Lines Matching full:ci
2 …y-machineinstrs -mattr=+load-store-opt < %s | FileCheck -enable-var-scope -check-prefixes=GCN,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], 0xb
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], 0xb
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
54 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[0:3], 0 addr64 offset:4
55 ; CI-NEXT: s_mov_b32 m0, -1
56 ; CI-NEXT: s_waitcnt vmcnt(0)
57 ; CI-NEXT: ds_write2_b32 v0, v2, v1 offset1:8
58 ; CI-NEXT: s_endpgm
84 ; CI-LABEL: simple_write2_two_val_f32_volatile_0:
85 ; CI: ; %bb.0:
86 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0xb
87 ; CI-NEXT: s_mov_b32 s3, 0xf000
88 ; CI-NEXT: s_mov_b32 s2, 0
89 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
90 ; CI-NEXT: v_mov_b32_e32 v1, 0
91 ; CI-NEXT: s_waitcnt lgkmcnt(0)
92 ; CI-NEXT: s_mov_b64 s[0:1], s[4:5]
93 ; CI-NEXT: s_mov_b64 s[4:5], s[6:7]
94 ; CI-NEXT: s_mov_b64 s[6:7], s[2:3]
95 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[0:3], 0 addr64
96 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[4:7], 0 addr64
97 ; CI-NEXT: s_mov_b32 m0, -1
98 ; CI-NEXT: s_waitcnt vmcnt(1)
99 ; CI-NEXT: ds_write_b32 v0, v2
100 ; CI-NEXT: s_waitcnt vmcnt(0)
101 ; CI-NEXT: ds_write_b32 v0, v1 offset:32
102 ; CI-NEXT: s_endpgm
130 ; CI-LABEL: simple_write2_two_val_f32_volatile_1:
131 ; CI: ; %bb.0:
132 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0xb
133 ; CI-NEXT: s_mov_b32 s3, 0xf000
134 ; CI-NEXT: s_mov_b32 s2, 0
135 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
136 ; CI-NEXT: v_mov_b32_e32 v1, 0
137 ; CI-NEXT: s_waitcnt lgkmcnt(0)
138 ; CI-NEXT: s_mov_b64 s[0:1], s[4:5]
139 ; CI-NEXT: s_mov_b64 s[4:5], s[6:7]
140 ; CI-NEXT: s_mov_b64 s[6:7], s[2:3]
141 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[0:3], 0 addr64
142 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[4:7], 0 addr64
143 ; CI-NEXT: s_mov_b32 m0, -1
144 ; CI-NEXT: s_waitcnt vmcnt(1)
145 ; CI-NEXT: ds_write_b32 v0, v2
146 ; CI-NEXT: s_waitcnt vmcnt(0)
147 ; CI-NEXT: ds_write_b32 v0, v1 offset:32
148 ; CI-NEXT: s_endpgm
181 ; CI-LABEL: simple_write2_two_val_subreg2_mixed_f32:
182 ; CI: ; %bb.0:
183 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
184 ; CI-NEXT: v_lshlrev_b32_e32 v1, 3, v0
185 ; CI-NEXT: s_mov_b32 s3, 0xf000
186 ; CI-NEXT: s_mov_b32 s2, 0
187 ; CI-NEXT: v_mov_b32_e32 v2, 0
188 ; CI-NEXT: s_waitcnt lgkmcnt(0)
189 ; CI-NEXT: buffer_load_dwordx2 v[3:4], v[1:2], s[0:3], 0 addr64
190 ; CI-NEXT: buffer_load_dwordx2 v[1:2], v[1:2], s[0:3], 0 addr64 offset:8
191 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
192 ; CI-NEXT: s_mov_b32 m0, -1
193 ; CI-NEXT: s_waitcnt vmcnt(0)
194 ; CI-NEXT: ds_write2_b32 v0, v3, v2 offset1:8
195 ; CI-NEXT: s_endpgm
225 ; CI-LABEL: simple_write2_two_val_subreg2_f32:
226 ; CI: ; %bb.0:
227 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
