Skip to content

Commit 460ec42

Browse files
committed
Code Formating
1 parent 15b3c66 commit 460ec42

File tree

4 files changed

+55
-90
lines changed

4 files changed

+55
-90
lines changed

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

Lines changed: 7 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -5257,12 +5257,9 @@ static uint32_t getIdentityValueFor32BitWaveReduction(unsigned Opc) {
52575257
case AMDGPU::S_ADD_I32:
52585258
case AMDGPU::S_SUB_I32:
52595259
case AMDGPU::S_OR_B32:
5260-
case AMDGPU::S_OR_B64:
52615260
case AMDGPU::S_XOR_B32:
5262-
case AMDGPU::S_XOR_B64:
52635261
return std::numeric_limits<uint32_t>::min();
52645262
case AMDGPU::S_AND_B32:
5265-
case AMDGPU::S_AND_B64:
52665263
return std::numeric_limits<uint32_t>::max();
52675264
default:
52685265
llvm_unreachable(
@@ -5282,7 +5279,11 @@ static uint64_t getIdentityValueFor64BitWaveReduction(unsigned Opc) {
52825279
return std::numeric_limits<int64_t>::min();
52835280
case AMDGPU::S_ADD_U64_PSEUDO:
52845281
case AMDGPU::S_SUB_U64_PSEUDO:
5282+
case AMDGPU::S_OR_B64:
5283+
case AMDGPU::S_XOR_B64:
52855284
return std::numeric_limits<uint64_t>::min();
5285+
case AMDGPU::S_AND_B64:
5286+
return std::numeric_limits<uint64_t>::max();
52865287
default:
52875288
llvm_unreachable(
52885289
"Unexpected opcode in getIdentityValueFor64BitWaveReduction");
@@ -5601,9 +5602,9 @@ static MachineBasicBlock *lowerWaveReduce(MachineInstr &MI,
56015602
.addReg(LaneValueHiReg)
56025603
.addImm(AMDGPU::sub1);
56035604
switch (Opc) {
5604-
case ::AMDGPU::S_OR_B64:
5605-
case ::AMDGPU::S_AND_B64:
5606-
case ::AMDGPU::S_XOR_B64: {
5605+
case AMDGPU::S_OR_B64:
5606+
case AMDGPU::S_AND_B64:
5607+
case AMDGPU::S_XOR_B64: {
56075608
NewAccumulator = BuildMI(*ComputeLoop, I, DL, TII->get(Opc), DstReg)
56085609
.addReg(Accumulator->getOperand(0).getReg())
56095610
.addReg(LaneValue->getOperand(0).getReg())

