• Home
  • Raw
  • Download

Lines Matching refs:INV

5 …-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
47 ; SKIP-CACHE-INV-LABEL: flat_singlethread_unordered_load:
48 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
49 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
50 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
51 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
52 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
53 ; SKIP-CACHE-INV-NEXT: flat_load_dword v0, v[0:1]
54 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
55 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s3
56 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
57 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[2:3], v0
58 ; SKIP-CACHE-INV-NEXT: s_endpgm
106 ; SKIP-CACHE-INV-LABEL: flat_singlethread_monotonic_load:
107 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
108 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
109 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
110 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
111 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
112 ; SKIP-CACHE-INV-NEXT: flat_load_dword v0, v[0:1]
113 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
114 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s3
115 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
116 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[2:3], v0
117 ; SKIP-CACHE-INV-NEXT: s_endpgm
165 ; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_load:
166 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
167 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
168 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
169 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
170 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
171 ; SKIP-CACHE-INV-NEXT: flat_load_dword v0, v[0:1]
172 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
173 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s3
174 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
175 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[2:3], v0
176 ; SKIP-CACHE-INV-NEXT: s_endpgm
224 ; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_load:
225 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
226 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
227 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
228 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
229 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
230 ; SKIP-CACHE-INV-NEXT: flat_load_dword v0, v[0:1]
231 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
232 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s3
233 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
234 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[2:3], v0
235 ; SKIP-CACHE-INV-NEXT: s_endpgm
279 ; SKIP-CACHE-INV-LABEL: flat_singlethread_unordered_store:
280 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
281 ; SKIP-CACHE-INV-NEXT: s_load_dword s2, s[0:1], 0x9
282 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
283 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
284 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
285 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
286 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
287 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
288 ; SKIP-CACHE-INV-NEXT: s_endpgm
331 ; SKIP-CACHE-INV-LABEL: flat_singlethread_monotonic_store:
332 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
333 ; SKIP-CACHE-INV-NEXT: s_load_dword s2, s[0:1], 0x9
334 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
335 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
336 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
337 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
338 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
339 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
340 ; SKIP-CACHE-INV-NEXT: s_endpgm
383 ; SKIP-CACHE-INV-LABEL: flat_singlethread_release_store:
384 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
385 ; SKIP-CACHE-INV-NEXT: s_load_dword s2, s[0:1], 0x9
386 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
387 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
388 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
389 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
390 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
391 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
392 ; SKIP-CACHE-INV-NEXT: s_endpgm
435 ; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_store:
436 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
437 ; SKIP-CACHE-INV-NEXT: s_load_dword s2, s[0:1], 0x9
438 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
439 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
440 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
441 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
442 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
443 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
444 ; SKIP-CACHE-INV-NEXT: s_endpgm
487 ; SKIP-CACHE-INV-LABEL: flat_singlethread_monotonic_atomicrmw:
488 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
489 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
490 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
491 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
492 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
493 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
494 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
495 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v[0:1], v2
496 ; SKIP-CACHE-INV-NEXT: s_endpgm
539 ; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_atomicrmw:
540 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
541 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
542 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
543 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
544 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
545 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
546 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
547 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v[0:1], v2
548 ; SKIP-CACHE-INV-NEXT: s_endpgm
591 ; SKIP-CACHE-INV-LABEL: flat_singlethread_release_atomicrmw:
592 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
593 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
594 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
595 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
596 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
597 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
598 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
599 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v[0:1], v2
600 ; SKIP-CACHE-INV-NEXT: s_endpgm
643 ; SKIP-CACHE-INV-LABEL: flat_singlethread_acq_rel_atomicrmw:
644 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
645 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
646 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
647 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
648 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
649 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
650 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
651 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v[0:1], v2
652 ; SKIP-CACHE-INV-NEXT: s_endpgm
695 ; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_atomicrmw:
696 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
697 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
698 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
699 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
700 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
701 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
702 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
703 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v[0:1], v2
704 ; SKIP-CACHE-INV-NEXT: s_endpgm
753 ; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_ret_atomicrmw:
754 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
755 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
756 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
757 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
758 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
759 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
760 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
761 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v2, v[0:1], v2 glc
762 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
763 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
764 ; SKIP-CACHE-INV-NEXT: s_endpgm
814 ; SKIP-CACHE-INV-LABEL: flat_singlethread_acq_rel_ret_atomicrmw:
815 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
816 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
817 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
818 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
819 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
820 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
821 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
822 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v2, v[0:1], v2 glc
823 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
824 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
825 ; SKIP-CACHE-INV-NEXT: s_endpgm
875 ; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_ret_atomicrmw:
876 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
877 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
878 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
879 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
880 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
881 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
882 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
883 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v2, v[0:1], v2 glc
884 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
885 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
886 ; SKIP-CACHE-INV-NEXT: s_endpgm
939 ; SKIP-CACHE-INV-LABEL: flat_singlethread_monotonic_monotonic_cmpxchg:
940 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
