Lines Matching +full:0 +full:x3900
8 ; SI: ; %bb.0: ; %entry
9 ; SI-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x9
10 ; SI-NEXT: s_mov_b32 s3, 0xf000
12 ; SI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x11
14 ; SI-NEXT: s_waitcnt lgkmcnt(0)
28 ; SI-NEXT: buffer_load_ushort v0, off, s[16:19], 0 glc
29 ; SI-NEXT: s_waitcnt vmcnt(0)
30 ; SI-NEXT: buffer_load_ushort v1, off, s[20:23], 0 glc
31 ; SI-NEXT: s_waitcnt vmcnt(0)
32 ; SI-NEXT: buffer_load_ushort v2, off, s[12:15], 0 glc
33 ; SI-NEXT: s_waitcnt vmcnt(0)
34 ; SI-NEXT: buffer_load_ushort v3, off, s[4:7], 0 glc
35 ; SI-NEXT: s_waitcnt vmcnt(0)
45 ; SI-NEXT: buffer_store_short v0, off, s[0:3], 0
49 ; VI: ; %bb.0: ; %entry
50 ; VI-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x24
51 ; VI-NEXT: s_mov_b32 s3, 0xf000
53 ; VI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x44
55 ; VI-NEXT: s_waitcnt lgkmcnt(0)
69 ; VI-NEXT: buffer_load_ushort v0, off, s[16:19], 0 glc
70 ; VI-NEXT: s_waitcnt vmcnt(0)
71 ; VI-NEXT: buffer_load_ushort v1, off, s[20:23], 0 glc
72 ; VI-NEXT: s_waitcnt vmcnt(0)
73 ; VI-NEXT: buffer_load_ushort v2, off, s[12:15], 0 glc
74 ; VI-NEXT: s_waitcnt vmcnt(0)
75 ; VI-NEXT: buffer_load_ushort v3, off, s[4:7], 0 glc
76 ; VI-NEXT: s_waitcnt vmcnt(0)
81 ; VI-NEXT: buffer_store_short v0, off, s[0:3], 0
85 ; GFX11: ; %bb.0: ; %entry
86 ; GFX11-NEXT: s_clause 0x1
87 ; GFX11-NEXT: s_load_b256 s[8:15], s[4:5], 0x24
88 ; GFX11-NEXT: s_load_b64 s[0:1], s[4:5], 0x44
90 ; GFX11-NEXT: s_mov_b32 s7, 0x31016000
99 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
106 ; GFX11-NEXT: buffer_load_u16 v0, off, s[16:19], 0 glc dlc
107 ; GFX11-NEXT: s_waitcnt vmcnt(0)
108 ; GFX11-NEXT: buffer_load_u16 v1, off, s[20:23], 0 glc dlc
109 ; GFX11-NEXT: s_waitcnt vmcnt(0)
110 ; GFX11-NEXT: buffer_load_u16 v2, off, s[24:27], 0 glc dlc
111 ; GFX11-NEXT: s_waitcnt vmcnt(0)
112 ; GFX11-NEXT: buffer_load_u16 v3, off, s[0:3], 0 glc dlc
113 ; GFX11-NEXT: s_waitcnt vmcnt(0)
118 ; GFX11-NEXT: buffer_store_b16 v0, off, s[4:7], 0
139 ; SI: ; %bb.0: ; %entry
140 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
141 ; SI-NEXT: s_mov_b32 s11, 0xf000
145 ; SI-NEXT: s_waitcnt lgkmcnt(0)
156 ; SI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
157 ; SI-NEXT: s_waitcnt vmcnt(0)
158 ; SI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
159 ; SI-NEXT: s_waitcnt vmcnt(0)
160 ; SI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
161 ; SI-NEXT: s_waitcnt vmcnt(0)
170 ; SI-NEXT: buffer_store_short v0, off, s[8:11], 0
174 ; VI: ; %bb.0: ; %entry
175 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
176 ; VI-NEXT: s_mov_b32 s11, 0xf000
180 ; VI-NEXT: s_waitcnt lgkmcnt(0)
191 ; VI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
192 ; VI-NEXT: s_waitcnt vmcnt(0)
193 ; VI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
194 ; VI-NEXT: s_waitcnt vmcnt(0)
195 ; VI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
196 ; VI-NEXT: s_waitcnt vmcnt(0)
201 ; VI-NEXT: buffer_store_short v0, off, s[8:11], 0
205 ; GFX11: ; %bb.0: ; %entry
206 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
208 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
215 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
222 ; GFX11-NEXT: buffer_load_u16 v0, off, s[12:15], 0 glc dlc
223 ; GFX11-NEXT: s_waitcnt vmcnt(0)
224 ; GFX11-NEXT: buffer_load_u16 v1, off, s[16:19], 0 glc dlc
225 ; GFX11-NEXT: s_waitcnt vmcnt(0)
226 ; GFX11-NEXT: buffer_load_u16 v2, off, s[20:23], 0 glc dlc
227 ; GFX11-NEXT: s_waitcnt vmcnt(0)
232 ; GFX11-NEXT: buffer_store_b16 v0, off, s[8:11], 0
242 %fcmp = fcmp olt half 0xH3800, %b.val
250 ; SI: ; %bb.0: ; %entry
251 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
252 ; SI-NEXT: s_mov_b32 s11, 0xf000
256 ; SI-NEXT: s_waitcnt lgkmcnt(0)
267 ; SI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
268 ; SI-NEXT: s_waitcnt vmcnt(0)
269 ; SI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
270 ; SI-NEXT: s_waitcnt vmcnt(0)
271 ; SI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
272 ; SI-NEXT: s_waitcnt vmcnt(0)
281 ; SI-NEXT: buffer_store_short v0, off, s[8:11], 0
285 ; VI: ; %bb.0: ; %entry
286 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
287 ; VI-NEXT: s_mov_b32 s11, 0xf000
291 ; VI-NEXT: s_waitcnt lgkmcnt(0)
302 ; VI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
303 ; VI-NEXT: s_waitcnt vmcnt(0)
304 ; VI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
305 ; VI-NEXT: s_waitcnt vmcnt(0)
306 ; VI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
307 ; VI-NEXT: s_waitcnt vmcnt(0)
312 ; VI-NEXT: buffer_store_short v0, off, s[8:11], 0
316 ; GFX11: ; %bb.0: ; %entry
317 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
319 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
326 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
333 ; GFX11-NEXT: buffer_load_u16 v0, off, s[12:15], 0 glc dlc
334 ; GFX11-NEXT: s_waitcnt vmcnt(0)
335 ; GFX11-NEXT: buffer_load_u16 v1, off, s[16:19], 0 glc dlc
336 ; GFX11-NEXT: s_waitcnt vmcnt(0)
337 ; GFX11-NEXT: buffer_load_u16 v2, off, s[20:23], 0 glc dlc
338 ; GFX11-NEXT: s_waitcnt vmcnt(0)
343 ; GFX11-NEXT: buffer_store_b16 v0, off, s[8:11], 0
353 %fcmp = fcmp olt half %a.val, 0xH3800
361 ; SI: ; %bb.0: ; %entry
362 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
363 ; SI-NEXT: s_mov_b32 s11, 0xf000
367 ; SI-NEXT: s_waitcnt lgkmcnt(0)
378 ; SI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
379 ; SI-NEXT: s_waitcnt vmcnt(0)
380 ; SI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
381 ; SI-NEXT: s_waitcnt vmcnt(0)
382 ; SI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
383 ; SI-NEXT: s_waitcnt vmcnt(0)
392 ; SI-NEXT: buffer_store_short v0, off, s[8:11], 0
396 ; VI: ; %bb.0: ; %entry
397 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
398 ; VI-NEXT: s_mov_b32 s11, 0xf000
