@@ -1269,9 +1269,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
1269
1269
; GFX8DAGISEL-LABEL: divergent_value_i64:
1270
1270
; GFX8DAGISEL: ; %bb.0: ; %entry
1271
1271
; GFX8DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1272
- ; GFX8DAGISEL-NEXT: s_mov_b32 s4 , -1
1272
+ ; GFX8DAGISEL-NEXT: s_mov_b64 s[4:5] , -1
1273
1273
; GFX8DAGISEL-NEXT: s_mov_b64 s[6:7], exec
1274
- ; GFX8DAGISEL-NEXT: s_mov_b32 s5, s4
1275
1274
; GFX8DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
1276
1275
; GFX8DAGISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
1277
1276
; GFX8DAGISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1290,9 +1289,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
1290
1289
; GFX8GISEL-LABEL: divergent_value_i64:
1291
1290
; GFX8GISEL: ; %bb.0: ; %entry
1292
1291
; GFX8GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1293
- ; GFX8GISEL-NEXT: s_mov_b32 s4 , -1
1292
+ ; GFX8GISEL-NEXT: s_mov_b64 s[4:5] , -1
1294
1293
; GFX8GISEL-NEXT: s_mov_b64 s[6:7], exec
1295
- ; GFX8GISEL-NEXT: s_mov_b32 s5, s4
1296
1294
; GFX8GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
1297
1295
; GFX8GISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
1298
1296
; GFX8GISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1311,9 +1309,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
1311
1309
; GFX9DAGISEL-LABEL: divergent_value_i64:
1312
1310
; GFX9DAGISEL: ; %bb.0: ; %entry
1313
1311
; GFX9DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1314
- ; GFX9DAGISEL-NEXT: s_mov_b32 s4 , -1
1312
+ ; GFX9DAGISEL-NEXT: s_mov_b64 s[4:5] , -1
1315
1313
; GFX9DAGISEL-NEXT: s_mov_b64 s[6:7], exec
1316
- ; GFX9DAGISEL-NEXT: s_mov_b32 s5, s4
1317
1314
; GFX9DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
1318
1315
; GFX9DAGISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
1319
1316
; GFX9DAGISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1332,9 +1329,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
1332
1329
; GFX9GISEL-LABEL: divergent_value_i64:
1333
1330
; GFX9GISEL: ; %bb.0: ; %entry
1334
1331
; GFX9GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1335
- ; GFX9GISEL-NEXT: s_mov_b32 s4 , -1
1332
+ ; GFX9GISEL-NEXT: s_mov_b64 s[4:5] , -1
1336
1333
; GFX9GISEL-NEXT: s_mov_b64 s[6:7], exec
1337
- ; GFX9GISEL-NEXT: s_mov_b32 s5, s4
1338
1334
; GFX9GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
1339
1335
; GFX9GISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
1340
1336
; GFX9GISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1353,9 +1349,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
1353
1349
; GFX1064DAGISEL-LABEL: divergent_value_i64:
1354
1350
; GFX1064DAGISEL: ; %bb.0: ; %entry
1355
1351
; GFX1064DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1356
- ; GFX1064DAGISEL-NEXT: s_mov_b32 s4 , -1
1352
+ ; GFX1064DAGISEL-NEXT: s_mov_b64 s[4:5] , -1
1357
1353
; GFX1064DAGISEL-NEXT: s_mov_b64 s[6:7], exec
1358
- ; GFX1064DAGISEL-NEXT: s_mov_b32 s5, s4
1359
1354
; GFX1064DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
1360
1355
; GFX1064DAGISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
1361
1356
; GFX1064DAGISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1373,9 +1368,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
1373
1368
; GFX1064GISEL-LABEL: divergent_value_i64:
1374
1369
; GFX1064GISEL: ; %bb.0: ; %entry
1375
1370
; GFX1064GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1376
- ; GFX1064GISEL-NEXT: s_mov_b32 s4 , -1
1371
+ ; GFX1064GISEL-NEXT: s_mov_b64 s[4:5] , -1
1377
1372
; GFX1064GISEL-NEXT: s_mov_b64 s[6:7], exec
1378
- ; GFX1064GISEL-NEXT: s_mov_b32 s5, s4
1379
1373
; GFX1064GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
1380
1374
; GFX1064GISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
1381
1375
; GFX1064GISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1393,9 +1387,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
1393
1387
; GFX1032DAGISEL-LABEL: divergent_value_i64:
1394
1388
; GFX1032DAGISEL: ; %bb.0: ; %entry
1395
1389
; GFX1032DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1396
- ; GFX1032DAGISEL-NEXT: s_mov_b32 s4 , -1
1390
+ ; GFX1032DAGISEL-NEXT: s_mov_b64 s[4:5] , -1
1397
1391
; GFX1032DAGISEL-NEXT: s_mov_b32 s6, exec_lo
1398
- ; GFX1032DAGISEL-NEXT: s_mov_b32 s5, s4
1399
1392