941 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
942 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
943 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
944 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
945 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
946 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
947 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
948 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
949 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
950 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
951 ; SKIP-CACHE-INV-NEXT: s_endpgm
1004 ; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_monotonic_cmpxchg:
1005 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1006 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1007 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1008 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1009 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
1010 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
1011 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1012 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1013 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1014 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1015 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
1016 ; SKIP-CACHE-INV-NEXT: s_endpgm
1069 ; SKIP-CACHE-INV-LABEL: flat_singlethread_release_monotonic_cmpxchg:
1070 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1071 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1072 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1073 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1074 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
1075 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
1076 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1077 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1078 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1079 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1080 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
1081 ; SKIP-CACHE-INV-NEXT: s_endpgm
1134 ; SKIP-CACHE-INV-LABEL: flat_singlethread_acq_rel_monotonic_cmpxchg:
1135 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1136 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1137 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1138 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1139 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
1140 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
1141 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1142 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1143 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1144 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1145 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
1146 ; SKIP-CACHE-INV-NEXT: s_endpgm
1199 ; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_monotonic_cmpxchg:
1200 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1201 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1202 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1203 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1204 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
1205 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
1206 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1207 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1208 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1209 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1210 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
1211 ; SKIP-CACHE-INV-NEXT: s_endpgm
1264 ; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_acquire_cmpxchg:
1265 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1266 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1267 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1268 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1269 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
1270 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
1271 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1272 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1273 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1274 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1275 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
1276 ; SKIP-CACHE-INV-NEXT: s_endpgm
1329 ; SKIP-CACHE-INV-LABEL: flat_singlethread_release_acquire_cmpxchg:
1330 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1331 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1332 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1333 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1334 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
1335 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
1336 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1337 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1338 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1339 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1340 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
1341 ; SKIP-CACHE-INV-NEXT: s_endpgm
1394 ; SKIP-CACHE-INV-LABEL: flat_singlethread_acq_rel_acquire_cmpxchg:
1395 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1396 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1397 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1398 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1399 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
1400 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
1401 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1402 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1403 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1404 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1405 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
1406 ; SKIP-CACHE-INV-NEXT: s_endpgm
1459 ; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_acquire_cmpxchg:
1460 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1461 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1462 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1463 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1464 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
1465 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
1466 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1467 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1468 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1469 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1470 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
1471 ; SKIP-CACHE-INV-NEXT: s_endpgm
1524 ; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_seq_cst_cmpxchg:
1525 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1526 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1527 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1528 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1529 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
1530 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
1531 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1532 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1533 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1534 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1535 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
1536 ; SKIP-CACHE-INV-NEXT: s_endpgm
1601 ; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_monotonic_ret_cmpxchg:
1602 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1603 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1604 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1605 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1606 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
1607 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
1608 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
1609 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1610 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
1611 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1612 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
1613 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1614 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1615 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
1616 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
1617 ; SKIP-CACHE-INV-NEXT: s_endpgm
1684 ; SKIP-CACHE-INV-LABEL: flat_singlethread_acq_rel_monotonic_ret_cmpxchg:
1685 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1686 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1687 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1688 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1689 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
1690 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
1691 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
1692 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1693 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
1694 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1695 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
1696 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1697 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1698 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
1699 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
1700 ; SKIP-CACHE-INV-NEXT: s_endpgm
1767 ; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_monotonic_ret_cmpxchg:
1768 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1769 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1770 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1771 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1772 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
1773 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
1774 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
1775 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1776 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
1777 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1778 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
1779 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1780 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1781 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