402 ; VI-NEXT: s_waitcnt lgkmcnt(0)
413 ; VI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
414 ; VI-NEXT: s_waitcnt vmcnt(0)
415 ; VI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
416 ; VI-NEXT: s_waitcnt vmcnt(0)
417 ; VI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
418 ; VI-NEXT: s_waitcnt vmcnt(0)
419 ; VI-NEXT: v_mov_b32_e32 v3, 0x3800
424 ; VI-NEXT: buffer_store_short v0, off, s[8:11], 0
428 ; GFX11: ; %bb.0: ; %entry
429 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
431 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
438 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
445 ; GFX11-NEXT: buffer_load_u16 v0, off, s[12:15], 0 glc dlc
446 ; GFX11-NEXT: s_waitcnt vmcnt(0)
447 ; GFX11-NEXT: buffer_load_u16 v1, off, s[16:19], 0 glc dlc
448 ; GFX11-NEXT: s_waitcnt vmcnt(0)
449 ; GFX11-NEXT: buffer_load_u16 v2, off, s[20:23], 0 glc dlc
450 ; GFX11-NEXT: s_waitcnt vmcnt(0)
454 ; GFX11-NEXT: v_cndmask_b32_e32 v0, 0x3800, v2, vcc_lo
455 ; GFX11-NEXT: buffer_store_b16 v0, off, s[8:11], 0
466 %r.val = select i1 %fcmp, half 0xH3800, half %d.val
473 ; SI: ; %bb.0: ; %entry
474 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
475 ; SI-NEXT: s_mov_b32 s11, 0xf000
479 ; SI-NEXT: s_waitcnt lgkmcnt(0)
490 ; SI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
491 ; SI-NEXT: s_waitcnt vmcnt(0)
492 ; SI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
493 ; SI-NEXT: s_waitcnt vmcnt(0)
494 ; SI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
495 ; SI-NEXT: s_waitcnt vmcnt(0)
504 ; SI-NEXT: buffer_store_short v0, off, s[8:11], 0
508 ; VI: ; %bb.0: ; %entry
509 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
510 ; VI-NEXT: s_mov_b32 s11, 0xf000
514 ; VI-NEXT: s_waitcnt lgkmcnt(0)
525 ; VI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
526 ; VI-NEXT: s_waitcnt vmcnt(0)
527 ; VI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
528 ; VI-NEXT: s_waitcnt vmcnt(0)
529 ; VI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
530 ; VI-NEXT: s_waitcnt vmcnt(0)
531 ; VI-NEXT: v_mov_b32_e32 v3, 0x3800
536 ; VI-NEXT: buffer_store_short v0, off, s[8:11], 0
540 ; GFX11: ; %bb.0: ; %entry
541 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
543 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
550 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
557 ; GFX11-NEXT: buffer_load_u16 v0, off, s[12:15], 0 glc dlc
558 ; GFX11-NEXT: s_waitcnt vmcnt(0)
559 ; GFX11-NEXT: buffer_load_u16 v1, off, s[16:19], 0 glc dlc
560 ; GFX11-NEXT: s_waitcnt vmcnt(0)
561 ; GFX11-NEXT: buffer_load_u16 v2, off, s[20:23], 0 glc dlc
562 ; GFX11-NEXT: s_waitcnt vmcnt(0)
566 ; GFX11-NEXT: v_cndmask_b32_e32 v0, 0x3800, v2, vcc_lo
567 ; GFX11-NEXT: buffer_store_b16 v0, off, s[8:11], 0
578 %r.val = select i1 %fcmp, half %c.val, half 0xH3800
585 ; SI: ; %bb.0: ; %entry
586 ; SI-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x9
587 ; SI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x11
588 ; SI-NEXT: s_mov_b32 s3, 0xf000
591 ; SI-NEXT: s_waitcnt lgkmcnt(0)
601 ; SI-NEXT: buffer_load_dword v1, off, s[16:19], 0
604 ; SI-NEXT: buffer_load_dword v0, off, s[4:7], 0
607 ; SI-NEXT: buffer_load_dword v2, off, s[20:23], 0
608 ; SI-NEXT: buffer_load_dword v3, off, s[12:15], 0
620 ; SI-NEXT: s_waitcnt vmcnt(0)
635 ; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
639 ; VI: ; %bb.0: ; %entry
640 ; VI-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x24
641 ; VI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x44
642 ; VI-NEXT: s_mov_b32 s3, 0xf000
646 ; VI-NEXT: s_waitcnt lgkmcnt(0)
658 ; VI-NEXT: buffer_load_dword v0, off, s[4:7], 0
659 ; VI-NEXT: buffer_load_dword v1, off, s[20:23], 0
660 ; VI-NEXT: buffer_load_dword v2, off, s[16:19], 0
662 ; VI-NEXT: buffer_load_dword v3, off, s[12:15], 0
672 ; VI-NEXT: s_waitcnt vmcnt(0)
679 ; VI-NEXT: buffer_store_dword v0, off, s[0:3], 0
683 ; GFX11: ; %bb.0: ; %entry
684 ; GFX11-NEXT: s_clause 0x1
685 ; GFX11-NEXT: s_load_b256 s[8:15], s[4:5], 0x24
686 ; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x44
688 ; GFX11-NEXT: s_mov_b32 s3, 0x31016000
697 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
704 ; GFX11-NEXT: buffer_load_b32 v0, off, s[4:7], 0
705 ; GFX11-NEXT: buffer_load_b32 v1, off, s[20:23], 0
706 ; GFX11-NEXT: buffer_load_b32 v2, off, s[16:19], 0
707 ; GFX11-NEXT: buffer_load_b32 v3, off, s[24:27], 0
717 ; GFX11-NEXT: s_waitcnt vmcnt(0)
722 ; GFX11-NEXT: v_dual_cndmask_b32 v1, v4, v7 :: v_dual_and_b32 v0, 0xffff, v0
724 ; GFX11-NEXT: buffer_store_b32 v0, off, s[0:3], 0
744 ; SI: ; %bb.0: ; %entry
745 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
746 ; SI-NEXT: s_mov_b32 s11, 0xf000
750 ; SI-NEXT: s_waitcnt lgkmcnt(0)
757 ; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0
762 ; SI-NEXT: buffer_load_dword v1, off, s[16:19], 0
763 ; SI-NEXT: buffer_load_dword v2, off, s[4:7], 0
764 ; SI-NEXT: s_mov_b32 s2, 0x3f200000
773 ; SI-NEXT: s_waitcnt vmcnt(0)
787 ; SI-NEXT: buffer_store_dword v0, off, s[8:11], 0
791 ; VI: ; %bb.0: ; %entry
792 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
793 ; VI-NEXT: s_mov_b32 s11, 0xf000
797 ; VI-NEXT: s_waitcnt lgkmcnt(0)
808 ; VI-NEXT: buffer_load_dword v0, off, s[12:15], 0
809 ; VI-NEXT: buffer_load_dword v1, off, s[16:19], 0
810 ; VI-NEXT: buffer_load_dword v2, off, s[4:7], 0
811 ; VI-NEXT: s_movk_i32 s2, 0x3900
817 ; VI-NEXT: s_waitcnt vmcnt(0)
825 ; VI-NEXT: buffer_store_dword v0, off, s[8:11], 0
829 ; GFX11: ; %bb.0: ; %entry
830 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
832 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
839 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
846 ; GFX11-NEXT: buffer_load_b32 v0, off, s[12:15], 0
847 ; GFX11-NEXT: buffer_load_b32 v1, off, s[16:19], 0
848 ; GFX11-NEXT: buffer_load_b32 v2, off, s[20:23], 0
856 ; GFX11-NEXT: s_waitcnt vmcnt(0)
