Skip to content

Commit 44e2454

Browse files
committed
Rebase against main
1 parent 869e4ce commit 44e2454

14 files changed

+35
-63
lines changed

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

Lines changed: 4 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -665,7 +665,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
665665
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
666666
; GFX1164DAGISEL-NEXT: v_readlane_b32 s6, v0, s5
667667
; GFX1164DAGISEL-NEXT: s_bitset0_b64 s[2:3], s5
668-
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
669668
; GFX1164DAGISEL-NEXT: s_add_i32 s4, s4, s6
670669
; GFX1164DAGISEL-NEXT: s_cmp_lg_u64 s[2:3], 0
671670
; GFX1164DAGISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -686,7 +685,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
686685
; GFX1164GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
687686
; GFX1164GISEL-NEXT: v_readlane_b32 s6, v0, s5
688687
; GFX1164GISEL-NEXT: s_bitset0_b64 s[2:3], s5
689-
; GFX1164GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
690688
; GFX1164GISEL-NEXT: s_add_i32 s4, s4, s6
691689
; GFX1164GISEL-NEXT: s_cmp_lg_u64 s[2:3], 0
692690
; GFX1164GISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -708,7 +706,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
708706
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
709707
; GFX1132DAGISEL-NEXT: v_readlane_b32 s5, v0, s4
710708
; GFX1132DAGISEL-NEXT: s_bitset0_b32 s3, s4
711-
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
712709
; GFX1132DAGISEL-NEXT: s_add_i32 s2, s2, s5
713710
; GFX1132DAGISEL-NEXT: s_cmp_lg_u32 s3, 0
714711
; GFX1132DAGISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -729,7 +726,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
729726
; GFX1132GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
730727
; GFX1132GISEL-NEXT: v_readlane_b32 s5, v0, s4
731728
; GFX1132GISEL-NEXT: s_bitset0_b32 s3, s4
732-
; GFX1132GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
733729
; GFX1132GISEL-NEXT: s_add_i32 s2, s2, s5
734730
; GFX1132GISEL-NEXT: s_cmp_lg_u32 s3, 0
735731
; GFX1132GISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -1075,7 +1071,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
10751071
; GFX1164DAGISEL-NEXT: s_mov_b32 s6, 0
10761072
; GFX1164DAGISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
10771073
; GFX1164DAGISEL-NEXT: s_ctz_i32_b64 s7, s[2:3]
1078-
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
1074+
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
10791075
; GFX1164DAGISEL-NEXT: v_readlane_b32 s8, v0, s7
10801076
; GFX1164DAGISEL-NEXT: s_bitset0_b64 s[2:3], s7
10811077
; GFX1164DAGISEL-NEXT: s_add_i32 s6, s6, s8
@@ -1116,7 +1112,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
11161112
; GFX1164GISEL-NEXT: s_mov_b32 s6, 0
11171113
; GFX1164GISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
11181114
; GFX1164GISEL-NEXT: s_ctz_i32_b64 s7, s[2:3]
1119-
; GFX1164GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
1115+
; GFX1164GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
11201116
; GFX1164GISEL-NEXT: v_readlane_b32 s8, v0, s7
11211117
; GFX1164GISEL-NEXT: s_bitset0_b64 s[2:3], s7
11221118
; GFX1164GISEL-NEXT: s_add_i32 s6, s6, s8
@@ -1158,7 +1154,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
11581154
; GFX1132DAGISEL-NEXT: s_mov_b32 s1, 0
11591155
; GFX1132DAGISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
11601156
; GFX1132DAGISEL-NEXT: s_ctz_i32_b32 s3, s2
1161-
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
1157+
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
11621158
; GFX1132DAGISEL-NEXT: v_readlane_b32 s6, v0, s3
11631159
; GFX1132DAGISEL-NEXT: s_bitset0_b32 s2, s3
11641160
; GFX1132DAGISEL-NEXT: s_add_i32 s1, s1, s6
@@ -1199,7 +1195,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
11991195
; GFX1132GISEL-NEXT: s_mov_b32 s0, 0
12001196
; GFX1132GISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
12011197
; GFX1132GISEL-NEXT: s_ctz_i32_b32 s3, s2
1202-
; GFX1132GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
1198+
; GFX1132GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
12031199
; GFX1132GISEL-NEXT: v_readlane_b32 s6, v0, s3
12041200
; GFX1132GISEL-NEXT: s_bitset0_b32 s2, s3
12051201
; GFX1132GISEL-NEXT: s_add_i32 s0, s0, s6

