32_e32 v0, 0x7b
1012 ; CI-NEXT: v_mov_b32_e32 v1, 0
1013 ; CI-NEXT: s_mov_b32 m0, -1
1014 ; CI-NEXT: ds_write_b8 v1, v0 offset:65
1015 ; CI-NEXT: v_mov_b32_e32 v0, 1
1016 ; CI-NEXT: ds_write_b8 v1, v0 offset:70
1017 ; CI-NEXT: v_mov_b32_e32 v0, 0xc8
1018 ; CI-NEXT: ds_write_b8 v1, v0 offset:69
1019 ; CI-NEXT: ds_write_b8 v1, v1 offset:68
1020 ; CI-NEXT: ds_write_b8 v1, v1 offset:67
1021 ; CI-NEXT: ds_write_b8 v1, v1 offset:66
1022 ; CI-NEXT: ds_write_b8 v1, v1 offset:72
1023 ; CI-NEXT: ds_write_b8 v1, v1 offset:71
1024 ; CI-NEXT: s_endpgm