859 ; GFX11-NEXT: v_cmp_lt_f16_e32 vcc_lo, 0x3900, v3
862 ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
865 ; GFX11-NEXT: buffer_store_b32 v0, off, s[8:11], 0
875 %fcmp = fcmp olt <2 x half> <half 0xH3800, half 0xH3900>, %b.val
883 ; SI: ; %bb.0: ; %entry
884 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
885 ; SI-NEXT: s_mov_b32 s11, 0xf000
889 ; SI-NEXT: s_waitcnt lgkmcnt(0)
896 ; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0
901 ; SI-NEXT: buffer_load_dword v1, off, s[16:19], 0
902 ; SI-NEXT: buffer_load_dword v2, off, s[4:7], 0
903 ; SI-NEXT: s_mov_b32 s2, 0x3f200000
912 ; SI-NEXT: s_waitcnt vmcnt(0)
926 ; SI-NEXT: buffer_store_dword v0, off, s[8:11], 0
930 ; VI: ; %bb.0: ; %entry
931 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
932 ; VI-NEXT: s_mov_b32 s11, 0xf000
936 ; VI-NEXT: s_waitcnt lgkmcnt(0)
947 ; VI-NEXT: buffer_load_dword v0, off, s[12:15], 0
948 ; VI-NEXT: buffer_load_dword v1, off, s[16:19], 0
949 ; VI-NEXT: buffer_load_dword v2, off, s[4:7], 0
950 ; VI-NEXT: s_movk_i32 s2, 0x3900
956 ; VI-NEXT: s_waitcnt vmcnt(0)
964 ; VI-NEXT: buffer_store_dword v0, off, s[8:11], 0
968 ; GFX11: ; %bb.0: ; %entry
969 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
971 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
978 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
985 ; GFX11-NEXT: buffer_load_b32 v0, off, s[12:15], 0
986 ; GFX11-NEXT: buffer_load_b32 v1, off, s[16:19], 0
987 ; GFX11-NEXT: buffer_load_b32 v2, off, s[20:23], 0
995 ; GFX11-NEXT: s_waitcnt vmcnt(0)
998 ; GFX11-NEXT: v_cmp_gt_f16_e32 vcc_lo, 0x3900, v3
1001 ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
1004 ; GFX11-NEXT: buffer_store_b32 v0, off, s[8:11], 0
1014 %fcmp = fcmp olt <2 x half> %a.val, <half 0xH3800, half 0xH3900>
1022 ; SI: ; %bb.0: ; %entry
1023 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
1024 ; SI-NEXT: s_mov_b32 s11, 0xf000
1028 ; SI-NEXT: s_waitcnt lgkmcnt(0)
1035 ; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0
1040 ; SI-NEXT: buffer_load_dword v1, off, s[16:19], 0
1041 ; SI-NEXT: buffer_load_dword v2, off, s[4:7], 0
1042 ; SI-NEXT: v_mov_b32_e32 v3, 0x3f200000
1051 ; SI-NEXT: s_waitcnt vmcnt(0)
1065 ; SI-NEXT: buffer_store_dword v0, off, s[8:11], 0
1069 ; VI: ; %bb.0: ; %entry
1070 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
1071 ; VI-NEXT: s_mov_b32 s11, 0xf000
1075 ; VI-NEXT: s_waitcnt lgkmcnt(0)
1085 ; VI-NEXT: buffer_load_dword v0, off, s[16:19], 0
1086 ; VI-NEXT: buffer_load_dword v1, off, s[12:15], 0
1088 ; VI-NEXT: buffer_load_dword v2, off, s[4:7], 0
1089 ; VI-NEXT: v_mov_b32_e32 v3, 0x3800
1090 ; VI-NEXT: v_mov_b32_e32 v4, 0x3900
1098 ; VI-NEXT: s_waitcnt vmcnt(0)
1105 ; VI-NEXT: buffer_store_dword v0, off, s[8:11], 0
1109 ; GFX11: ; %bb.0: ; %entry
1110 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
1112 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
1119 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
1126 ; GFX11-NEXT: buffer_load_b32 v0, off, s[16:19], 0
1127 ; GFX11-NEXT: buffer_load_b32 v1, off, s[12:15], 0
1128 ; GFX11-NEXT: buffer_load_b32 v2, off, s[20:23], 0
1136 ; GFX11-NEXT: s_waitcnt vmcnt(0)
1138 ; GFX11-NEXT: v_cndmask_b32_e32 v0, 0x3800, v2, vcc_lo
1141 ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
1142 ; GFX11-NEXT: v_cndmask_b32_e32 v1, 0x3900, v5, vcc_lo
1145 ; GFX11-NEXT: buffer_store_b32 v0, off, s[8:11], 0
1156 %r.val = select <2 x i1> %fcmp, <2 x half> <half 0xH3800, half 0xH3900>, <2 x half> %d.val
1163 ; SI: ; %bb.0: ; %entry
1164 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
1165 ; SI-NEXT: s_mov_b32 s11, 0xf000
1169 ; SI-NEXT: s_waitcnt lgkmcnt(0)
1176 ; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0
1181 ; SI-NEXT: buffer_load_dword v1, off, s[16:19], 0
1182 ; SI-NEXT: buffer_load_dword v2, off, s[4:7], 0
1183 ; SI-NEXT: v_mov_b32_e32 v3, 0x3f200000
1192 ; SI-NEXT: s_waitcnt vmcnt(0)
1206 ; SI-NEXT: buffer_store_dword v0, off, s[8:11], 0
1210 ; VI: ; %bb.0: ; %entry
1211 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
1212 ; VI-NEXT: s_mov_b32 s11, 0xf000
1216 ; VI-NEXT: s_waitcnt lgkmcnt(0)
1226 ; VI-NEXT: buffer_load_dword v0, off, s[16:19], 0
1227 ; VI-NEXT: buffer_load_dword v1, off, s[12:15], 0
1229 ; VI-NEXT: buffer_load_dword v2, off, s[4:7], 0
1230 ; VI-NEXT: v_mov_b32_e32 v3, 0x3800
1231 ; VI-NEXT: v_mov_b32_e32 v4, 0x3900
1239 ; VI-NEXT: s_waitcnt vmcnt(0)
1246 ; VI-NEXT: buffer_store_dword v0, off, s[8:11], 0
1250 ; GFX11: ; %bb.0: ; %entry
1251 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
1253 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
1260 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
1267 ; GFX11-NEXT: buffer_load_b32 v0, off, s[16:19], 0
1268 ; GFX11-NEXT: buffer_load_b32 v1, off, s[12:15], 0
1269 ; GFX11-NEXT: buffer_load_b32 v2, off, s[20:23], 0
1277 ; GFX11-NEXT: s_waitcnt vmcnt(0)
1279 ; GFX11-NEXT: v_cndmask_b32_e32 v0, 0x3800, v2, vcc_lo
1282 ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
1283 ; GFX11-NEXT: v_cndmask_b32_e32 v1, 0x3900, v5, vcc_lo
1286 ; GFX11-NEXT: buffer_store_b32 v0, off, s[8:11], 0
1297 %r.val = select <2 x i1> %fcmp, <2 x half> %c.val, <2 x half> <half 0xH3800, half 0xH3900>
1304 ; SI: ; %bb.0:
1305 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1322 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v8
1334 ; VI: ; %bb.0:
1335 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1336 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v4
1342 ; GFX11: ; %bb.0:
1343 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1344 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v4
1347 %cmp = icmp eq i32 %cond, 0
1354 ; SI: ; %bb.0:
1355 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1372 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v8
1374 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v9