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

Lines changed: 4 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -459,7 +459,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
459459
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
460460
; GFX1164DAGISEL-NEXT: v_readlane_b32 s6, v0, s5
461461
; GFX1164DAGISEL-NEXT: s_bitset0_b64 s[2:3], s5
462-
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
463462
; GFX1164DAGISEL-NEXT: s_and_b32 s4, s4, s6
464463
; GFX1164DAGISEL-NEXT: s_cmp_lg_u64 s[2:3], 0
465464
; GFX1164DAGISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -480,7 +479,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
480479
; GFX1164GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
481480
; GFX1164GISEL-NEXT: v_readlane_b32 s6, v0, s5
482481
; GFX1164GISEL-NEXT: s_bitset0_b64 s[2:3], s5
483-
; GFX1164GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
484482
; GFX1164GISEL-NEXT: s_and_b32 s4, s4, s6
485483
; GFX1164GISEL-NEXT: s_cmp_lg_u64 s[2:3], 0
486484
; GFX1164GISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -502,7 +500,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
502500
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
503501
; GFX1132DAGISEL-NEXT: v_readlane_b32 s5, v0, s4
504502
; GFX1132DAGISEL-NEXT: s_bitset0_b32 s3, s4
505-
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
506503
; GFX1132DAGISEL-NEXT: s_and_b32 s2, s2, s5
507504
; GFX1132DAGISEL-NEXT: s_cmp_lg_u32 s3, 0
508505
; GFX1132DAGISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -523,7 +520,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
523520
; GFX1132GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
524521
; GFX1132GISEL-NEXT: v_readlane_b32 s5, v0, s4
525522
; GFX1132GISEL-NEXT: s_bitset0_b32 s3, s4
526-
; GFX1132GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
527523
; GFX1132GISEL-NEXT: s_and_b32 s2, s2, s5
528524
; GFX1132GISEL-NEXT: s_cmp_lg_u32 s3, 0
529525
; GFX1132GISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -840,7 +836,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
840836
; GFX1164DAGISEL-NEXT: s_mov_b32 s6, -1
841837
; GFX1164DAGISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
842838
; GFX1164DAGISEL-NEXT: s_ctz_i32_b64 s7, s[2:3]
843-
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
839+
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
844840
; GFX1164DAGISEL-NEXT: v_readlane_b32 s8, v0, s7
845841
; GFX1164DAGISEL-NEXT: s_bitset0_b64 s[2:3], s7
846842
; GFX1164DAGISEL-NEXT: s_and_b32 s6, s6, s8
@@ -878,7 +874,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
878874
; GFX1164GISEL-NEXT: s_mov_b32 s6, -1
879875
; GFX1164GISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
880876
; GFX1164GISEL-NEXT: s_ctz_i32_b64 s7, s[2:3]
881-
; GFX1164GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
877+
; GFX1164GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
882878
; GFX1164GISEL-NEXT: v_readlane_b32 s8, v0, s7
883879
; GFX1164GISEL-NEXT: s_bitset0_b64 s[2:3], s7
884880
; GFX1164GISEL-NEXT: s_and_b32 s6, s6, s8
@@ -915,7 +911,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
915911
; GFX1132DAGISEL-NEXT: s_mov_b32 s1, -1
916912
; GFX1132DAGISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
917913
; GFX1132DAGISEL-NEXT: s_ctz_i32_b32 s3, s2
918-
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
914+
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
919915
; GFX1132DAGISEL-NEXT: v_readlane_b32 s6, v0, s3
920916
; GFX1132DAGISEL-NEXT: s_bitset0_b32 s2, s3
921917
; GFX1132DAGISEL-NEXT: s_and_b32 s1, s1, s6
@@ -953,7 +949,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
953949
; GFX1132GISEL-NEXT: s_mov_b32 s0, -1
954950
; GFX1132GISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
955951
; GFX1132GISEL-NEXT: s_ctz_i32_b32 s3, s2
956-
; GFX1132GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
952+
; GFX1132GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
957953
; GFX1132GISEL-NEXT: v_readlane_b32 s6, v0, s3
958954
; GFX1132GISEL-NEXT: s_bitset0_b32 s2, s3
959955
; GFX1132GISEL-NEXT: s_and_b32 s0, s0, s6

