Skip to content

Commit d05b07a

Browse files
anjenneranjenner
committed
Add some metadata tests
1 parent ef23b55 commit d05b07a

File tree

2 files changed

+592
-0
lines changed

2 files changed

+592
-0
lines changed

‎llvm/test/CodeGen/AMDGPU/atomicrmw_usub_cond.ll‎

Lines changed: 242 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1312,5 +1312,247 @@ define amdgpu_kernel void @global_atomic_usub_cond_sgpr_base_offset_nortn(ptr ad
13121312
ret void
13131313
}
13141314

1315+
define i32 @global_atomic_usub_cond__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, i32 %data) {
1316+
; GFX9-SDAG-LABEL: global_atomic_usub_cond__amdgpu_no_remote_memory:
1317+
; GFX9-SDAG: ; %bb.0:
1318+
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1319+
; GFX9-SDAG-NEXT: global_load_dword v3, v[0:1], off
1320+
; GFX9-SDAG-NEXT: s_mov_b64 s[4:5], 0
1321+
; GFX9-SDAG-NEXT: .LBB15_1: ; %atomicrmw.start
1322+
; GFX9-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1
1323+
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
1324+
; GFX9-SDAG-NEXT: v_mov_b32_e32 v4, v3
1325+
; GFX9-SDAG-NEXT: v_sub_u32_e32 v3, v4, v2
1326+
; GFX9-SDAG-NEXT: v_cmp_ge_u32_e32 vcc, v4, v2
1327+
; GFX9-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
1328+
; GFX9-SDAG-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
1329+
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
1330+
; GFX9-SDAG-NEXT: buffer_wbinvl1_vol
1331+
; GFX9-SDAG-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
1332+
; GFX9-SDAG-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
1333+
; GFX9-SDAG-NEXT: s_andn2_b64 exec, exec, s[4:5]
1334+
; GFX9-SDAG-NEXT: s_cbranch_execnz .LBB15_1
1335+
; GFX9-SDAG-NEXT: ; %bb.2: ; %atomicrmw.end
1336+
; GFX9-SDAG-NEXT: s_or_b64 exec, exec, s[4:5]
1337+
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, v3
1338+
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
1339+
;
1340+
; GFX12-SDAG-LABEL: global_atomic_usub_cond__amdgpu_no_remote_memory:
1341+
; GFX12-SDAG: ; %bb.0:
1342+
; GFX12-SDAG-NEXT: s_wait_loadcnt_dscnt 0x0
1343+
; GFX12-SDAG-NEXT: s_wait_expcnt 0x0
1344+
; GFX12-SDAG-NEXT: s_wait_samplecnt 0x0
1345+
; GFX12-SDAG-NEXT: s_wait_bvhcnt 0x0
1346+
; GFX12-SDAG-NEXT: s_wait_kmcnt 0x0
1347+
; GFX12-SDAG-NEXT: s_wait_storecnt 0x0
1348+
; GFX12-SDAG-NEXT: global_atomic_cond_sub_u32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
1349+
; GFX12-SDAG-NEXT: s_wait_loadcnt 0x0
1350+
; GFX12-SDAG-NEXT: global_inv scope:SCOPE_DEV
1351+
; GFX12-SDAG-NEXT: s_wait_loadcnt 0x0
1352+
; GFX12-SDAG-NEXT: s_setpc_b64 s[30:31]
1353+
;
1354+
; GFX9-GISEL-LABEL: global_atomic_usub_cond__amdgpu_no_remote_memory:
1355+
; GFX9-GISEL: ; %bb.0:
1356+
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1357+
; GFX9-GISEL-NEXT: global_load_dword v3, v[0:1], off
1358+
; GFX9-GISEL-NEXT: s_mov_b64 s[4:5], 0
1359+
; GFX9-GISEL-NEXT: .LBB15_1: ; %atomicrmw.start
1360+
; GFX9-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1
1361+
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
1362+
; GFX9-GISEL-NEXT: v_mov_b32_e32 v4, v3
1363+
; GFX9-GISEL-NEXT: v_sub_u32_e32 v3, v4, v2
1364+
; GFX9-GISEL-NEXT: v_cmp_ge_u32_e32 vcc, v4, v2
1365+
; GFX9-GISEL-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
1366+
; GFX9-GISEL-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
1367+
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
1368+
; GFX9-GISEL-NEXT: buffer_wbinvl1_vol
