Lines Matching refs:SKIP
6 …-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
56 ; SKIP-CACHE-INV-LABEL: local_wavefront_unordered_load:
57 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
58 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
59 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
60 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
61 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
62 ; SKIP-CACHE-INV-NEXT: ds_read_b32 v0, v0
63 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
64 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
65 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v1, v0
66 ; SKIP-CACHE-INV-NEXT: s_endpgm
122 ; SKIP-CACHE-INV-LABEL: local_wavefront_monotonic_load:
123 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
124 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
125 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
126 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
127 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
128 ; SKIP-CACHE-INV-NEXT: ds_read_b32 v0, v0
129 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
130 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
131 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v1, v0
132 ; SKIP-CACHE-INV-NEXT: s_endpgm
188 ; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_load:
189 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
190 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
191 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
192 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
193 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
194 ; SKIP-CACHE-INV-NEXT: ds_read_b32 v0, v0
195 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
196 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
197 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v1, v0
198 ; SKIP-CACHE-INV-NEXT: s_endpgm
254 ; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_load:
255 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
256 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
257 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
258 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
259 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
260 ; SKIP-CACHE-INV-NEXT: ds_read_b32 v0, v0
261 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
262 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
263 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v1, v0
264 ; SKIP-CACHE-INV-NEXT: s_endpgm
312 ; SKIP-CACHE-INV-LABEL: local_wavefront_unordered_store:
313 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
314 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
315 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
316 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
317 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s1
318 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
319 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
320 ; SKIP-CACHE-INV-NEXT: s_endpgm
367 ; SKIP-CACHE-INV-LABEL: local_wavefront_monotonic_store:
368 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
369 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
370 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
371 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
372 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s1
373 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
374 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
375 ; SKIP-CACHE-INV-NEXT: s_endpgm
422 ; SKIP-CACHE-INV-LABEL: local_wavefront_release_store:
423 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
424 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
425 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
426 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
427 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s1
428 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
429 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
430 ; SKIP-CACHE-INV-NEXT: s_endpgm
477 ; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_store:
478 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
479 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
480 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
481 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
482 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s1
483 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
484 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
485 ; SKIP-CACHE-INV-NEXT: s_endpgm
532 ; SKIP-CACHE-INV-LABEL: local_wavefront_monotonic_atomicrmw:
533 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
534 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
535 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
536 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
537 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
538 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
539 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v0, v0, v1
540 ; SKIP-CACHE-INV-NEXT: s_endpgm
587 ; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_atomicrmw:
588 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
589 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
590 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
591 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
592 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
593 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
594 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v0, v0, v1
595 ; SKIP-CACHE-INV-NEXT: s_endpgm
642 ; SKIP-CACHE-INV-LABEL: local_wavefront_release_atomicrmw:
643 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
644 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
645 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
646 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
647 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
648 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
649 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v0, v0, v1
650 ; SKIP-CACHE-INV-NEXT: s_endpgm
697 ; SKIP-CACHE-INV-LABEL: local_wavefront_acq_rel_atomicrmw:
698 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
699 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
700 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
701 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
702 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
703 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
704 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v0, v0, v1
705 ; SKIP-CACHE-INV-NEXT: s_endpgm
752 ; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_atomicrmw:
753 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
754 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
755 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
756 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
757 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
758 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
759 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v0, v0, v1
760 ; SKIP-CACHE-INV-NEXT: s_endpgm
815 ; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_ret_atomicrmw:
816 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
817 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
818 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
819 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
820 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
821 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
822 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v1, v0, v1
823 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
824 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
825 ; SKIP-CACHE-INV-NEXT: s_endpgm
881 ; SKIP-CACHE-INV-LABEL: local_wavefront_acq_rel_ret_atomicrmw:
882 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
883 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
884 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
885 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
886 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
887 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
888 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v1, v0, v1
889 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
890 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
891 ; SKIP-CACHE-INV-NEXT: s_endpgm
947 ; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_ret_atomicrmw:
948 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
949 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
950 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
951 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
952 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
953 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
954 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v1, v0, v1
955 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
956 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
957 ; SKIP-CACHE-INV-NEXT: s_endpgm
1009 ; SKIP-CACHE-INV-LABEL: local_wavefront_monotonic_monotonic_cmpxchg:
1010 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1011 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1012 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1013 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1014 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1015 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1016 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1017 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1018 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
1019 ; SKIP-CACHE-INV-NEXT: s_endpgm
1071 ; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_monotonic_cmpxchg:
1072 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1073 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1074 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1075 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1076 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1077 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1078 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1079 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1080 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
1081 ; SKIP-CACHE-INV-NEXT: s_endpgm
1133 ; SKIP-CACHE-INV-LABEL: local_wavefront_release_monotonic_cmpxchg:
1134 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1135 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1136 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1137 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1138 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1139 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1140 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1141 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1142 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