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

Lines changed: 4 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -459,7 +459,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
459459
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
460460
; GFX1164DAGISEL-NEXT: v_readlane_b32 s6, v0, s5
461461
; GFX1164DAGISEL-NEXT: s_bitset0_b64 s[2:3], s5
462-
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
463462
; GFX1164DAGISEL-NEXT: s_max_i32 s4, s4, s6
464463
; GFX1164DAGISEL-NEXT: s_cmp_lg_u64 s[2:3], 0
465464
; GFX1164DAGISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -480,7 +479,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
480479
; GFX1164GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
481480
; GFX1164GISEL-NEXT: v_readlane_b32 s6, v0, s5
482481
; GFX1164GISEL-NEXT: s_bitset0_b64 s[2:3], s5
483-
; GFX1164GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
484482
; GFX1164GISEL-NEXT: s_max_i32 s4, s4, s6
485483
; GFX1164GISEL-NEXT: s_cmp_lg_u64 s[2:3], 0
486484
; GFX1164GISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -502,7 +500,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
502500
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
503501
; GFX1132DAGISEL-NEXT: v_readlane_b32 s5, v0, s4
504502
; GFX1132DAGISEL-NEXT: s_bitset0_b32 s3, s4
505-
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
506503
; GFX1132DAGISEL-NEXT: s_max_i32 s2, s2, s5
507504
; GFX1132DAGISEL-NEXT: s_cmp_lg_u32 s3, 0
508505
; GFX1132DAGISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -523,7 +520,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
523520
; GFX1132GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
524521
; GFX1132GISEL-NEXT: v_readlane_b32 s5, v0, s4
525522
; GFX1132GISEL-NEXT: s_bitset0_b32 s3, s4
526-
; GFX1132GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
527523
; GFX1132GISEL-NEXT: s_max_i32 s2, s2, s5
528524
; GFX1132GISEL-NEXT: s_cmp_lg_u32 s3, 0
529525
; GFX1132GISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -840,7 +836,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
840836
; GFX1164DAGISEL-NEXT: s_brev_b32 s6, 1
841837
; GFX1164DAGISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
842838
; GFX1164DAGISEL-NEXT: s_ctz_i32_b64 s7, s[2:3]
843-
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
839+
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
844840
; GFX1164DAGISEL-NEXT: v_readlane_b32 s8, v0, s7
845841
; GFX1164DAGISEL-NEXT: s_bitset0_b64 s[2:3], s7
846842
; GFX1164DAGISEL-NEXT: s_max_i32 s6, s6, s8
@@ -878,7 +874,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
878874
; GFX1164GISEL-NEXT: s_brev_b32 s6, 1
879875
; GFX1164GISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
880876
; GFX1164GISEL-NEXT: s_ctz_i32_b64 s7, s[2:3]
881-
; GFX1164GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
877+
; GFX1164GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
882878
; GFX1164GISEL-NEXT: v_readlane_b32 s8, v0, s7
883879
; GFX1164GISEL-NEXT: s_bitset0_b64 s[2:3], s7
884880
; GFX1164GISEL-NEXT: s_max_i32 s6, s6, s8
@@ -915,7 +911,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
915911
; GFX1132DAGISEL-NEXT: s_brev_b32 s1, 1
916912
; GFX1132DAGISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
917913
; GFX1132DAGISEL-NEXT: s_ctz_i32_b32 s3, s2
918-
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
914+
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
919915
; GFX1132DAGISEL-NEXT: v_readlane_b32 s6, v0, s3
920916
; GFX1132DAGISEL-NEXT: s_bitset0_b32 s2, s3
921917
; GFX1132DAGISEL-NEXT: s_max_i32 s1, s1, s6
@@ -953,7 +949,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
953949
; GFX1132GISEL-NEXT: s_brev_b32 s0, 1
954950
; GFX1132GISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
955951
; GFX1132GISEL-NEXT: s_ctz_i32_b32 s3, s2
956-
; GFX1132GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
952+
; GFX1132GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
957953
; GFX1132GISEL-NEXT: v_readlane_b32 s6, v0, s3
958954
; GFX1132GISEL-NEXT: s_bitset0_b32 s2, s3
959955
; GFX1132GISEL-NEXT: s_max_i32 s0, s0, s6