; GFX1032DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
1400
1393
; GFX1032DAGISEL-NEXT: s_ff1_i32_b32 s7, s6
1401
1394
; GFX1032DAGISEL-NEXT: v_readlane_b32 s8, v2, s7
@@ -1413,9 +1406,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
1413
1406
; GFX1032GISEL-LABEL: divergent_value_i64:
1414
1407
; GFX1032GISEL: ; %bb.0: ; %entry
1415
1408
; GFX1032GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1416
- ; GFX1032GISEL-NEXT: s_mov_b32 s4 , -1
1409
+ ; GFX1032GISEL-NEXT: s_mov_b64 s[4:5] , -1
1417
1410
; GFX1032GISEL-NEXT: s_mov_b32 s6, exec_lo
1418
- ; GFX1032GISEL-NEXT: s_mov_b32 s5, s4
1419
1411
; GFX1032GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
1420
1412
; GFX1032GISEL-NEXT: s_ff1_i32_b32 s7, s6
1421
1413
; GFX1032GISEL-NEXT: v_readlane_b32 s8, v2, s7
@@ -1433,12 +1425,11 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
1433
1425
; GFX1164DAGISEL-LABEL: divergent_value_i64:
1434
1426
; GFX1164DAGISEL: ; %bb.0: ; %entry
1435
1427
; GFX1164DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1436
- ; GFX1164DAGISEL-NEXT: s_mov_b32 s0 , -1
1428
+ ; GFX1164DAGISEL-NEXT: s_mov_b64 s[0:1] , -1
1437
1429
; GFX1164DAGISEL-NEXT: s_mov_b64 s[2:3], exec
1438
- ; GFX1164DAGISEL-NEXT: s_mov_b32 s1, s0
1439
1430
; GFX1164DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
1431
+ ; GFX1164DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
1440
1432
; GFX1164DAGISEL-NEXT: s_ctz_i32_b64 s6, s[2:3]
1441
- ; GFX1164DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
1442
1433
; GFX1164DAGISEL-NEXT: v_readlane_b32 s4, v2, s6
1443
1434
; GFX1164DAGISEL-NEXT: v_readlane_b32 s5, v3, s6
1444
1435
; GFX1164DAGISEL-NEXT: s_bitset0_b64 s[2:3], s6
@@ -1454,12 +1445,11 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
1454
1445
; GFX1164GISEL-LABEL: divergent_value_i64:
1455
1446
; GFX1164GISEL: ; %bb.0: ; %entry
1456
1447
; GFX1164GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1457
- ; GFX1164GISEL-NEXT: s_mov_b32 s0 , -1
1448
+ ; GFX1164GISEL-NEXT: s_mov_b64 s[0:1] , -1
1458
1449
; GFX1164GISEL-NEXT: s_mov_b64 s[2:3], exec
1459
- ; GFX1164GISEL-NEXT: s_mov_b32 s1, s0
1460
1450
; GFX1164GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
1451
+ ; GFX1164GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
1461
1452
; GFX1164GISEL-NEXT: s_ctz_i32_b64 s6, s[2:3]
1462
- ; GFX1164GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
1463
1453
; GFX1164GISEL-NEXT: v_readlane_b32 s4, v2, s6
1464
1454
; GFX1164GISEL-NEXT: v_readlane_b32 s5, v3, s6
1465
1455
; GFX1164GISEL-NEXT: s_bitset0_b64 s[2:3], s6
@@ -1475,12 +1465,11 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
1475
1465
; GFX1132DAGISEL-LABEL: divergent_value_i64:
1476
1466
; GFX1132DAGISEL: ; %bb.0: ; %entry
1477
1467
; GFX1132DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1478
- ; GFX1132DAGISEL-NEXT: s_mov_b32 s0 , -1
1468
+ ; GFX1132DAGISEL-NEXT: s_mov_b64 s[0:1] , -1
1479
1469
; GFX1132DAGISEL-NEXT: s_mov_b32 s2, exec_lo
1480
- ; GFX1132DAGISEL-NEXT: s_mov_b32 s1, s0
1481
1470
; GFX1132DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
1471
+ ; GFX1132DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
1482
1472
; GFX1132DAGISEL-NEXT: s_ctz_i32_b32 s3, s2
1483
- ; GFX1132DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
1484
1473
; GFX1132DAGISEL-NEXT: v_readlane_b32 s4, v2, s3
1485
1474
; GFX1132DAGISEL-NEXT: v_readlane_b32 s5, v3, s3
1486
1475
; GFX1132DAGISEL-NEXT: s_bitset0_b32 s2, s3
@@ -1495,12 +1484,11 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
1495
1484
; GFX1132GISEL-LABEL: divergent_value_i64:
1496
1485
; GFX1132GISEL: ; %bb.0: ; %entry
1497
1486
; GFX1132GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1498
- ; GFX1132GISEL-NEXT: s_mov_b32 s0 , -1
1487
+ ; GFX1132GISEL-NEXT: s_mov_b64 s[0:1] , -1
1499
1488
; GFX1132GISEL-NEXT: s_mov_b32 s2, exec_lo
1500
- ; GFX1132GISEL-NEXT: s_mov_b32 s1, s0
1501
1489
; GFX1132GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
1490
+ ; GFX1132GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
1502
1491
; GFX1132GISEL-NEXT: s_ctz_i32_b32 s3, s2
1503
- ; GFX1132GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
1504
1492
; GFX1132GISEL-NEXT: v_readlane_b32 s4, v2, s3
1505
1493
; GFX1132GISEL-NEXT: v_readlane_b32 s5, v3, s3
1506
1494
; GFX1132GISEL-NEXT: s_bitset0_b32 s2, s3
0 commit comments