1369+
; GFX9-GISEL-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
1370+
; GFX9-GISEL-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
1371+
; GFX9-GISEL-NEXT: s_andn2_b64 exec, exec, s[4:5]
1372+
; GFX9-GISEL-NEXT: s_cbranch_execnz .LBB15_1
1373+
; GFX9-GISEL-NEXT: ; %bb.2: ; %atomicrmw.end
1374+
; GFX9-GISEL-NEXT: s_or_b64 exec, exec, s[4:5]
1375+
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, v3
1376+
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
1377+
;
1378+
; GFX12-GISEL-LABEL: global_atomic_usub_cond__amdgpu_no_remote_memory:
1379+
; GFX12-GISEL: ; %bb.0:
1380+
; GFX12-GISEL-NEXT: s_wait_loadcnt_dscnt 0x0
1381+
; GFX12-GISEL-NEXT: s_wait_expcnt 0x0
1382+
; GFX12-GISEL-NEXT: s_wait_samplecnt 0x0
1383+
; GFX12-GISEL-NEXT: s_wait_bvhcnt 0x0
1384+
; GFX12-GISEL-NEXT: s_wait_kmcnt 0x0
1385+
; GFX12-GISEL-NEXT: s_wait_storecnt 0x0
1386+
; GFX12-GISEL-NEXT: global_atomic_cond_sub_u32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
1387+
; GFX12-GISEL-NEXT: s_wait_loadcnt 0x0
1388+
; GFX12-GISEL-NEXT: global_inv scope:SCOPE_DEV
1389+
; GFX12-GISEL-NEXT: s_wait_loadcnt 0x0
1390+
; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31]
1391+
%ret = atomicrmw usub_cond ptr addrspace(1) %ptr, i32 %data syncscope("agent") seq_cst, align 4, !amdgpu.no.remote.memory !0
1392+
ret i32 %ret
1393+
}
1394+
1395+
define i32 @global_atomic_usub_cond__amdgpu_no_fine_grained_memory(ptr addrspace(1) %ptr, i32 %data) {
1396+
; GFX9-SDAG-LABEL: global_atomic_usub_cond__amdgpu_no_fine_grained_memory:
1397+
; GFX9-SDAG: ; %bb.0:
1398+
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1399+
; GFX9-SDAG-NEXT: global_load_dword v3, v[0:1], off
1400+
; GFX9-SDAG-NEXT: s_mov_b64 s[4:5], 0
1401+
; GFX9-SDAG-NEXT: .LBB16_1: ; %atomicrmw.start
1402+
; GFX9-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1
1403+
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
1404+
; GFX9-SDAG-NEXT: v_mov_b32_e32 v4, v3
1405+
; GFX9-SDAG-NEXT: v_sub_u32_e32 v3, v4, v2
1406+
; GFX9-SDAG-NEXT: v_cmp_ge_u32_e32 vcc, v4, v2
1407+
; GFX9-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
1408+
; GFX9-SDAG-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
1409+
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
1410+
; GFX9-SDAG-NEXT: buffer_wbinvl1_vol
1411+
; GFX9-SDAG-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
1412+
; GFX9-SDAG-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
1413+
; GFX9-SDAG-NEXT: s_andn2_b64 exec, exec, s[4:5]
1414+
; GFX9-SDAG-NEXT: s_cbranch_execnz .LBB16_1
1415+
; GFX9-SDAG-NEXT: ; %bb.2: ; %atomicrmw.end
1416+
; GFX9-SDAG-NEXT: s_or_b64 exec, exec, s[4:5]
1417+
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, v3
1418+
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
1419+
;
1420+
; GFX12-SDAG-LABEL: global_atomic_usub_cond__amdgpu_no_fine_grained_memory:
1421+
; GFX12-SDAG: ; %bb.0:
1422+
; GFX12-SDAG-NEXT: s_wait_loadcnt_dscnt 0x0
1423+
; GFX12-SDAG-NEXT: s_wait_expcnt 0x0
1424+
; GFX12-SDAG-NEXT: s_wait_samplecnt 0x0
1425+
; GFX12-SDAG-NEXT: s_wait_bvhcnt 0x0
1426+
; GFX12-SDAG-NEXT: s_wait_kmcnt 0x0
1427+
; GFX12-SDAG-NEXT: s_wait_storecnt 0x0
1428+
; GFX12-SDAG-NEXT: global_atomic_cond_sub_u32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
1429+
; GFX12-SDAG-NEXT: s_wait_loadcnt 0x0
1430+
; GFX12-SDAG-NEXT: global_inv scope:SCOPE_DEV
1431+
; GFX12-SDAG-NEXT: s_wait_loadcnt 0x0
1432+
