Lines Matching refs:INV
6 …-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
59 ; SKIP-CACHE-INV-LABEL: global_wavefront_unordered_load:
60 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
61 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
62 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
63 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
64 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
65 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s0, s4
66 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s5
67 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[0:3], 0
68 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s4, s6
69 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s7
70 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s2
71 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s3
72 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
73 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
74 ; SKIP-CACHE-INV-NEXT: s_endpgm
133 ; SKIP-CACHE-INV-LABEL: global_wavefront_monotonic_load:
134 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
135 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
136 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
137 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
138 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
139 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s0, s4
140 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s5
141 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[0:3], 0
142 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s4, s6
143 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s7
144 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s2
145 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s3
146 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
147 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
148 ; SKIP-CACHE-INV-NEXT: s_endpgm
207 ; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_load:
208 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
209 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
210 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
211 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
212 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
213 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s0, s4
214 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s5
215 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[0:3], 0
216 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s4, s6
217 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s7
218 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s2
219 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s3
220 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
221 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
222 ; SKIP-CACHE-INV-NEXT: s_endpgm
281 ; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_load:
282 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
283 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
284 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
285 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
286 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
287 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s0, s4
288 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s5
289 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[0:3], 0
290 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s4, s6
291 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s7
292 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s2
293 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s3
294 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
295 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
296 ; SKIP-CACHE-INV-NEXT: s_endpgm
349 ; SKIP-CACHE-INV-LABEL: global_wavefront_unordered_store:
350 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
351 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[0:1], 0x9
352 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
353 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
354 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
355 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
356 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
357 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
358 ; SKIP-CACHE-INV-NEXT: s_endpgm
410 ; SKIP-CACHE-INV-LABEL: global_wavefront_monotonic_store:
411 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
412 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[0:1], 0x9
413 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
414 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
415 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
416 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
417 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
418 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
419 ; SKIP-CACHE-INV-NEXT: s_endpgm
471 ; SKIP-CACHE-INV-LABEL: global_wavefront_release_store:
472 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
473 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[0:1], 0x9
474 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
475 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
476 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
477 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
478 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
479 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
480 ; SKIP-CACHE-INV-NEXT: s_endpgm
532 ; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_store:
533 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
534 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[0:1], 0x9
535 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
536 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
537 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
538 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
539 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
540 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
541 ; SKIP-CACHE-INV-NEXT: s_endpgm
593 ; SKIP-CACHE-INV-LABEL: global_wavefront_monotonic_atomicrmw:
594 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
595 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
596 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
597 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
598 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
599 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
600 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
601 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
602 ; SKIP-CACHE-INV-NEXT: s_endpgm
654 ; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_atomicrmw:
655 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
656 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
657 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
658 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
659 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
660 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
661 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
662 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
663 ; SKIP-CACHE-INV-NEXT: s_endpgm
715 ; SKIP-CACHE-INV-LABEL: global_wavefront_release_atomicrmw:
716 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
717 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
718 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
719 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
720 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
721 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
722 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
723 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
724 ; SKIP-CACHE-INV-NEXT: s_endpgm
776 ; SKIP-CACHE-INV-LABEL: global_wavefront_acq_rel_atomicrmw:
777 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
778 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
779 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
780 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
781 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
782 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
783 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
784 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
785 ; SKIP-CACHE-INV-NEXT: s_endpgm
837 ; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_atomicrmw:
838 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
839 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
840 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
841 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
842 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
843 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
844 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
845 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
846 ; SKIP-CACHE-INV-NEXT: s_endpgm
906 ; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_ret_atomicrmw:
907 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
908 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
909 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
910 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
911 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
912 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
913 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
914 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0 glc
915 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
916 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
917 ; SKIP-CACHE-INV-NEXT: s_endpgm
978 ; SKIP-CACHE-INV-LABEL: global_wavefront_acq_rel_ret_atomicrmw:
979 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
980 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
981 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
982 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
983 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
984 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
985 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
986 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0 glc
987 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
988 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
989 ; SKIP-CACHE-INV-NEXT: s_endpgm
1050 ; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_ret_atomicrmw:
1051 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1052 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1053 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
1054 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1055 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1056 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1057 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1058 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0 glc
1059 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
1060 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
1061 ; SKIP-CACHE-INV-NEXT: s_endpgm
1120 ; SKIP-CACHE-INV-LABEL: global_wavefront_monotonic_monotonic_cmpxchg:
1121 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1122 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1123 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1124 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1125 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1126 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1127 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1128 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
1129 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
1130 ; SKIP-CACHE-INV-NEXT: s_endpgm
1189 ; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_monotonic_cmpxchg:
1190 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1191 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1192 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1193 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1194 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1195 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1196 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1197 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
1198 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
1199 ; SKIP-CACHE-INV-NEXT: s_endpgm
1258 ; SKIP-CACHE-INV-LABEL: global_wavefront_release_monotonic_cmpxchg:
1259 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1260 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1261 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1262 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1263 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1264 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1265 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1266 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
1267 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
1268 ; SKIP-CACHE-INV-NEXT: s_endpgm
1327 ; SKIP-CACHE-INV-LABEL: global_wavefront_acq_rel_monotonic_cmpxchg:
1328 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1329 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1330 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1331 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1332 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1333 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1334 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1335 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
1336 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
1337 ; SKIP-CACHE-INV-NEXT: s_endpgm
1396 ; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_monotonic_cmpxchg:
1397 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1398 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1399 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1400 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1401 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1402 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1403 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1404 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
1405 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
1406 ; SKIP-CACHE-INV-NEXT: s_endpgm
1465 ; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_acquire_cmpxchg:
1466 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1467 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1468 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1469 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1470 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1471 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1472 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1473 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
1474 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
1475 ; SKIP-CACHE-INV-NEXT: s_endpgm
1534 ; SKIP-CACHE-INV-LABEL: global_wavefront_release_acquire_cmpxchg:
1535 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1536 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1537 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1538 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1539 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1540 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1541 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1542 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
1543 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
1544 ; SKIP-CACHE-INV-NEXT: s_endpgm
1603 ; SKIP-CACHE-INV-LABEL: global_wavefront_acq_rel_acquire_cmpxchg:
1604 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1605 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1606 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1607 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1608 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1609 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1610 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1611 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
1612 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
1613 ; SKIP-CACHE-INV-NEXT: s_endpgm
1672 ; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_acquire_cmpxchg:
1673 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1674 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1675 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1676 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1677 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1678 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1679 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1680 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
1681 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
1682 ; SKIP-CACHE-INV-NEXT: s_endpgm
1741 ; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_seq_cst_cmpxchg:
1742 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1743 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1744 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1745 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1746 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1747 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1748 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1749 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
1750 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
1751 ; SKIP-CACHE-INV-NEXT: s_endpgm
1820 ; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_monotonic_ret_cmpxchg:
1821 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1822 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1823 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1824 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1825 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1826 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1827 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1828 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
1829 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
1830 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
1831 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
1832 ; SKIP-CACHE-INV-NEXT: s_endpgm
1903 ; SKIP-CACHE-INV-LABEL: global_wavefront_acq_rel_monotonic_ret_cmpxchg:
1904 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1905 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1906 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1907 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1908 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1909 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1910 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1911 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
1912 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
1913 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
1914 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
1915 ; SKIP-CACHE-INV-NEXT: s_endpgm
1986 ; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_monotonic_ret_cmpxchg:
1987 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1988 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
1989 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
1990 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
1991 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1992 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1993 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
1994 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
1995 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
1996 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
1997 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
1998 ; SKIP-CACHE-INV-NEXT: s_endpgm
2069 ; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_acquire_ret_cmpxchg:
2070 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2071 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
2072 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2073 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
2074 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
2075 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2076 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2077 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2078 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
2079 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
2080 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
2081 ; SKIP-CACHE-INV-NEXT: s_endpgm
2152 ; SKIP-CACHE-INV-LABEL: global_wavefront_release_acquire_ret_cmpxchg:
2153 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2154 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
2155 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2156 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
2157 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
2158 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2159 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2160 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2161 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
2162 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
2163 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
2164 ; SKIP-CACHE-INV-NEXT: s_endpgm
2235 ; SKIP-CACHE-INV-LABEL: global_wavefront_acq_rel_acquire_ret_cmpxchg:
2236 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2237 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
2238 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2239 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
2240 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
2241 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2242 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2243 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2244 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
2245 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
2246 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
2247 ; SKIP-CACHE-INV-NEXT: s_endpgm
2318 ; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_acquire_ret_cmpxchg:
2319 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2320 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
2321 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2322 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
2323 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
2324 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2325 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2326 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2327 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
2328 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
2329 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
2330 ; SKIP-CACHE-INV-NEXT: s_endpgm
2401 ; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_seq_cst_ret_cmpxchg:
2402 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2403 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
2404 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2405 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
2406 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
2407 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2408 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
2409 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
2410 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
2411 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
2412 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
2413 ; SKIP-CACHE-INV-NEXT: s_endpgm
2474 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_unordered_load:
2475 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2476 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
2477 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
2478 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
2479 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2480 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s0, s4
2481 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s5
2482 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[0:3], 0
2483 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s4, s6
2484 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s7
2485 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s2
2486 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s3
2487 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
2488 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
2489 ; SKIP-CACHE-INV-NEXT: s_endpgm
2548 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_monotonic_load:
2549 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2550 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
2551 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
2552 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
2553 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2554 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s0, s4
2555 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s5
2556 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[0:3], 0
2557 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s4, s6
2558 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s7
2559 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s2
2560 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s3
2561 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
2562 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
2563 ; SKIP-CACHE-INV-NEXT: s_endpgm
2622 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_load:
2623 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2624 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
2625 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
2626 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
2627 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2628 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s0, s4
2629 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s5
2630 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[0:3], 0
2631 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s4, s6
2632 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s7
2633 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s2
2634 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s3
2635 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
2636 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
2637 ; SKIP-CACHE-INV-NEXT: s_endpgm
2696 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_load:
2697 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2698 ; SKIP-CACHE-INV-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
2699 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
2700 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
2701 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2702 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s0, s4
2703 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s5
2704 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[0:3], 0
2705 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s4, s6
2706 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s7
2707 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s2
2708 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s3
2709 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
2710 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
2711 ; SKIP-CACHE-INV-NEXT: s_endpgm
2764 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_unordered_store:
2765 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2766 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[0:1], 0x9
2767 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2768 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
2769 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
2770 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2771 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
2772 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
2773 ; SKIP-CACHE-INV-NEXT: s_endpgm
2825 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_monotonic_store:
2826 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2827 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[0:1], 0x9
2828 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2829 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
2830 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
2831 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2832 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
2833 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
2834 ; SKIP-CACHE-INV-NEXT: s_endpgm
2886 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_release_store:
2887 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2888 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[0:1], 0x9
2889 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2890 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
2891 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
2892 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2893 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
2894 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
2895 ; SKIP-CACHE-INV-NEXT: s_endpgm
2947 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_store:
2948 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2949 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[0:1], 0x9
2950 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
2951 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, 0xf000
2952 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, -1
2953 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2954 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
2955 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
2956 ; SKIP-CACHE-INV-NEXT: s_endpgm
3008 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_monotonic_atomicrmw:
3009 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3010 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3011 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3012 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3013 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3014 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3015 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3016 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
3017 ; SKIP-CACHE-INV-NEXT: s_endpgm
3069 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_atomicrmw:
3070 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3071 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3072 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3073 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3074 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3075 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3076 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3077 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
3078 ; SKIP-CACHE-INV-NEXT: s_endpgm
3130 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_release_atomicrmw:
3131 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3132 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3133 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3134 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3135 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3136 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3137 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3138 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
3139 ; SKIP-CACHE-INV-NEXT: s_endpgm
3191 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acq_rel_atomicrmw:
3192 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3193 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3194 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3195 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3196 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3197 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3198 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3199 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
3200 ; SKIP-CACHE-INV-NEXT: s_endpgm
3252 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_atomicrmw:
3253 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3254 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3255 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3256 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3257 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3258 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3259 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3260 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
3261 ; SKIP-CACHE-INV-NEXT: s_endpgm
3321 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_ret_atomicrmw:
3322 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3323 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3324 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3325 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3326 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3327 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3328 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3329 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0 glc
3330 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
3331 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
3332 ; SKIP-CACHE-INV-NEXT: s_endpgm
3393 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acq_rel_ret_atomicrmw:
3394 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3395 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3396 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3397 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3398 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3399 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3400 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3401 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0 glc
3402 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
3403 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
3404 ; SKIP-CACHE-INV-NEXT: s_endpgm
3465 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_ret_atomicrmw:
3466 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3467 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3468 ; SKIP-CACHE-INV-NEXT: s_load_dword s0, s[0:1], 0xb
3469 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3470 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3471 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3472 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3473 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[4:7], 0 glc
3474 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
3475 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
3476 ; SKIP-CACHE-INV-NEXT: s_endpgm
3535 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_monotonic_monotonic_cmpxchg:
3536 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3537 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3538 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3539 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3540 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3541 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3542 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3543 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
3544 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
3545 ; SKIP-CACHE-INV-NEXT: s_endpgm