1376 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v10
1378 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v11
1383 ; VI: ; %bb.0:
1384 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1387 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v7
1391 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v5
1393 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v6
1395 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v4
1404 ; GFX11: ; %bb.0:
1405 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1408 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v7
1413 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v5
1415 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v4
1417 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v6
1419 ; GFX11-NEXT: v_perm_b32 v0, v5, v0, 0x5040100
1421 ; GFX11-NEXT: v_perm_b32 v1, v7, v1, 0x5040100
1430 ; SI: ; %bb.0:
1431 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1464 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v16
1484 ; VI: ; %bb.0:
1485 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1486 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v8
1494 ; GFX11: ; %bb.0:
1495 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1496 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v8
1500 %cmp = icmp eq i32 %cond, 0
1507 ; SI: ; %bb.0:
1508 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1511 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v16
1527 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v17
1541 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v18
1545 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v19
1549 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v20
1551 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v21
1553 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v22
1555 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v23
1560 ; VI: ; %bb.0:
1561 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1564 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v15
1568 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v13
1572 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v11
1576 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v9
1578 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v14
1580 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v12
1582 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v10
1584 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v8
1597 ; GFX11: ; %bb.0:
1598 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1601 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v15
1609 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v13
1612 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v11
1614 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v9
1616 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v12
1618 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v8
1620 ; GFX11-NEXT: v_perm_b32 v2, v13, v2, 0x5040100
1622 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v10
1623 ; GFX11-NEXT: v_perm_b32 v0, v9, v0, 0x5040100
1625 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v14
1627 ; GFX11-NEXT: v_perm_b32 v1, v11, v1, 0x5040100
1629 ; GFX11-NEXT: v_perm_b32 v3, v15, v3, 0x5040100
1638 ; SI: ; %bb.0:
1639 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1655 ; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32
1663 ; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:4
1707 ; SI-NEXT: s_waitcnt vmcnt(0)
1708 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v26
1744 ; VI: ; %bb.0:
1745 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1746 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v16
1758 ; GFX11: ; %bb.0:
1759 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1760 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v16
1766 %cmp = icmp eq i32 %cond, 0
1773 ; SI: ; %bb.0:
1774 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1775 ; SI-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:4
1781 ; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:8
1782 ; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:12
1783 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
1784 ; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:20
1785 ; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:24
1786 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:28
1787 ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:32
1820 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v36
1823 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v38
1827 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v35
1829 ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:48
1835 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v32
1836 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
1837 ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:44
1841 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v33
1842 ; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:40
1847 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v34
1851 ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:56
1853 ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:52
1855 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
1858 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v16
1859 ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:60
1860 ; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32
1865 ; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:64
1869 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v32
1874 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v33
1877 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v19
1881 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v20
1886 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v17