; GFX12-SDAG-NEXT: s_setpc_b64 s[30:31]
1433+
;
1434+
; GFX9-GISEL-LABEL: global_atomic_usub_cond__amdgpu_no_fine_grained_memory:
1435+
; GFX9-GISEL: ; %bb.0:
1436+
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1437+
; GFX9-GISEL-NEXT: global_load_dword v3, v[0:1], off
1438+
; GFX9-GISEL-NEXT: s_mov_b64 s[4:5], 0
1439+
; GFX9-GISEL-NEXT: .LBB16_1: ; %atomicrmw.start
1440+
; GFX9-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1
1441+
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
1442+
; GFX9-GISEL-NEXT: v_mov_b32_e32 v4, v3
1443+
; GFX9-GISEL-NEXT: v_sub_u32_e32 v3, v4, v2
1444+
; GFX9-GISEL-NEXT: v_cmp_ge_u32_e32 vcc, v4, v2
1445+
; GFX9-GISEL-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
1446+
; GFX9-GISEL-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
1447+
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
1448+
; GFX9-GISEL-NEXT: buffer_wbinvl1_vol
1449+
; GFX9-GISEL-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
1450+
; GFX9-GISEL-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
1451+
; GFX9-GISEL-NEXT: s_andn2_b64 exec, exec, s[4:5]
1452+
; GFX9-GISEL-NEXT: s_cbranch_execnz .LBB16_1
1453+
; GFX9-GISEL-NEXT: ; %bb.2: ; %atomicrmw.end
1454+
; GFX9-GISEL-NEXT: s_or_b64 exec, exec, s[4:5]
1455+
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, v3
1456+
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
1457+
;
1458+
; GFX12-GISEL-LABEL: global_atomic_usub_cond__amdgpu_no_fine_grained_memory:
1459+
; GFX12-GISEL: ; %bb.0:
1460+
; GFX12-GISEL-NEXT: s_wait_loadcnt_dscnt 0x0
1461+
; GFX12-GISEL-NEXT: s_wait_expcnt 0x0
1462+
; GFX12-GISEL-NEXT: s_wait_samplecnt 0x0
1463+
; GFX12-GISEL-NEXT: s_wait_bvhcnt 0x0
1464+
; GFX12-GISEL-NEXT: s_wait_kmcnt 0x0
1465+
; GFX12-GISEL-NEXT: s_wait_storecnt 0x0
1466+
; GFX12-GISEL-NEXT: global_atomic_cond_sub_u32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
1467+
; GFX12-GISEL-NEXT: s_wait_loadcnt 0x0
1468+
; GFX12-GISEL-NEXT: global_inv scope:SCOPE_DEV
1469+
; GFX12-GISEL-NEXT: s_wait_loadcnt 0x0
1470+
; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31]
1471+
%ret = atomicrmw usub_cond ptr addrspace(1) %ptr, i32 %data syncscope("agent") seq_cst, align 4, !amdgpu.no.fine.grained.memory !0
1472+
ret i32 %ret
1473+
}
1474+
1475+
define i32 @global_atomic_usub_cond__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr, i32 %data) {
1476+
; GFX9-SDAG-LABEL: global_atomic_usub_cond__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
1477+
; GFX9-SDAG: ; %bb.0:
1478+
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1479+
; GFX9-SDAG-NEXT: global_load_dword v3, v[0:1], off
1480+
; GFX9-SDAG-NEXT: s_mov_b64 s[4:5], 0
1481+
; GFX9-SDAG-NEXT: .LBB17_1: ; %atomicrmw.start
1482+
; GFX9-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1
1483+
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
1484+
; GFX9-SDAG-NEXT: v_mov_b32_e32 v4, v3
1485+
; GFX9-SDAG-NEXT: v_sub_u32_e32 v3, v4, v2
1486+
; GFX9-SDAG-NEXT: v_cmp_ge_u32_e32 vcc, v4, v2
1487+
; GFX9-SDAG-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
1488+
; GFX9-SDAG-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
1489+
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
1490+
; GFX9-SDAG-NEXT: buffer_wbinvl1_vol
1491+
; GFX9-SDAG-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
1492+
; GFX9-SDAG-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
1493+
; GFX9-SDAG-NEXT: s_andn2_b64 exec, exec, s[4:5]