1782 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
1783 ; SKIP-CACHE-INV-NEXT: s_endpgm
1850 ; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_acquire_ret_cmpxchg:
1851 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1852 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
1853 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1854 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1855 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
1856 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
1857 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
1858 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1859 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
1860 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1861 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
1862 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1863 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1864 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
1865 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
1866 ; SKIP-CACHE-INV-NEXT: s_endpgm
1933 ; SKIP-CACHE-INV-LABEL: flat_singlethread_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_dwordx2 s[0:1], s[0:1], 0xb
1937 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1938 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
1939 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
1940 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
1941 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
1942 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
1943 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
1944 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
1945 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
1946 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
1947 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
1948 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
1949 ; SKIP-CACHE-INV-NEXT: s_endpgm
2016 ; SKIP-CACHE-INV-LABEL: flat_singlethread_acq_rel_acquire_ret_cmpxchg:
2017 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2018 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
2019 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2020 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2021 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
2022 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
2023 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
2024 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
2025 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
2026 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
2027 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
2028 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
2029 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
2030 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
2031 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
2032 ; SKIP-CACHE-INV-NEXT: s_endpgm
2099 ; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_acquire_ret_cmpxchg:
2100 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2101 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
2102 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2103 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2104 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
2105 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
2106 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
2107 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
2108 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
2109 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
2110 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
2111 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
2112 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
2113 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
2114 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
2115 ; SKIP-CACHE-INV-NEXT: s_endpgm
2182 ; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_seq_cst_ret_cmpxchg:
2183 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2184 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
2185 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2186 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2187 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
2188 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
2189 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
2190 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
2191 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
2192 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
2193 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
2194 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
2195 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
2196 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
2197 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
2198 ; SKIP-CACHE-INV-NEXT: s_endpgm
2248 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_unordered_load:
2249 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2250 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
2251 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2252 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2253 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2254 ; SKIP-CACHE-INV-NEXT: flat_load_dword v0, v[0:1]
2255 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
2256 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s3
2257 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
2258 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[2:3], v0
2259 ; SKIP-CACHE-INV-NEXT: s_endpgm
2307 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_monotonic_load:
2308 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2309 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
2310 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2311 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2312 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2313 ; SKIP-CACHE-INV-NEXT: flat_load_dword v0, v[0:1]
2314 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
2315 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s3
2316 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
2317 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[2:3], v0
2318 ; SKIP-CACHE-INV-NEXT: s_endpgm
2366 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_load:
2367 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2368 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
2369 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2370 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2371 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2372 ; SKIP-CACHE-INV-NEXT: flat_load_dword v0, v[0:1]
2373 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
2374 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s3
2375 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
2376 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[2:3], v0
2377 ; SKIP-CACHE-INV-NEXT: s_endpgm
2425 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_load:
2426 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2427 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
2428 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2429 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2430 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2431 ; SKIP-CACHE-INV-NEXT: flat_load_dword v0, v[0:1]
2432 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
2433 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s3
2434 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
2435 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[2:3], v0
2436 ; SKIP-CACHE-INV-NEXT: s_endpgm
2480 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_unordered_store:
2481 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2482 ; SKIP-CACHE-INV-NEXT: s_load_dword s2, s[0:1], 0x9
2483 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2484 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2485 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
2486 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2487 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2488 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
2489 ; SKIP-CACHE-INV-NEXT: s_endpgm
2532 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_monotonic_store:
2533 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2534 ; SKIP-CACHE-INV-NEXT: s_load_dword s2, s[0:1], 0x9
2535 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2536 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2537 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
2538 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2539 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2540 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
2541 ; SKIP-CACHE-INV-NEXT: s_endpgm
2584 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_release_store:
2585 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2586 ; SKIP-CACHE-INV-NEXT: s_load_dword s2, s[0:1], 0x9
2587 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2588 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2589 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
2590 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2591 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2592 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
2593 ; SKIP-CACHE-INV-NEXT: s_endpgm
2636 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_store:
2637 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2638 ; SKIP-CACHE-INV-NEXT: s_load_dword s2, s[0:1], 0x9
2639 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2640 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2641 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s2
2642 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2643 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2644 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
2645 ; SKIP-CACHE-INV-NEXT: s_endpgm
2688 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_monotonic_atomicrmw:
2689 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2690 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
2691 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
2692 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2693 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
2694 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
2695 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
2696 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v[0:1], v2
2697 ; SKIP-CACHE-INV-NEXT: s_endpgm
2740 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_atomicrmw:
2741 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2742 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
2743 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
2744 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2745 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
2746 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
2747 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
2748 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v[0:1], v2
2749 ; SKIP-CACHE-INV-NEXT: s_endpgm
2792 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_release_atomicrmw:
2793 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2794 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
2795 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
2796 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2797 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
2798 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
2799 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
2800 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v[0:1], v2
2801 ; SKIP-CACHE-INV-NEXT: s_endpgm
2844 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acq_rel_atomicrmw:
2845 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2846 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
2847 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
2848 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2849 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
2850 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
2851 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
2852 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v[0:1], v2
2853 ; SKIP-CACHE-INV-NEXT: s_endpgm
2896 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_atomicrmw:
2897 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2898 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
2899 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
2900 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2901 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
2902 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
2903 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
2904 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v[0:1], v2
2905 ; SKIP-CACHE-INV-NEXT: s_endpgm
2954 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_ret_atomicrmw:
2955 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2956 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
2957 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
2958 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2959 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
2960 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
2961 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
2962 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v2, v[0:1], v2 glc
2963 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
2964 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
2965 ; SKIP-CACHE-INV-NEXT: s_endpgm
3015 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acq_rel_ret_atomicrmw:
3016 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3017 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3018 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3019 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3020 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3021 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3022 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3023 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v2, v[0:1], v2 glc
3024 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
3025 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
3026 ; SKIP-CACHE-INV-NEXT: s_endpgm
3076 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_ret_atomicrmw:
3077 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3078 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3079 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3080 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3081 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3082 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3083 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3084 ; SKIP-CACHE-INV-NEXT: flat_atomic_swap v2, v[0:1], v2 glc
3085 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
3086 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
3087 ; SKIP-CACHE-INV-NEXT: s_endpgm
3140 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_monotonic_monotonic_cmpxchg:
3141 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3142 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3143 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3144 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3145 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
3146 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
3147 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3148 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
3149 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3150 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3151 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3152 ; SKIP-CACHE-INV-NEXT: s_endpgm
3205 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_monotonic_cmpxchg:
3206 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3207 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3208 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3209 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3210 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
3211 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
3212 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3213 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
3214 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3215 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3216 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3217 ; SKIP-CACHE-INV-NEXT: s_endpgm
3270 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_release_monotonic_cmpxchg:
3271 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3272 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3273 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3274 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3275 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
3276 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
3277 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3278 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
3279 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3280 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3281 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3282 ; SKIP-CACHE-INV-NEXT: s_endpgm
3335 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acq_rel_monotonic_cmpxchg:
3336 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3337 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3338 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3339 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3340 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
3341 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
3342 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3343 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
3344 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3345 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3346 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3347 ; SKIP-CACHE-INV-NEXT: s_endpgm
3400 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_monotonic_cmpxchg:
3401 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3402 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3403 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3404 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3405 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
3406 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
3407 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3408 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
3409 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3410 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3411 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3412 ; SKIP-CACHE-INV-NEXT: s_endpgm
3465 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_acquire_cmpxchg:
3466 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3467 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3468 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3469 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3470 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
3471 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
3472 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3473 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
3474 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3475 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3476 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3477 ; SKIP-CACHE-INV-NEXT: s_endpgm
3530 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_release_acquire_cmpxchg:
3531 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3532 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3533 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3534 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3535 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
3536 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
3537 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3538 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
3539 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3540 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3541 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3542 ; SKIP-CACHE-INV-NEXT: s_endpgm