1889 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v18
1896 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v16
1898 ; SI-NEXT: s_waitcnt vmcnt(0)
1899 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v22
1904 ; VI: ; %bb.0:
1905 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1907 ; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
1909 ; VI-NEXT: v_writelane_b32 v31, s30, 0
1911 ; VI-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v16
1912 ; VI-NEXT: v_cmp_eq_u32_e64 s[18:19], 0, v17
1913 ; VI-NEXT: v_cmp_eq_u32_e64 s[30:31], 0, v29
1916 ; VI-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v18
1917 ; VI-NEXT: v_cmp_eq_u32_e64 s[28:29], 0, v27
1921 ; VI-NEXT: v_cmp_eq_u32_e64 s[20:21], 0, v19
1922 ; VI-NEXT: v_cmp_eq_u32_e64 s[26:27], 0, v25
1926 ; VI-NEXT: v_cmp_eq_u32_e64 s[8:9], 0, v20
1927 ; VI-NEXT: v_cmp_eq_u32_e64 s[24:25], 0, v23
1931 ; VI-NEXT: v_cmp_eq_u32_e64 s[22:23], 0, v21
1935 ; VI-NEXT: v_cmp_eq_u32_e64 s[10:11], 0, v22
1943 ; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32
1945 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v30
1953 ; VI-NEXT: v_cmp_eq_u32_e64 s[12:13], 0, v24
1955 ; VI-NEXT: v_cmp_eq_u32_e64 s[14:15], 0, v26
1957 ; VI-NEXT: v_cmp_eq_u32_e64 s[16:17], 0, v28
1961 ; VI-NEXT: v_readlane_b32 s30, v31, 0
1962 ; VI-NEXT: s_waitcnt vmcnt(0)
1963 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v8
1980 ; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
1982 ; VI-NEXT: s_waitcnt vmcnt(0)
1986 ; GFX11: ; %bb.0:
1987 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1989 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v30
1995 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v28
2001 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v26
2007 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v24
2013 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v22
2015 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v20
2017 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v18
2019 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v16
2021 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v29
2023 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v27
2025 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v25
2027 ; GFX11-NEXT: v_perm_b32 v6, v8, v6, 0x5040100
2028 ; GFX11-NEXT: v_perm_b32 v5, v9, v5, 0x5040100
2030 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v19
2032 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v17
2034 ; GFX11-NEXT: v_perm_b32 v4, v10, v4, 0x5040100
2035 ; GFX11-NEXT: v_perm_b32 v1, v11, v1, 0x5040100
2037 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v21
2039 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v23
2042 ; GFX11-NEXT: v_perm_b32 v2, v13, v2, 0x5040100
2043 ; GFX11-NEXT: v_perm_b32 v3, v14, v3, 0x5040100
2044 ; GFX11-NEXT: s_waitcnt vmcnt(0)
2045 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v31
2047 ; GFX11-NEXT: v_perm_b32 v0, v12, v0, 0x5040100
2049 ; GFX11-NEXT: v_perm_b32 v7, v11, v7, 0x5040100
2058 ; SI: ; %bb.0:
2059 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2066 ; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32
2090 ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:56
2092 ; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:48
2094 ; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:40
2096 ; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:32
2098 ; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24
2100 ; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16
2102 ; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8
2127 ; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:124
2130 ; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:116
2131 ; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:108
2132 ; SI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:100
2133 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:92
2134 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:84
2135 ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:80
2136 ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72
2137 ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:64
2142 ; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:128
2174 ; SI-NEXT: s_waitcnt vmcnt(0)
2178 ; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:120
2179 ; SI-NEXT: s_waitcnt vmcnt(0)
2183 ; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:112
2184 ; SI-NEXT: s_waitcnt vmcnt(0)
2188 ; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:104
2189 ; SI-NEXT: s_waitcnt vmcnt(0)
2193 ; SI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:96
2194 ; SI-NEXT: s_waitcnt vmcnt(0)
2198 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:88
2199 ; SI-NEXT: s_waitcnt vmcnt(0)
2203 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76
2206 ; SI-NEXT: s_waitcnt vmcnt(0)
2209 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
2212 ; SI-NEXT: s_waitcnt vmcnt(0)
2215 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
2218 ; SI-NEXT: s_waitcnt vmcnt(0)
2221 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:52
2222 ; SI-NEXT: s_waitcnt vmcnt(0)
2225 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
2226 ; SI-NEXT: s_waitcnt vmcnt(0)
2229 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
2230 ; SI-NEXT: s_waitcnt vmcnt(0)
2233 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:28
2234 ; SI-NEXT: s_waitcnt vmcnt(0)
2237 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:20
2238 ; SI-NEXT: s_waitcnt vmcnt(0)
2241 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:12
2242 ; SI-NEXT: s_waitcnt vmcnt(0)
2245 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
2246 ; SI-NEXT: s_waitcnt vmcnt(0)
2249 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:132
2250 ; SI-NEXT: s_waitcnt vmcnt(0)