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.and.ll

Lines changed: 16 additions & 28 deletions
Original file line numberDiff line numberDiff line change
@@ -1269,9 +1269,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
12691269
; GFX8DAGISEL-LABEL: divergent_value_i64:
12701270
; GFX8DAGISEL: ; %bb.0: ; %entry
12711271
; 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
12731273
; GFX8DAGISEL-NEXT: s_mov_b64 s[6:7], exec
1274-
; GFX8DAGISEL-NEXT: s_mov_b32 s5, s4
12751274
; GFX8DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
12761275
; GFX8DAGISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
12771276
; GFX8DAGISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1290,9 +1289,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
12901289
; GFX8GISEL-LABEL: divergent_value_i64:
12911290
; GFX8GISEL: ; %bb.0: ; %entry
12921291
; 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
12941293
; GFX8GISEL-NEXT: s_mov_b64 s[6:7], exec
1295-
; GFX8GISEL-NEXT: s_mov_b32 s5, s4
12961294
; GFX8GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
12971295
; GFX8GISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
12981296
; GFX8GISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1311,9 +1309,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
13111309
; GFX9DAGISEL-LABEL: divergent_value_i64:
13121310
; GFX9DAGISEL: ; %bb.0: ; %entry
13131311
; 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
13151313
; GFX9DAGISEL-NEXT: s_mov_b64 s[6:7], exec
1316-
; GFX9DAGISEL-NEXT: s_mov_b32 s5, s4
13171314
; GFX9DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
13181315
; GFX9DAGISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
13191316
; GFX9DAGISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1332,9 +1329,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
13321329
; GFX9GISEL-LABEL: divergent_value_i64:
13331330
; GFX9GISEL: ; %bb.0: ; %entry
13341331
; 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
13361333
; GFX9GISEL-NEXT: s_mov_b64 s[6:7], exec
1337-
; GFX9GISEL-NEXT: s_mov_b32 s5, s4
13381334
; GFX9GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
13391335
; GFX9GISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
13401336
; GFX9GISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1353,9 +1349,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
13531349
; GFX1064DAGISEL-LABEL: divergent_value_i64:
13541350
; GFX1064DAGISEL: ; %bb.0: ; %entry
13551351
; 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
13571353
; GFX1064DAGISEL-NEXT: s_mov_b64 s[6:7], exec
1358-
; GFX1064DAGISEL-NEXT: s_mov_b32 s5, s4
13591354
; GFX1064DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
13601355
; GFX1064DAGISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
13611356
; GFX1064DAGISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1373,9 +1368,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
13731368
; GFX1064GISEL-LABEL: divergent_value_i64:
13741369
; GFX1064GISEL: ; %bb.0: ; %entry
13751370
; 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
13771372
; GFX1064GISEL-NEXT: s_mov_b64 s[6:7], exec
1378-
; GFX1064GISEL-NEXT: s_mov_b32 s5, s4
13791373
; GFX1064GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
13801374
; GFX1064GISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
13811375
; GFX1064GISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1393,9 +1387,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
13931387
; GFX1032DAGISEL-LABEL: divergent_value_i64:
13941388