1143 ; SKIP-CACHE-INV-NEXT: s_endpgm
1195 ; SKIP-CACHE-INV-LABEL: local_wavefront_acq_rel_monotonic_cmpxchg:
1196 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1197 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1198 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1199 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1200 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1201 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1202 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1203 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1204 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
1205 ; SKIP-CACHE-INV-NEXT: s_endpgm
1257 ; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_monotonic_cmpxchg:
1258 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1259 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1260 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1261 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1262 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1263 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1264 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1265 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1266 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
1267 ; SKIP-CACHE-INV-NEXT: s_endpgm
1319 ; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_acquire_cmpxchg:
1320 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1321 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1322 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1323 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1324 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1325 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1326 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1327 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1328 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
1329 ; SKIP-CACHE-INV-NEXT: s_endpgm
1381 ; SKIP-CACHE-INV-LABEL: local_wavefront_release_acquire_cmpxchg:
1382 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1383 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1384 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1385 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1386 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1387 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1388 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1389 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1390 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
1391 ; SKIP-CACHE-INV-NEXT: s_endpgm
1443 ; SKIP-CACHE-INV-LABEL: local_wavefront_acq_rel_acquire_cmpxchg:
1444 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1445 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1446 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1447 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1448 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1449 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1450 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1451 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1452 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
1453 ; SKIP-CACHE-INV-NEXT: s_endpgm
1505 ; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_acquire_cmpxchg:
1506 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1507 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1508 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1509 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1510 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1511 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1512 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1513 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1514 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
1515 ; SKIP-CACHE-INV-NEXT: s_endpgm
1567 ; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_seq_cst_cmpxchg:
1568 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1569 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1570 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1571 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1572 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1573 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1574 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1575 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1576 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
1577 ; SKIP-CACHE-INV-NEXT: s_endpgm
1637 ; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_monotonic_ret_cmpxchg:
1638 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1639 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1640 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1641 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1642 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1643 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1644 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1645 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1646 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
1647 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1648 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
1649 ; SKIP-CACHE-INV-NEXT: s_endpgm
1711 ; SKIP-CACHE-INV-LABEL: local_wavefront_acq_rel_monotonic_ret_cmpxchg:
1712 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1713 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1714 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1715 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1716 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1717 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1718 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1719 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1720 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
1721 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1722 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
1723 ; SKIP-CACHE-INV-NEXT: s_endpgm
1785 ; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_monotonic_ret_cmpxchg:
1786 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1787 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1788 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1789 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1790 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1791 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1792 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1793 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1794 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
1795 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1796 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
1797 ; SKIP-CACHE-INV-NEXT: s_endpgm
1859 ; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_acquire_ret_cmpxchg:
1860 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1861 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1862 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1863 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1864 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1865 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1866 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1867 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1868 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
1869 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1870 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
1871 ; SKIP-CACHE-INV-NEXT: s_endpgm
1933 ; SKIP-CACHE-INV-LABEL: local_wavefront_release_acquire_ret_cmpxchg:
1934 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1935 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1936 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1937 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
1938 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1939 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1940 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
1941 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
1942 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
1943 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1944 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
1945 ; SKIP-CACHE-INV-NEXT: s_endpgm
2007 ; SKIP-CACHE-INV-LABEL: local_wavefront_acq_rel_acquire_ret_cmpxchg:
2008 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2009 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
2010 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
2011 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2012 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2013 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
2014 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
2015 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
2016 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
2017 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2018 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
2019 ; SKIP-CACHE-INV-NEXT: s_endpgm
2081 ; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_acquire_ret_cmpxchg:
2082 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2083 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
2084 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
2085 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2086 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2087 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
2088 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
2089 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
2090 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
2091 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2092 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
2093 ; SKIP-CACHE-INV-NEXT: s_endpgm
2155 ; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_seq_cst_ret_cmpxchg:
2156 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2157 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
2158 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
2159 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2160 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2161 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
2162 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
2163 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
2164 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
2165 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2166 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
2167 ; SKIP-CACHE-INV-NEXT: s_endpgm
2225 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_unordered_load:
2226 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2227 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2228 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2229 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2230 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2231 ; SKIP-CACHE-INV-NEXT: ds_read_b32 v0, v0
2232 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2233 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2234 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v1, v0
2235 ; SKIP-CACHE-INV-NEXT: s_endpgm
2291 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_monotonic_load:
2292 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2293 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2294 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2295 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2296 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2297 ; SKIP-CACHE-INV-NEXT: ds_read_b32 v0, v0
2298 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2299 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2300 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v1, v0
2301 ; SKIP-CACHE-INV-NEXT: s_endpgm
2357 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_load:
2358 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2359 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2360 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2361 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2362 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2363 ; SKIP-CACHE-INV-NEXT: ds_read_b32 v0, v0
2364 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2365 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2366 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v1, v0
2367 ; SKIP-CACHE-INV-NEXT: s_endpgm
2423 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_load:
2424 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2425 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2426 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2427 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2428 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2429 ; SKIP-CACHE-INV-NEXT: ds_read_b32 v0, v0
2430 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2431 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2432 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v1, v0
2433 ; SKIP-CACHE-INV-NEXT: s_endpgm
2481 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_unordered_store:
2482 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2483 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2484 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2485 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2486 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s1
2487 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
2488 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
2489 ; SKIP-CACHE-INV-NEXT: s_endpgm
2536 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_monotonic_store:
2537 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2538 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2539 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2540 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2541 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s1
2542 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
2543 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
2544 ; SKIP-CACHE-INV-NEXT: s_endpgm
2591 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_release_store:
2592 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2593 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2594 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2595 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2596 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s1
2597 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
2598 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
2599 ; SKIP-CACHE-INV-NEXT: s_endpgm
2646 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_store:
2647 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2648 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2649 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2650 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2651 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s1
2652 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
2653 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
2654 ; SKIP-CACHE-INV-NEXT: s_endpgm
2701 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_monotonic_atomicrmw:
2702 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2703 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2704 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2705 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2706 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2707 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2708 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v0, v0, v1
2709 ; SKIP-CACHE-INV-NEXT: s_endpgm
2756 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_atomicrmw:
2757 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2758 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2759 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2760 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2761 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2762 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2763 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v0, v0, v1
2764 ; SKIP-CACHE-INV-NEXT: s_endpgm
2811 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_release_atomicrmw:
2812 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2813 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2814 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2815 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2816 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2817 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2818 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v0, v0, v1
2819 ; SKIP-CACHE-INV-NEXT: s_endpgm
2866 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acq_rel_atomicrmw:
2867 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2868 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2869 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2870 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2871 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2872 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2873 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v0, v0, v1
2874 ; SKIP-CACHE-INV-NEXT: s_endpgm
2921 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_atomicrmw:
2922 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2923 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2924 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2925 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2926 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2927 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2928 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v0, v0, v1
2929 ; SKIP-CACHE-INV-NEXT: s_endpgm
2984 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_ret_atomicrmw:
2985 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2986 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
2987 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
2988 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2989 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2990 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2991 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v1, v0, v1
2992 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2993 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
2994 ; SKIP-CACHE-INV-NEXT: s_endpgm
3050 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acq_rel_ret_atomicrmw:
3051 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3052 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
3053 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3054 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3055 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3056 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
3057 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v1, v0, v1
3058 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3059 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
3060 ; SKIP-CACHE-INV-NEXT: s_endpgm
3116 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_ret_atomicrmw:
3117 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3118 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x9
3119 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3120 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3121 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3122 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
3123 ; SKIP-CACHE-INV-NEXT: ds_wrxchg_rtn_b32 v1, v0, v1
3124 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3125 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
3126 ; SKIP-CACHE-INV-NEXT: s_endpgm
3178 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_monotonic_monotonic_cmpxchg:
3179 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3180 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3181 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3182 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3183 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3184 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3185 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
3186 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
3187 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
3188 ; SKIP-CACHE-INV-NEXT: s_endpgm
3240 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_monotonic_cmpxchg:
3241 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3242 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3243 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3244 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3245 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3246 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3247 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
3248 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
3249 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
3250 ; SKIP-CACHE-INV-NEXT: s_endpgm
3302 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_release_monotonic_cmpxchg:
3303 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3304 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3305 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3306 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3307 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3308 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3309 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