2251 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v32
2319 ; VI: ; %bb.0:
2320 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2321 ; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
2322 ; VI-NEXT: s_waitcnt vmcnt(0)
2323 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2325 ; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32
2340 ; VI-NEXT: s_waitcnt vmcnt(0)
2345 ; GFX11: ; %bb.0:
2346 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2347 ; GFX11-NEXT: s_clause 0x1
2351 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v31
2359 ; GFX11-NEXT: s_waitcnt vmcnt(0)
2362 %cmp = icmp eq i32 %cond, 0
2369 ; SI: ; %bb.0:
2370 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2371 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:132
2431 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
2436 ; SI-NEXT: v_cmp_eq_u32_e64 s[16:17], 0, v31
2437 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:136
2441 ; SI-NEXT: s_waitcnt vmcnt(0)
2442 ; SI-NEXT: v_cmp_eq_u32_e64 s[14:15], 0, v31
2443 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:140
2444 ; SI-NEXT: s_waitcnt vmcnt(0)
2445 ; SI-NEXT: v_cmp_eq_u32_e64 s[12:13], 0, v31
2446 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:144
2447 ; SI-NEXT: s_waitcnt vmcnt(0)
2448 ; SI-NEXT: v_cmp_eq_u32_e64 s[10:11], 0, v31
2449 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:148
2450 ; SI-NEXT: s_waitcnt vmcnt(0)
2451 ; SI-NEXT: v_cmp_eq_u32_e64 s[8:9], 0, v31
2452 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:152
2453 ; SI-NEXT: s_waitcnt vmcnt(0)
2454 ; SI-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v31
2455 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:156
2456 ; SI-NEXT: s_waitcnt vmcnt(0)
2457 ; SI-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v31
2458 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:160
2459 ; SI-NEXT: s_waitcnt vmcnt(0)
2460 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2461 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
2462 ; SI-NEXT: s_waitcnt vmcnt(0)
2466 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8
2467 ; SI-NEXT: s_waitcnt vmcnt(0)
2471 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:12
2472 ; SI-NEXT: s_waitcnt vmcnt(0)
2476 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:16
2477 ; SI-NEXT: s_waitcnt vmcnt(0)
2481 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:20
2482 ; SI-NEXT: s_waitcnt vmcnt(0)
2486 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:24
2487 ; SI-NEXT: s_waitcnt vmcnt(0)
2491 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:28
2492 ; SI-NEXT: s_waitcnt vmcnt(0)
2496 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:32
2497 ; SI-NEXT: s_waitcnt vmcnt(0)
2501 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:164
2502 ; SI-NEXT: s_waitcnt vmcnt(0)
2503 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2504 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:36
2505 ; SI-NEXT: s_waitcnt vmcnt(0)
2509 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:168
2510 ; SI-NEXT: s_waitcnt vmcnt(0)
2511 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2512 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:40
2513 ; SI-NEXT: s_waitcnt vmcnt(0)
2517 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:172
2518 ; SI-NEXT: s_waitcnt vmcnt(0)
2519 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2520 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:44
2521 ; SI-NEXT: s_waitcnt vmcnt(0)
2525 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:176
2526 ; SI-NEXT: s_waitcnt vmcnt(0)
2527 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2528 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:48
2529 ; SI-NEXT: s_waitcnt vmcnt(0)
2533 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:180
2534 ; SI-NEXT: s_waitcnt vmcnt(0)
2535 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2536 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:52
2537 ; SI-NEXT: s_waitcnt vmcnt(0)
2541 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:184
2542 ; SI-NEXT: s_waitcnt vmcnt(0)
2543 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2544 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:56
2545 ; SI-NEXT: s_waitcnt vmcnt(0)
2549 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:188
2550 ; SI-NEXT: s_waitcnt vmcnt(0)
2551 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2552 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:60
2553 ; SI-NEXT: s_waitcnt vmcnt(0)
2557 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:192
2558 ; SI-NEXT: s_waitcnt vmcnt(0)
2559 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2560 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:64
2561 ; SI-NEXT: s_waitcnt vmcnt(0)
2565 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:196
2566 ; SI-NEXT: s_waitcnt vmcnt(0)
2567 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2568 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:68
2569 ; SI-NEXT: s_waitcnt vmcnt(0)
2573 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:200
2574 ; SI-NEXT: s_waitcnt vmcnt(0)
2575 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2576 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:72
2577 ; SI-NEXT: s_waitcnt vmcnt(0)
2581 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:204
2582 ; SI-NEXT: s_waitcnt vmcnt(0)
2583 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2584 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:76
2585 ; SI-NEXT: s_waitcnt vmcnt(0)
2589 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:208
2590 ; SI-NEXT: s_waitcnt vmcnt(0)
2591 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2592 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:80
2593 ; SI-NEXT: s_waitcnt vmcnt(0)
2597 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:212
2598 ; SI-NEXT: s_waitcnt vmcnt(0)