; GFX1032DAGISEL: ; %bb.0: ; %entry
13951389
; 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
13971391
; GFX1032DAGISEL-NEXT: s_mov_b32 s6, exec_lo
1398-
; GFX1032DAGISEL-NEXT: s_mov_b32 s5, s4
13991392
; GFX1032DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
14001393
; GFX1032DAGISEL-NEXT: s_ff1_i32_b32 s7, s6
14011394
; GFX1032DAGISEL-NEXT: v_readlane_b32 s8, v2, s7
@@ -1413,9 +1406,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
14131406
; GFX1032GISEL-LABEL: divergent_value_i64:
14141407
; GFX1032GISEL: ; %bb.0: ; %entry
14151408
; 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
14171410
; GFX1032GISEL-NEXT: s_mov_b32 s6, exec_lo
1418-
; GFX1032GISEL-NEXT: s_mov_b32 s5, s4
14191411
; GFX1032GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
14201412
; GFX1032GISEL-NEXT: s_ff1_i32_b32 s7, s6
14211413
; GFX1032GISEL-NEXT: v_readlane_b32 s8, v2, s7
@@ -1433,12 +1425,11 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
14331425
; GFX1164DAGISEL-LABEL: divergent_value_i64:
14341426
; GFX1164DAGISEL: ; %bb.0: ; %entry
14351427
; 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
14371429
; GFX1164DAGISEL-NEXT: s_mov_b64 s[2:3], exec
1438-
; GFX1164DAGISEL-NEXT: s_mov_b32 s1, s0
14391430
; 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)
14401432
; GFX1164DAGISEL-NEXT: s_ctz_i32_b64 s6, s[2:3]
1441-
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
14421433
; GFX1164DAGISEL-NEXT: v_readlane_b32 s4, v2, s6
14431434
; GFX1164DAGISEL-NEXT: v_readlane_b32 s5, v3, s6
14441435
; 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) {
14541445
; GFX1164GISEL-LABEL: divergent_value_i64:
14551446
; GFX1164GISEL: ; %bb.0: ; %entry
14561447
; 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
14581449
; GFX1164GISEL-NEXT: s_mov_b64 s[2:3], exec
1459-
; GFX1164GISEL-NEXT: s_mov_b32 s1, s0
14601450
; 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)
14611452
; GFX1164GISEL-NEXT: s_ctz_i32_b64 s6, s[2:3]
1462-
; GFX1164GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
14631453
; GFX1164GISEL-NEXT: v_readlane_b32 s4, v2, s6
14641454
; GFX1164GISEL-NEXT: v_readlane_b32 s5, v3, s6
14651455
; 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) {
14751465
; GFX1132DAGISEL-LABEL: divergent_value_i64:
14761466
; GFX1132DAGISEL: ; %bb.0: ; %entry
14771467
; 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
14791469
; GFX1132DAGISEL-NEXT: s_mov_b32 s2, exec_lo
1480-
; GFX1132DAGISEL-NEXT: s_mov_b32 s1, s0
14811470
; 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)
14821472
; GFX1132DAGISEL-NEXT: s_ctz_i32_b32 s3, s2
1483-
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
14841473
; GFX1132DAGISEL-NEXT: v_readlane_b32 s4, v2, s3
14851474
; GFX1132DAGISEL-NEXT: v_readlane_b32 s5, v3, s3
14861475
; GFX1132DAGISEL-NEXT: s_bitset0_b32 s2, s3
@@ -1495,12 +1484,11 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
14951484
; GFX1132GISEL-LABEL: divergent_value_i64:
14961485
; GFX1132GISEL: ; %bb.0: ; %entry
14971486
; 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
14991488
; GFX1132GISEL-NEXT: s_mov_b32 s2, exec_lo
1500-
; GFX1132GISEL-NEXT: s_mov_b32 s1, s0
15011489
; 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)
15021491
; GFX1132GISEL-NEXT: s_ctz_i32_b32 s3, s2
1503-
; GFX1132GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
15041492
; GFX1132GISEL-NEXT: v_readlane_b32 s4, v2, s3
15051493
; GFX1132GISEL-NEXT: v_readlane_b32 s5, v3, s3
15061494
; GFX1132GISEL-NEXT: s_bitset0_b32 s2, s3