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

Lines changed: 4 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -459,7 +459,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
459459
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
460460
; GFX1164DAGISEL-NEXT: v_readlane_b32 s6, v0, s5
461461
; GFX1164DAGISEL-NEXT: s_bitset0_b64 s[2:3], s5
462-
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
463462
; GFX1164DAGISEL-NEXT: s_min_i32 s4, s4, s6
464463
; GFX1164DAGISEL-NEXT: s_cmp_lg_u64 s[2:3], 0
465464
; GFX1164DAGISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -480,7 +479,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
480479
; GFX1164GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
481480
; GFX1164GISEL-NEXT: v_readlane_b32 s6, v0, s5
482481
; GFX1164GISEL-NEXT: s_bitset0_b64 s[2:3], s5
483-
; GFX1164GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
484482
; GFX1164GISEL-NEXT: s_min_i32 s4, s4, s6
485483
; GFX1164GISEL-NEXT: s_cmp_lg_u64 s[2:3], 0
486484
; GFX1164GISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -502,7 +500,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
502500
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
503501
; GFX1132DAGISEL-NEXT: v_readlane_b32 s5, v0, s4
504502
; GFX1132DAGISEL-NEXT: s_bitset0_b32 s3, s4
505-
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
506503
; GFX1132DAGISEL-NEXT: s_min_i32 s2, s2, s5
507504
; GFX1132DAGISEL-NEXT: s_cmp_lg_u32 s3, 0
508505
; GFX1132DAGISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -523,7 +520,6 @@ define amdgpu_kernel void @divergent_value(ptr addrspace(1) %out) {
523520
; GFX1132GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
524521
; GFX1132GISEL-NEXT: v_readlane_b32 s5, v0, s4
525522
; GFX1132GISEL-NEXT: s_bitset0_b32 s3, s4
526-
; GFX1132GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
527523
; GFX1132GISEL-NEXT: s_min_i32 s2, s2, s5
528524
; GFX1132GISEL-NEXT: s_cmp_lg_u32 s3, 0
529525
; GFX1132GISEL-NEXT: s_cbranch_scc1 .LBB3_1
@@ -840,7 +836,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
840836
; GFX1164DAGISEL-NEXT: s_brev_b32 s6, -2
841837
; GFX1164DAGISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
842838
; GFX1164DAGISEL-NEXT: s_ctz_i32_b64 s7, s[2:3]
843-
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
839+
; GFX1164DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
844840
; GFX1164DAGISEL-NEXT: v_readlane_b32 s8, v0, s7
845841
; GFX1164DAGISEL-NEXT: s_bitset0_b64 s[2:3], s7
846842
; GFX1164DAGISEL-NEXT: s_min_i32 s6, s6, s8
@@ -878,7 +874,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
878874
; GFX1164GISEL-NEXT: s_brev_b32 s6, -2
879875
; GFX1164GISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
880876
; GFX1164GISEL-NEXT: s_ctz_i32_b64 s7, s[2:3]
881-
; GFX1164GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
877+
; GFX1164GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
882878
; GFX1164GISEL-NEXT: v_readlane_b32 s8, v0, s7
883879
; GFX1164GISEL-NEXT: s_bitset0_b64 s[2:3], s7
884880
; GFX1164GISEL-NEXT: s_min_i32 s6, s6, s8
@@ -915,7 +911,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
915911
; GFX1132DAGISEL-NEXT: s_brev_b32 s1, -2
916912
; GFX1132DAGISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
917913
; GFX1132DAGISEL-NEXT: s_ctz_i32_b32 s3, s2
918-
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
914+
; GFX1132DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
919915
; GFX1132DAGISEL-NEXT: v_readlane_b32 s6, v0, s3
920916
; GFX1132DAGISEL-NEXT: s_bitset0_b32 s2, s3
921917
; GFX1132DAGISEL-NEXT: s_min_i32 s1, s1, s6
@@ -953,7 +949,7 @@ define amdgpu_kernel void @divergent_cfg(ptr addrspace(1) %out, i32 %in) {
953949
; GFX1132GISEL-NEXT: s_brev_b32 s0, -2
954950
; GFX1132GISEL-NEXT: .LBB4_4: ; =>This Inner Loop Header: Depth=1
955951
; GFX1132GISEL-NEXT: s_ctz_i32_b32 s3, s2
956-
; GFX1132GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
952+
; GFX1132GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
957953
; GFX1132GISEL-NEXT: v_readlane_b32 s6, v0, s3
958954
; GFX1132GISEL-NEXT: s_bitset0_b32 s2, s3
959955
; GFX1132GISEL-NEXT: s_min_i32 s0, s0, s6

0 commit comments

Comments
 (0)