228 ; CI-NEXT: v_lshlrev_b32_e32 v1, 3, v0
229 ; CI-NEXT: s_mov_b32 s3, 0xf000
230 ; CI-NEXT: s_mov_b32 s2, 0
231 ; CI-NEXT: v_mov_b32_e32 v2, 0
232 ; CI-NEXT: s_waitcnt lgkmcnt(0)
233 ; CI-NEXT: buffer_load_dwordx2 v[1:2], v[1:2], s[0:3], 0 addr64
234 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
235 ; CI-NEXT: s_mov_b32 m0, -1
236 ; CI-NEXT: s_waitcnt vmcnt(0)
237 ; CI-NEXT: ds_write2_b32 v0, v1, v2 offset1:8
238 ; CI-NEXT: s_endpgm
264 ; CI-LABEL: simple_write2_two_val_subreg4_f32:
265 ; CI: ; %bb.0:
266 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
267 ; CI-NEXT: v_lshlrev_b32_e32 v1, 4, v0
268 ; CI-NEXT: s_mov_b32 s3, 0xf000
269 ; CI-NEXT: s_mov_b32 s2, 0
270 ; CI-NEXT: v_mov_b32_e32 v2, 0
271 ; CI-NEXT: s_waitcnt lgkmcnt(0)
272 ; CI-NEXT: buffer_load_dwordx4 v[1:4], v[1:2], s[0:3], 0 addr64
273 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
274 ; CI-NEXT: s_mov_b32 m0, -1
275 ; CI-NEXT: s_waitcnt vmcnt(0)
276 ; CI-NEXT: ds_write2_b32 v0, v1, v4 offset1:8
277 ; CI-NEXT: s_endpgm
303 ; CI-LABEL: simple_write2_two_val_max_offset_f32:
304 ; CI: ; %bb.0:
305 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
306 ; CI-NEXT: s_mov_b32 s3, 0xf000
307 ; CI-NEXT: s_mov_b32 s2, 0
308 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
309 ; CI-NEXT: v_mov_b32_e32 v1, 0
310 ; CI-NEXT: s_waitcnt lgkmcnt(0)
311 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[0:3], 0 addr64
312 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[0:3], 0 addr64 offset:4
313 ; CI-NEXT: s_mov_b32 m0, -1
314 ; CI-NEXT: s_waitcnt vmcnt(0)
315 ; CI-NEXT: ds_write2_b32 v0, v2, v1 offset1:255
316 ; CI-NEXT: s_endpgm
342 ; CI-LABEL: simple_write2_two_val_too_far_f32:
343 ; CI: ; %bb.0:
344 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0xb
345 ; CI-NEXT: s_mov_b32 s3, 0xf000
346 ; CI-NEXT: s_mov_b32 s2, 0
347 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
348 ; CI-NEXT: v_mov_b32_e32 v1, 0
349 ; CI-NEXT: s_waitcnt lgkmcnt(0)
350 ; CI-NEXT: s_mov_b64 s[0:1], s[4:5]
351 ; CI-NEXT: s_mov_b64 s[4:5], s[6:7]
352 ; CI-NEXT: s_mov_b64 s[6:7], s[2:3]
353 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[0:3], 0 addr64
354 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[4:7], 0 addr64
355 ; CI-NEXT: s_mov_b32 m0, -1
356 ; CI-NEXT: s_waitcnt vmcnt(1)
357 ; CI-NEXT: ds_write_b32 v0, v2
358 ; CI-NEXT: s_waitcnt vmcnt(0)
359 ; CI-NEXT: ds_write_b32 v0, v1 offset:1028
360 ; CI-NEXT: s_endpgm
388 ; CI-LABEL: simple_write2_two_val_f32_x2:
389 ; CI: ; %bb.0:
390 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0xb
391 ; CI-NEXT: s_mov_b32 s3, 0xf000
392 ; CI-NEXT: s_mov_b32 s2, 0
393 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
394 ; CI-NEXT: v_mov_b32_e32 v1, 0
395 ; CI-NEXT: s_waitcnt lgkmcnt(0)
396 ; CI-NEXT: s_mov_b64 s[0:1], s[4:5]
397 ; CI-NEXT: s_mov_b64 s[4:5], s[6:7]
398 ; CI-NEXT: s_mov_b64 s[6:7], s[2:3]