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.reduce.or.ll

Lines changed: 16 additions & 28 deletions
Original file line numberDiff line numberDiff line change
@@ -1269,9 +1269,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
12691269
; GFX8DAGISEL-LABEL: divergent_value_i64:
12701270
; GFX8DAGISEL: ; %bb.0: ; %entry
12711271
; GFX8DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1272-
; GFX8DAGISEL-NEXT: s_mov_b32 s4, 0
1272+
; GFX8DAGISEL-NEXT: s_mov_b64 s[4:5], 0
12731273
; GFX8DAGISEL-NEXT: s_mov_b64 s[6:7], exec
1274-
; GFX8DAGISEL-NEXT: s_mov_b32 s5, s4
12751274
; GFX8DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
12761275
; GFX8DAGISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
12771276
; GFX8DAGISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1290,9 +1289,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
12901289
; GFX8GISEL-LABEL: divergent_value_i64:
12911290
; GFX8GISEL: ; %bb.0: ; %entry
12921291
; GFX8GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1293-
; GFX8GISEL-NEXT: s_mov_b32 s4, 0
1292+
; GFX8GISEL-NEXT: s_mov_b64 s[4:5], 0
12941293
; GFX8GISEL-NEXT: s_mov_b64 s[6:7], exec
1295-
; GFX8GISEL-NEXT: s_mov_b32 s5, s4
12961294
; GFX8GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
12971295
; GFX8GISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
12981296
; GFX8GISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1311,9 +1309,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
13111309
; GFX9DAGISEL-LABEL: divergent_value_i64:
13121310
; GFX9DAGISEL: ; %bb.0: ; %entry
13131311
; GFX9DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1314-
; GFX9DAGISEL-NEXT: s_mov_b32 s4, 0
1312+
; GFX9DAGISEL-NEXT: s_mov_b64 s[4:5], 0
13151313
; GFX9DAGISEL-NEXT: s_mov_b64 s[6:7], exec
1316-
; GFX9DAGISEL-NEXT: s_mov_b32 s5, s4
13171314
; GFX9DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
13181315
; GFX9DAGISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
13191316
; GFX9DAGISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1332,9 +1329,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
13321329
; GFX9GISEL-LABEL: divergent_value_i64:
13331330
; GFX9GISEL: ; %bb.0: ; %entry
13341331
; GFX9GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1335-
; GFX9GISEL-NEXT: s_mov_b32 s4, 0
1332+
; GFX9GISEL-NEXT: s_mov_b64 s[4:5], 0
13361333
; GFX9GISEL-NEXT: s_mov_b64 s[6:7], exec
1337-
; GFX9GISEL-NEXT: s_mov_b32 s5, s4
13381334
; GFX9GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
13391335
; GFX9GISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
13401336
; GFX9GISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1353,9 +1349,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
13531349
; GFX1064DAGISEL-LABEL: divergent_value_i64:
13541350
; GFX1064DAGISEL: ; %bb.0: ; %entry
13551351
; GFX1064DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1356-
; GFX1064DAGISEL-NEXT: s_mov_b32 s4, 0
1352+
; GFX1064DAGISEL-NEXT: s_mov_b64 s[4:5], 0
13571353
; GFX1064DAGISEL-NEXT: s_mov_b64 s[6:7], exec
1358-
; GFX1064DAGISEL-NEXT: s_mov_b32 s5, s4
13591354
; GFX1064DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
13601355
; GFX1064DAGISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
13611356
; GFX1064DAGISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1373,9 +1368,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
13731368
; GFX1064GISEL-LABEL: divergent_value_i64:
13741369
; GFX1064GISEL: ; %bb.0: ; %entry
13751370
; GFX1064GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1376-
; GFX1064GISEL-NEXT: s_mov_b32 s4, 0
1371+
; GFX1064GISEL-NEXT: s_mov_b64 s[4:5], 0
13771372
; GFX1064GISEL-NEXT: s_mov_b64 s[6:7], exec
1378-
; GFX1064GISEL-NEXT: s_mov_b32 s5, s4
13791373
; GFX1064GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
13801374
; GFX1064GISEL-NEXT: s_ff1_i32_b64 s10, s[6:7]
13811375
; GFX1064GISEL-NEXT: v_readlane_b32 s8, v2, s10
@@ -1393,9 +1387,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
13931387
; GFX1032DAGISEL-LABEL: divergent_value_i64:
13941388
; GFX1032DAGISEL: ; %bb.0: ; %entry
13951389