2599 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2600 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:84
2601 ; SI-NEXT: s_waitcnt vmcnt(0)
2605 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:216
2606 ; SI-NEXT: s_waitcnt vmcnt(0)
2607 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2608 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:88
2609 ; SI-NEXT: s_waitcnt vmcnt(0)
2613 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:220
2614 ; SI-NEXT: s_waitcnt vmcnt(0)
2615 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2616 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:92
2617 ; SI-NEXT: s_waitcnt vmcnt(0)
2621 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:224
2622 ; SI-NEXT: s_waitcnt vmcnt(0)
2623 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2624 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:96
2625 ; SI-NEXT: s_waitcnt vmcnt(0)
2629 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:228
2630 ; SI-NEXT: s_waitcnt vmcnt(0)
2631 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2632 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:100
2633 ; SI-NEXT: s_waitcnt vmcnt(0)
2637 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:232
2638 ; SI-NEXT: s_waitcnt vmcnt(0)
2639 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2640 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:104
2641 ; SI-NEXT: s_waitcnt vmcnt(0)
2645 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:236
2646 ; SI-NEXT: s_waitcnt vmcnt(0)
2647 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2648 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:108
2649 ; SI-NEXT: s_waitcnt vmcnt(0)
2653 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:240
2654 ; SI-NEXT: s_waitcnt vmcnt(0)
2655 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2656 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:112
2657 ; SI-NEXT: s_waitcnt vmcnt(0)
2661 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:244
2662 ; SI-NEXT: s_waitcnt vmcnt(0)
2663 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2664 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:116
2665 ; SI-NEXT: s_waitcnt vmcnt(0)
2669 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:248
2670 ; SI-NEXT: s_waitcnt vmcnt(0)
2671 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2672 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:120
2673 ; SI-NEXT: s_waitcnt vmcnt(0)
2677 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:252
2678 ; SI-NEXT: s_waitcnt vmcnt(0)
2679 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2680 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:124
2681 ; SI-NEXT: s_waitcnt vmcnt(0)
2685 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:256
2686 ; SI-NEXT: s_waitcnt vmcnt(0)
2687 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2688 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32
2689 ; SI-NEXT: s_waitcnt vmcnt(0)
2696 ; VI: ; %bb.0:
2697 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2698 ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
2699 ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
2700 ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
2701 ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
2702 ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
2703 ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
2704 ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
2705 ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
2706 ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
2707 ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
2708 ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
2709 ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
2710 ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
2711 ; VI-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:120
2712 ; VI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:112
2713 ; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:104
2714 ; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:96
2715 ; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
2716 ; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:80
2717 ; VI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:72
2718 ; VI-NEXT: buffer_load_dword v37, off, s[0:3], s32
2719 ; VI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:128
2720 ; VI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:64
2721 ; VI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:56
2722 ; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:48
2723 ; VI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:40
2724 ; VI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:32
2725 ; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:24
2726 ; VI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:16
2727 ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8
2728 ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:124
2729 ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:116
2742 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v36
2744 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v35
2746 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v34
2748 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v33
2750 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v32
2754 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2759 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v39
2765 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v55
2770 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v50
2775 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v53
2780 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v54
2785 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v52
2790 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v51
2794 ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:108