3604 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_monotonic_cmpxchg:
3605 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3606 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3607 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3608 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3609 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3610 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3611 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3612 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
3613 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
3614 ; SKIP-CACHE-INV-NEXT: s_endpgm
3673 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_release_monotonic_cmpxchg:
3674 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3675 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3676 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3677 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3678 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3679 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3680 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3681 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
3682 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
3683 ; SKIP-CACHE-INV-NEXT: s_endpgm
3742 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acq_rel_monotonic_cmpxchg:
3743 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3744 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3745 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3746 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3747 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3748 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3749 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3750 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
3751 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
3752 ; SKIP-CACHE-INV-NEXT: s_endpgm
3811 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_monotonic_cmpxchg:
3812 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3813 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3814 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3815 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3816 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3817 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3818 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3819 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
3820 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
3821 ; SKIP-CACHE-INV-NEXT: s_endpgm
3880 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_acquire_cmpxchg:
3881 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3882 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3883 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3884 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3885 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3886 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3887 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3888 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
3889 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
3890 ; SKIP-CACHE-INV-NEXT: s_endpgm
3949 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_release_acquire_cmpxchg:
3950 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3951 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
3952 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
3953 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
3954 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
3955 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3956 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
3957 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
3958 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
3959 ; SKIP-CACHE-INV-NEXT: s_endpgm
4018 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acq_rel_acquire_cmpxchg:
4019 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4020 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
4021 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4022 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
4023 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
4024 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4025 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
4026 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
4027 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
4028 ; SKIP-CACHE-INV-NEXT: s_endpgm
4087 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_acquire_cmpxchg:
4088 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4089 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
4090 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4091 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
4092 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
4093 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4094 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
4095 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
4096 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
4097 ; SKIP-CACHE-INV-NEXT: s_endpgm
4156 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
4157 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4158 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
4159 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4160 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
4161 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
4162 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4163 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
4164 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
4165 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
4166 ; SKIP-CACHE-INV-NEXT: s_endpgm
4235 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
4236 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4237 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
4238 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4239 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
4240 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
4241 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4242 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
4243 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
4244 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
4245 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
4246 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
4247 ; SKIP-CACHE-INV-NEXT: s_endpgm
4318 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
4319 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4320 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
4321 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4322 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
4323 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
4324 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4325 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
4326 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
4327 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
4328 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
4329 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
4330 ; SKIP-CACHE-INV-NEXT: s_endpgm
4401 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
4402 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4403 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
4404 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4405 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
4406 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
4407 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4408 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
4409 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
4410 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
4411 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
4412 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
4413 ; SKIP-CACHE-INV-NEXT: s_endpgm
4484 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_acquire_ret_cmpxchg:
4485 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4486 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
4487 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4488 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
4489 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
4490 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4491 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
4492 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
4493 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
4494 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
4495 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
4496 ; SKIP-CACHE-INV-NEXT: s_endpgm
4567 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_release_acquire_ret_cmpxchg:
4568 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4569 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
4570 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4571 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
4572 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
4573 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4574 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
4575 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
4576 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
4577 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
4578 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
4579 ; SKIP-CACHE-INV-NEXT: s_endpgm
4650 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
4651 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4652 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
4653 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4654 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
4655 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
4656 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4657 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
4658 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
4659 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
4660 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
4661 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
4662 ; SKIP-CACHE-INV-NEXT: s_endpgm
4733 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
4734 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4735 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
4736 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4737 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
4738 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
4739 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4740 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
4741 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
4742 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
4743 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
4744 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
4745 ; SKIP-CACHE-INV-NEXT: s_endpgm
4816 ; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
4817 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4818 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x9
4819 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xb
4820 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, 0xf000
4821 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
4822 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4823 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s0
4824 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, s1
4825 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
4826 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
4827 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[4:7], 0
4828 ; SKIP-CACHE-INV-NEXT: s_endpgm