3595 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acq_rel_acquire_cmpxchg:
3596 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3597 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3598 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3599 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3600 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
3601 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
3602 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3603 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
3604 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3605 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3606 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3607 ; SKIP-CACHE-INV-NEXT: s_endpgm
3660 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_acquire_cmpxchg:
3661 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3662 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3663 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3664 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3665 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
3666 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
3667 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3668 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
3669 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3670 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3671 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3672 ; SKIP-CACHE-INV-NEXT: s_endpgm
3725 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
3726 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3727 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3728 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3729 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3730 ; SKIP-CACHE-INV-NEXT: s_add_u32 s2, s2, 16
3731 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s3, s3, 0
3732 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3733 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
3734 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3735 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3736 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3737 ; SKIP-CACHE-INV-NEXT: s_endpgm
3802 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
3803 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3804 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3805 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3806 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3807 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
3808 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
3809 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
3810 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
3811 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
3812 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3813 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
3814 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3815 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3816 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
3817 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
3818 ; SKIP-CACHE-INV-NEXT: s_endpgm
3885 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
3886 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3887 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3888 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3889 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3890 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
3891 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
3892 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
3893 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
3894 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
3895 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3896 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
3897 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3898 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3899 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
3900 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
3901 ; SKIP-CACHE-INV-NEXT: s_endpgm
3968 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
3969 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3970 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
3971 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3972 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3973 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
3974 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
3975 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
3976 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
3977 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
3978 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
3979 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
3980 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
3981 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
3982 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
3983 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
3984 ; SKIP-CACHE-INV-NEXT: s_endpgm
4051 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_acquire_ret_cmpxchg:
4052 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4053 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
4054 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4055 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4056 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
4057 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
4058 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
4059 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
4060 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
4061 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
4062 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
4063 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
4064 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
4065 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
4066 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
4067 ; SKIP-CACHE-INV-NEXT: s_endpgm
4134 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_release_acquire_ret_cmpxchg:
4135 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4136 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
4137 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4138 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4139 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
4140 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
4141 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
4142 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
4143 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
4144 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
4145 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
4146 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
4147 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
4148 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
4149 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
4150 ; SKIP-CACHE-INV-NEXT: s_endpgm
4217 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
4218 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4219 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
4220 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4221 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4222 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
4223 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
4224 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
4225 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
4226 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
4227 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
4228 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
4229 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
4230 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
4231 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
4232 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
4233 ; SKIP-CACHE-INV-NEXT: s_endpgm
4300 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
4301 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4302 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
4303 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4304 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4305 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
4306 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
4307 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
4308 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
4309 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
4310 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
4311 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
4312 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
4313 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
4314 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
4315 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
4316 ; SKIP-CACHE-INV-NEXT: s_endpgm
4383 ; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
4384 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4385 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x9
4386 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4387 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4388 ; SKIP-CACHE-INV-NEXT: s_add_u32 s4, s2, 16
4389 ; SKIP-CACHE-INV-NEXT: s_addc_u32 s5, s3, 0
4390 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
4391 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v3, s1
4392 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s5
4393 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s0
4394 ; SKIP-CACHE-INV-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
4395 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s2
4396 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s3
4397 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
4398 ; SKIP-CACHE-INV-NEXT: flat_store_dword v[0:1], v2
4399 ; SKIP-CACHE-INV-NEXT: s_endpgm