399 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[0:3], 0 addr64
400 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[4:7], 0 addr64
401 ; CI-NEXT: s_mov_b32 m0, -1
402 ; CI-NEXT: s_waitcnt vmcnt(0)
403 ; CI-NEXT: ds_write2_b32 v0, v2, v1 offset1:8
404 ; CI-NEXT: ds_write2_b32 v0, v2, v1 offset0:11 offset1:27
405 ; CI-NEXT: s_endpgm
444 ; CI-LABEL: simple_write2_two_val_f32_x2_nonzero_base:
445 ; CI: ; %bb.0:
446 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0xb
447 ; CI-NEXT: s_mov_b32 s3, 0xf000
448 ; CI-NEXT: s_mov_b32 s2, 0
449 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
450 ; CI-NEXT: v_mov_b32_e32 v1, 0
451 ; CI-NEXT: s_waitcnt lgkmcnt(0)
452 ; CI-NEXT: s_mov_b64 s[0:1], s[4:5]
453 ; CI-NEXT: s_mov_b64 s[4:5], s[6:7]
454 ; CI-NEXT: s_mov_b64 s[6:7], s[2:3]
455 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[0:3], 0 addr64
456 ; CI-NEXT: buffer_load_dword v1, v[0:1], s[4:7], 0 addr64
457 ; CI-NEXT: s_mov_b32 m0, -1
458 ; CI-NEXT: s_waitcnt vmcnt(0)
459 ; CI-NEXT: ds_write2_b32 v0, v2, v1 offset0:3 offset1:8
460 ; CI-NEXT: ds_write2_b32 v0, v2, v1 offset0:11 offset1:27
461 ; CI-NEXT: s_endpgm
500 ; CI-LABEL: write2_ptr_subreg_arg_two_val_f32:
501 ; CI: ; %bb.0:
502 ; CI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0xb
503 ; CI-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0xf
504 ; CI-NEXT: s_mov_b32 s3, 0xf000
505 ; CI-NEXT: s_mov_b32 s2, 0
506 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v0
507 ; CI-NEXT: s_waitcnt lgkmcnt(0)
508 ; CI-NEXT: s_mov_b64 s[0:1], s[4:5]
509 ; CI-NEXT: v_mov_b32_e32 v1, 0
510 ; CI-NEXT: s_mov_b64 s[4:5], s[6:7]
511 ; CI-NEXT: s_mov_b64 s[6:7], s[2:3]
512 ; CI-NEXT: buffer_load_dword v2, v[0:1], s[0:3], 0 addr64
513 ; CI-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64
514 ; CI-NEXT: v_mov_b32_e32 v1, s8
515 ; CI-NEXT: s_mov_b32 m0, -1
516 ; CI-NEXT: v_mov_b32_e32 v3, s9
517 ; CI-NEXT: s_waitcnt vmcnt(1)
518 ; CI-NEXT: ds_write_b32 v1, v2 offset:32
519 ; CI-NEXT: s_waitcnt vmcnt(0)
520 ; CI-NEXT: ds_write_b32 v3, v0 offset:32
521 ; CI-NEXT: s_endpgm
560 ; CI-LABEL: simple_write2_one_val_f64:
561 ; CI: ; %bb.0:
562 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
563 ; CI-NEXT: s_mov_b32 s3, 0xf000
564 ; CI-NEXT: s_mov_b32 s2, 0
565 ; CI-NEXT: v_lshlrev_b32_e32 v0, 3, v0
566 ; CI-NEXT: v_mov_b32_e32 v1, 0
567 ; CI-NEXT: s_waitcnt lgkmcnt(0)
568 ; CI-NEXT: buffer_load_dwordx2 v[1:2], v[0:1], s[0:3], 0 addr64
569 ; CI-NEXT: s_mov_b32 m0, -1
570 ; CI-NEXT: s_waitcnt vmcnt(0)
571 ; CI-NEXT: ds_write2_b64 v0, v[1:2], v[1:2] offset1:8
572 ; CI-NEXT: s_endpgm
595 ; CI-LABEL: misaligned_simple_write2_one_val_f64:
596 ; CI: ; %bb.0:
597 ; CI-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0xb
598 ; CI-NEXT: s_load_dword s0, s[0:1], 0xd
599 ; CI-NEXT: s_mov_b32 s7, 0xf000
600 ; CI-NEXT: s_mov_b32 s6, 0
601 ; CI-NEXT: v_lshlrev_b32_e32 v0, 3, v0
602 ; CI-NEXT: v_mov_b32_e32 v1, 0