1494+
; GFX9-SDAG-NEXT: s_cbranch_execnz .LBB17_1
1495+
; GFX9-SDAG-NEXT: ; %bb.2: ; %atomicrmw.end
1496+
; GFX9-SDAG-NEXT: s_or_b64 exec, exec, s[4:5]
1497+
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, v3
1498+
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
1499+
;
1500+
; GFX12-SDAG-LABEL: global_atomic_usub_cond__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
1501+
; GFX12-SDAG: ; %bb.0:
1502+
; GFX12-SDAG-NEXT: s_wait_loadcnt_dscnt 0x0
1503+
; GFX12-SDAG-NEXT: s_wait_expcnt 0x0
1504+
; GFX12-SDAG-NEXT: s_wait_samplecnt 0x0
1505+
; GFX12-SDAG-NEXT: s_wait_bvhcnt 0x0
1506+
; GFX12-SDAG-NEXT: s_wait_kmcnt 0x0
1507+
; GFX12-SDAG-NEXT: s_wait_storecnt 0x0
1508+
; GFX12-SDAG-NEXT: global_atomic_cond_sub_u32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
1509+
; GFX12-SDAG-NEXT: s_wait_loadcnt 0x0
1510+
; GFX12-SDAG-NEXT: global_inv scope:SCOPE_DEV
1511+
; GFX12-SDAG-NEXT: s_wait_loadcnt 0x0
1512+
; GFX12-SDAG-NEXT: s_setpc_b64 s[30:31]
1513+
;
1514+
; GFX9-GISEL-LABEL: global_atomic_usub_cond__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
1515+
; GFX9-GISEL: ; %bb.0:
1516+
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1517+
; GFX9-GISEL-NEXT: global_load_dword v3, v[0:1], off
1518+
; GFX9-GISEL-NEXT: s_mov_b64 s[4:5], 0
1519+
; GFX9-GISEL-NEXT: .LBB17_1: ; %atomicrmw.start
1520+
; GFX9-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1
1521+
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
1522+
; GFX9-GISEL-NEXT: v_mov_b32_e32 v4, v3
1523+
; GFX9-GISEL-NEXT: v_sub_u32_e32 v3, v4, v2
1524+
; GFX9-GISEL-NEXT: v_cmp_ge_u32_e32 vcc, v4, v2
1525+
; GFX9-GISEL-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
1526+
; GFX9-GISEL-NEXT: global_atomic_cmpswap v3, v[0:1], v[3:4], off glc
1527+
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
1528+
; GFX9-GISEL-NEXT: buffer_wbinvl1_vol
1529+
; GFX9-GISEL-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4
1530+
; GFX9-GISEL-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
1531+
; GFX9-GISEL-NEXT: s_andn2_b64 exec, exec, s[4:5]
1532+
; GFX9-GISEL-NEXT: s_cbranch_execnz .LBB17_1
1533+
; GFX9-GISEL-NEXT: ; %bb.2: ; %atomicrmw.end
1534+
; GFX9-GISEL-NEXT: s_or_b64 exec, exec, s[4:5]
1535+
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, v3
1536+
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
1537+
;
1538+
; GFX12-GISEL-LABEL: global_atomic_usub_cond__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
1539+
; GFX12-GISEL: ; %bb.0:
1540+
; GFX12-GISEL-NEXT: s_wait_loadcnt_dscnt 0x0
1541+
; GFX12-GISEL-NEXT: s_wait_expcnt 0x0
1542+
; GFX12-GISEL-NEXT: s_wait_samplecnt 0x0
1543+
; GFX12-GISEL-NEXT: s_wait_bvhcnt 0x0
1544+
; GFX12-GISEL-NEXT: s_wait_kmcnt 0x0
1545+
; GFX12-GISEL-NEXT: s_wait_storecnt 0x0
1546+
; GFX12-GISEL-NEXT: global_atomic_cond_sub_u32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
1547+
; GFX12-GISEL-NEXT: s_wait_loadcnt 0x0
1548+
; GFX12-GISEL-NEXT: global_inv scope:SCOPE_DEV
1549+
; GFX12-GISEL-NEXT: s_wait_loadcnt 0x0
1550+
; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31]
1551+
%ret = atomicrmw usub_cond ptr addrspace(1) %ptr, i32 %data syncscope("agent") seq_cst, align 4, !amdgpu.no.fine.grained.memory !0, !amdgpu.no.remote.memory !0
1552+
ret i32 %ret
1553+
}
1554+
13151555
attributes #0 = { nounwind willreturn }
13161556
attributes #1 = { argmemonly nounwind }
1557+
1558+
!0 = !{}

0 commit comments

Comments
 (0)