2796 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v49
2798 ; VI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:100
2799 ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:92
2802 ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:84
2803 ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:76
2805 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v48
2807 ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:68
2808 ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:60
2809 ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:52
2813 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v40
2814 ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44
2816 ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:36
2818 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v41
2819 ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:28
2821 ; VI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:20
2823 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v42
2824 ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:12
2826 ; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:4
2828 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v44
2831 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v55
2834 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v45
2837 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v47
2840 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v56
2843 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v43
2846 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v57
2849 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v59
2852 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v40
2855 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v58
2858 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v41
2861 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v37
2864 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v42
2866 ; VI-NEXT: s_waitcnt vmcnt(0)
2867 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v30
2870 ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
2871 ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
2872 ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
2873 ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
2874 ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
2875 ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
2876 ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
2877 ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
2878 ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
2879 ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
2880 ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
2881 ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
2882 ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
2914 ; VI-NEXT: s_waitcnt vmcnt(0)
2918 ; GFX11: ; %bb.0:
2919 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2920 ; GFX11-NEXT: s_clause 0x1f
2983 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v31
2987 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v32
2992 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v34
2996 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v35
2999 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v36
3002 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v37
3005 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v38
3008 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v39
3011 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v48
3014 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v49
3017 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v50
3020 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v51
3023 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v52
3026 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v53
3029 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v54
3032 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v55
3035 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v64
3038 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v65
3040 ; GFX11-NEXT: v_perm_b32 v14, v97, v14, 0x5040100
3043 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v66
3046 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v67
3048 ; GFX11-NEXT: v_perm_b32 v12, v34, v12, 0x5040100
3051 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v68
3054 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v69
3055 ; GFX11-NEXT: v_perm_b32 v13, v99, v13, 0x5040100
3057 ; GFX11-NEXT: v_perm_b32 v10, v36, v10, 0x5040100
3060 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v70
3063 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v71
3064 ; GFX11-NEXT: v_perm_b32 v11, v35, v11, 0x5040100
3066 ; GFX11-NEXT: v_perm_b32 v8, v38, v8, 0x5040100
3069 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v80
3072 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v81
3073 ; GFX11-NEXT: v_perm_b32 v9, v37, v9, 0x5040100
3075 ; GFX11-NEXT: v_perm_b32 v6, v48, v6, 0x5040100
3078 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v82
3081 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v83
3082 ; GFX11-NEXT: v_perm_b32 v7, v39, v7, 0x5040100
3084 ; GFX11-NEXT: v_perm_b32 v4, v50, v4, 0x5040100
3087 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v84
3090 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v85
3093 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v86
3094 ; GFX11-NEXT: v_perm_b32 v5, v49, v5, 0x5040100
3096 ; GFX11-NEXT: v_perm_b32 v0, v32, v0, 0x5040100
3098 ; GFX11-NEXT: s_waitcnt vmcnt(0)
3099 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v87
3100 ; GFX11-NEXT: v_perm_b32 v3, v51, v3, 0x5040100
3101 ; GFX11-NEXT: v_perm_b32 v2, v52, v2, 0x5040100
3103 ; GFX11-NEXT: v_perm_b32 v1, v31, v1, 0x5040100
3105 ; GFX11-NEXT: v_perm_b32 v15, v16, v15, 0x5040100