; GFX1032DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1396-
; GFX1032DAGISEL-NEXT: s_mov_b32 s4, 0
1390+
; GFX1032DAGISEL-NEXT: s_mov_b64 s[4:5], 0
13971391
; GFX1032DAGISEL-NEXT: s_mov_b32 s6, exec_lo
1398-
; GFX1032DAGISEL-NEXT: s_mov_b32 s5, s4
13991392
; GFX1032DAGISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
14001393
; GFX1032DAGISEL-NEXT: s_ff1_i32_b32 s7, s6
14011394
; GFX1032DAGISEL-NEXT: v_readlane_b32 s8, v2, s7
@@ -1413,9 +1406,8 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
14131406
; GFX1032GISEL-LABEL: divergent_value_i64:
14141407
; GFX1032GISEL: ; %bb.0: ; %entry
14151408
; GFX1032GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1416-
; GFX1032GISEL-NEXT: s_mov_b32 s4, 0
1409+
; GFX1032GISEL-NEXT: s_mov_b64 s[4:5], 0
14171410
; GFX1032GISEL-NEXT: s_mov_b32 s6, exec_lo
1418-
; GFX1032GISEL-NEXT: s_mov_b32 s5, s4
14191411
; GFX1032GISEL-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
14201412
; GFX1032GISEL-NEXT: s_ff1_i32_b32 s7, s6
14211413
; GFX1032GISEL-NEXT: v_readlane_b32 s8, v2, s7
@@ -1433,12 +1425,11 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
14331425
; GFX1164DAGISEL-LABEL: divergent_value_i64:
14341426
; GFX1164DAGISEL: ; %bb.0: ; %entry
14351427
; GFX1164DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1436-
; GFX1164DAGISEL-NEXT: s_mov_b32 s0, 0
1428+
; GFX1164DAGISEL-NEXT: s_mov_b64 s[0:1], 0
14371429
; GFX1164DAGISEL-NEXT: s_mov_b64 s[2:3], exec
1438-
; GFX1164DAGISEL-NEXT: s_mov_b32 s1, s0
14391430
; 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)
14401432
; GFX1164DAGISEL-NEXT: s_ctz_i32_b64 s6, s[2:3]
1441-
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
14421433
; GFX1164DAGISEL-NEXT: v_readlane_b32 s4, v2, s6
14431434
; GFX1164DAGISEL-NEXT: v_readlane_b32 s5, v3, s6
14441435
; 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) {
14541445
; GFX1164GISEL-LABEL: divergent_value_i64:
14551446
; GFX1164GISEL: ; %bb.0: ; %entry
14561447
; GFX1164GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1457-
; GFX1164GISEL-NEXT: s_mov_b32 s0, 0
1448+
; GFX1164GISEL-NEXT: s_mov_b64 s[0:1], 0
14581449
; GFX1164GISEL-NEXT: s_mov_b64 s[2:3], exec
1459-
; GFX1164GISEL-NEXT: s_mov_b32 s1, s0
14601450
; 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)
14611452
; GFX1164GISEL-NEXT: s_ctz_i32_b64 s6, s[2:3]
1462-
; GFX1164GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
14631453
; GFX1164GISEL-NEXT: v_readlane_b32 s4, v2, s6
14641454
; GFX1164GISEL-NEXT: v_readlane_b32 s5, v3, s6
14651455
; 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) {
14751465
; GFX1132DAGISEL-LABEL: divergent_value_i64:
14761466
; GFX1132DAGISEL: ; %bb.0: ; %entry
14771467
; GFX1132DAGISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1478-
; GFX1132DAGISEL-NEXT: s_mov_b32 s0, 0
1468+
; GFX1132DAGISEL-NEXT: s_mov_b64 s[0:1], 0
14791469
; GFX1132DAGISEL-NEXT: s_mov_b32 s2, exec_lo
1480-
; GFX1132DAGISEL-NEXT: s_mov_b32 s1, s0
14811470
; 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)
14821472
; GFX1132DAGISEL-NEXT: s_ctz_i32_b32 s3, s2
1483-
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
14841473
; GFX1132DAGISEL-NEXT: v_readlane_b32 s4, v2, s3
14851474
; GFX1132DAGISEL-NEXT: v_readlane_b32 s5, v3, s3
14861475
; GFX1132DAGISEL-NEXT: s_bitset0_b32 s2, s3
@@ -1495,12 +1484,11 @@ define void @divergent_value_i64(ptr addrspace(1) %out, i64 %id.x) {
14951484
; GFX1132GISEL-LABEL: divergent_value_i64:
14961485
; GFX1132GISEL: ; %bb.0: ; %entry
14971486
; GFX1132GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1498-
; GFX1132GISEL-NEXT: s_mov_b32 s0, 0
1487+
; GFX1132GISEL-NEXT: s_mov_b64 s[0:1], 0
14991488
; GFX1132GISEL-NEXT: s_mov_b32 s2, exec_lo
1500-
; GFX1132GISEL-NEXT: s_mov_b32 s1, s0
15011489
; 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)
15021491
; GFX1132GISEL-NEXT: s_ctz_i32_b32 s3, s2
1503-
; GFX1132GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
15041492
; GFX1132GISEL-NEXT: v_readlane_b32 s4, v2, s3
15051493
; GFX1132GISEL-NEXT: v_readlane_b32 s5, v3, s3
15061494
; GFX1132GISEL-NEXT: s_bitset0_b32 s2, s3

0 commit comments

Comments
 (0)