603 ; CI-NEXT: s_waitcnt lgkmcnt(0)
604 ; CI-NEXT: buffer_load_dwordx2 v[1:2], v[0:1], s[4:7], 0 addr64
605 ; CI-NEXT: v_add_i32_e32 v0, vcc, s0, v0
606 ; CI-NEXT: s_mov_b32 m0, -1
607 ; CI-NEXT: s_waitcnt vmcnt(0)
608 ; CI-NEXT: ds_write2_b32 v0, v1, v2 offset1:1
609 ; CI-NEXT: ds_write2_b32 v0, v1, v2 offset0:14 offset1:15
610 ; CI-NEXT: s_endpgm
636 ; CI-LABEL: unaligned_offset_simple_write2_one_val_f64:
637 ; CI: ; %bb.0:
638 ; CI-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0xb
639 ; CI-NEXT: s_load_dword s0, s[0:1], 0xd
640 ; CI-NEXT: s_mov_b32 s7, 0xf000
641 ; CI-NEXT: s_mov_b32 s6, 0
642 ; CI-NEXT: v_lshlrev_b32_e32 v0, 3, v0
643 ; CI-NEXT: v_mov_b32_e32 v1, 0
644 ; CI-NEXT: s_waitcnt lgkmcnt(0)
645 ; CI-NEXT: buffer_load_dwordx2 v[1:2], v[0:1], s[4:7], 0 addr64
646 ; CI-NEXT: v_add_i32_e32 v0, vcc, s0, v0
647 ; CI-NEXT: s_mov_b32 m0, -1
648 ; CI-NEXT: s_waitcnt vmcnt(0)
649 ; CI-NEXT: v_lshrrev_b32_e32 v3, 24, v1
650 ; CI-NEXT: ds_write_b8 v0, v1 offset:5
651 ; CI-NEXT: v_lshrrev_b32_e32 v4, 16, v1
652 ; CI-NEXT: v_lshrrev_b32_e32 v5, 8, v1
653 ; CI-NEXT: ds_write_b8 v0, v2 offset:13
654 ; CI-NEXT: ds_write_b8 v0, v1 offset:9
655 ; CI-NEXT: v_lshrrev_b32_e32 v1, 24, v2
656 ; CI-NEXT: v_lshrrev_b32_e32 v6, 16, v2
657 ; CI-NEXT: v_lshrrev_b32_e32 v2, 8, v2
658 ; CI-NEXT: ds_write_b8 v0, v3 offset:8
659 ; CI-NEXT: ds_write_b8 v0, v4 offset:7
660 ; CI-NEXT: ds_write_b8 v0, v5 offset:6
661 ; CI-NEXT: ds_write_b8 v0, v1 offset:16
662 ; CI-NEXT: ds_write_b8 v0, v6 offset:15
663 ; CI-NEXT: ds_write_b8 v0, v2 offset:14
664 ; CI-NEXT: ds_write_b8 v0, v3 offset:12
665 ; CI-NEXT: ds_write_b8 v0, v4 offset:11
666 ; CI-NEXT: ds_write_b8 v0, v5 offset:10
667 ; CI-NEXT: s_endpgm
725 ; CI-LABEL: simple_write2_two_val_f64:
726 ; CI: ; %bb.0:
727 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
728 ; CI-NEXT: s_mov_b32 s3, 0xf000
729 ; CI-NEXT: s_mov_b32 s2, 0
730 ; CI-NEXT: v_lshlrev_b32_e32 v0, 3, v0
731 ; CI-NEXT: v_mov_b32_e32 v1, 0
732 ; CI-NEXT: s_waitcnt lgkmcnt(0)
733 ; CI-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64
734 ; CI-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[0:3], 0 addr64 offset:8
735 ; CI-NEXT: s_mov_b32 m0, -1
736 ; CI-NEXT: s_waitcnt vmcnt(0)
737 ; CI-NEXT: ds_write2_b64 v0, v[2:3], v[4:5] offset1:8
738 ; CI-NEXT: s_endpgm
766 ; CI-LABEL: store_constant_adjacent_offsets:
767 ; CI: ; %bb.0:
768 ; CI-NEXT: s_movk_i32 s0, 0x7b
769 ; CI-NEXT: v_mov_b32_e32 v0, 0
770 ; CI-NEXT: v_mov_b32_e32 v1, s0
771 ; CI-NEXT: v_mov_b32_e32 v2, s0
772 ; CI-NEXT: s_mov_b32 m0, -1
773 ; CI-NEXT: ds_write2_b32 v0, v1, v2 offset1:1
774 ; CI-NEXT: s_endpgm
790 ; CI-LABEL: store_constant_disjoint_offsets:
791 ; CI: ; %bb.0:
792 ; CI-NEXT: v_mov_b32_e32 v0, 0x7b
793 ; CI-NEXT: v_mov_b32_e32 v1, 0
794 ; CI-NEXT: s_mov_b32 m0, -1
795 ; CI-NEXT: ds_write2_b32 v1, v0, v0 offset1:2
796 ; CI-NEXT: s_endpgm