3310 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
3311 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
3312 ; SKIP-CACHE-INV-NEXT: s_endpgm
3364 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acq_rel_monotonic_cmpxchg:
3365 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3366 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3367 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3368 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3369 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3370 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3371 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
3372 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
3373 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
3374 ; SKIP-CACHE-INV-NEXT: s_endpgm
3426 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_monotonic_cmpxchg:
3427 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3428 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3429 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3430 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3431 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3432 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3433 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
3434 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
3435 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
3436 ; SKIP-CACHE-INV-NEXT: s_endpgm
3488 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_acquire_cmpxchg:
3489 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3490 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3491 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3492 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3493 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3494 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3495 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
3496 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
3497 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
3498 ; SKIP-CACHE-INV-NEXT: s_endpgm
3550 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_release_acquire_cmpxchg:
3551 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3552 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3553 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3554 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3555 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3556 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3557 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
3558 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
3559 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
3560 ; SKIP-CACHE-INV-NEXT: s_endpgm
3612 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acq_rel_acquire_cmpxchg:
3613 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3614 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3615 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3616 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3617 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3618 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3619 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
3620 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
3621 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
3622 ; SKIP-CACHE-INV-NEXT: s_endpgm
3674 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_acquire_cmpxchg:
3675 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3676 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3677 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3678 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3679 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3680 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3681 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
3682 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
3683 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
3684 ; SKIP-CACHE-INV-NEXT: s_endpgm
3736 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
3737 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3738 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3739 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3740 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3741 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3742 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3743 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
3744 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
3745 ; SKIP-CACHE-INV-NEXT: ds_cmpst_b32 v0, v1, v2 offset:16
3746 ; SKIP-CACHE-INV-NEXT: s_endpgm
3806 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
3807 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3808 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3809 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3810 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3811 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3812 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3813 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
3814 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
3815 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
3816 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3817 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
3818 ; SKIP-CACHE-INV-NEXT: s_endpgm
3880 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
3881 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3882 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3883 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3884 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3885 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3886 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3887 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
3888 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
3889 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
3890 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3891 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
3892 ; SKIP-CACHE-INV-NEXT: s_endpgm
3954 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
3955 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3956 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3957 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3958 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
3959 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3960 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3961 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
3962 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
3963 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
3964 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3965 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
3966 ; SKIP-CACHE-INV-NEXT: s_endpgm
4028 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_acquire_ret_cmpxchg:
4029 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4030 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
4031 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
4032 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
4033 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4034 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
4035 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
4036 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
4037 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
4038 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4039 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
4040 ; SKIP-CACHE-INV-NEXT: s_endpgm
4102 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_release_acquire_ret_cmpxchg:
4103 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4104 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
4105 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
4106 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
4107 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4108 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
4109 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
4110 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
4111 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
4112 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4113 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
4114 ; SKIP-CACHE-INV-NEXT: s_endpgm
4176 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
4177 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4178 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
4179 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
4180 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
4181 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4182 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
4183 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
4184 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
4185 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
4186 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4187 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
4188 ; SKIP-CACHE-INV-NEXT: s_endpgm
4250 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
4251 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4252 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
4253 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
4254 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
4255 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4256 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
4257 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
4258 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
4259 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
4260 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4261 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
4262 ; SKIP-CACHE-INV-NEXT: s_endpgm
4324 ; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
4325 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4326 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
4327 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
4328 ; SKIP-CACHE-INV-NEXT: s_mov_b32 m0, -1
4329 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4330 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
4331 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s0
4332 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s3
4333 ; SKIP-CACHE-INV-NEXT: ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
4334 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4335 ; SKIP-CACHE-INV-NEXT: ds_write_b32 v0, v1
4336 ; SKIP-CACHE-INV-NEXT: s_endpgm