812 ; CI-LABEL: store_misaligned64_constant_offsets:
813 ; CI: ; %bb.0:
814 ; CI-NEXT: v_mov_b32_e32 v0, 0
815 ; CI-NEXT: v_mov_b32_e32 v1, 0x7b
816 ; CI-NEXT: s_mov_b32 m0, -1
817 ; CI-NEXT: ds_write2_b32 v0, v1, v0 offset1:1
818 ; CI-NEXT: ds_write2_b32 v0, v1, v0 offset0:2 offset1:3
819 ; CI-NEXT: s_endpgm
845 ; CI-LABEL: store_misaligned64_constant_large_offsets:
846 ; CI: ; %bb.0:
847 ; CI-NEXT: v_mov_b32_e32 v0, 0x4000
848 ; CI-NEXT: v_mov_b32_e32 v1, 0x7b
849 ; CI-NEXT: v_mov_b32_e32 v2, 0
850 ; CI-NEXT: s_mov_b32 m0, -1
851 ; CI-NEXT: ds_write2_b32 v0, v1, v2 offset1:1
852 ; CI-NEXT: v_mov_b32_e32 v0, 0x7ff8
853 ; CI-NEXT: ds_write2_b32 v0, v1, v2 offset1:1
854 ; CI-NEXT: s_endpgm
874 ; CI-LABEL: write2_sgemm_sequence:
875 ; CI: ; %bb.0:
876 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
877 ; CI-NEXT: s_lshl_b32 s2, s2, 2
878 ; CI-NEXT: s_add_i32 s3, s2, 0xc20
879 ; CI-NEXT: v_mov_b32_e32 v0, s3
880 ; CI-NEXT: s_addk_i32 s2, 0xc60
881 ; CI-NEXT: s_waitcnt lgkmcnt(0)
882 ; CI-NEXT: s_load_dword s0, s[0:1], 0x0
883 ; CI-NEXT: s_mov_b32 m0, -1
884 ; CI-NEXT: s_waitcnt lgkmcnt(0)
885 ; CI-NEXT: v_mov_b32_e32 v2, s0
886 ; CI-NEXT: v_mov_b32_e32 v3, s0
887 ; CI-NEXT: ds_write2_b32 v0, v2, v3 offset1:1
888 ; CI-NEXT: v_mov_b32_e32 v0, s2
889 ; CI-NEXT: ds_write2_b32 v0, v2, v3 offset1:1
890 ; CI-NEXT: v_lshlrev_b32_e32 v0, 2, v1
891 ; CI-NEXT: ds_write2_b32 v0, v2, v3 offset1:1
892 ; CI-NEXT: ds_write2_b32 v0, v2, v3 offset0:32 offset1:33
893 ; CI-NEXT: ds_write2_b32 v0, v2, v3 offset0:64 offset1:65
894 ; CI-NEXT: s_endpgm
951 ; CI-LABEL: simple_write2_v4f32_superreg_align4:
952 ; CI: ; %bb.0:
953 ; CI-NEXT: s_load_dword s4, s[0:1], 0x9
954 ; CI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
955 ; CI-NEXT: v_lshlrev_b32_e32 v0, 4, v0
956 ; CI-NEXT: s_mov_b32 m0, -1
957 ; CI-NEXT: s_waitcnt lgkmcnt(0)
958 ; CI-NEXT: v_add_i32_e32 v0, vcc, s4, v0
959 ; CI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x0
960 ; CI-NEXT: s_waitcnt lgkmcnt(0)
961 ; CI-NEXT: v_mov_b32_e32 v1, s0
962 ; CI-NEXT: v_mov_b32_e32 v2, s1
963 ; CI-NEXT: ds_write2_b32 v0, v1, v2 offset1:1
964 ; CI-NEXT: v_mov_b32_e32 v3, s2
965 ; CI-NEXT: v_mov_b32_e32 v1, s3
966 ; CI-NEXT: ds_write2_b32 v0, v3, v1 offset0:2 offset1:3
967 ; CI-NEXT: s_endpgm
1010 ; CI-LABEL: write2_v2i32_align1_odd_offset:
1011 ; CI: ; %bb.0: ; %entry
1012 ; CI-NEXT: v_mov_b32_e32 v0, 0x7b
1013 ; CI-NEXT: v_mov_b32_e32 v1, 0
1014 ; CI-NEXT: s_mov_b32 m0, -1
1015 ; CI-NEXT: ds_write_b8 v1, v0 offset:65
1016 ; CI-NEXT: v_mov_b32_e32 v0, 1
1017 ; CI-NEXT: ds_write_b8 v1, v0 offset:70
1018 ; CI-NEXT: v_mov_b32_e32 v0, 0xc8
1019 ; CI-NEXT: ds_write_b8 v1, v0 offset:69
1020 ; CI-NEXT: ds_write_b8 v1, v1 offset:68
1021 ; CI-NEXT: ds_write_b8 v1, v1 offset:67
1022 ; CI-NEXT: ds_write_b8 v1, v1 offset:66
1023 ; CI-NEXT: ds_write_b8 v1, v1 offset:72
1024 ; CI-NEXT: ds_write_b8 v1, v1 offset:71
1025 ; CI-NEXT: s_endpgm