@@ -1070,37 +1070,35 @@ define amdgpu_kernel void @v8i8_phi_zeroinit(ptr addrspace(1) %src1, ptr addrspa
1070
1070
; FEATURE-NEXT: v_lshlrev_b32_e32 v5, 3, v0
1071
1071
; FEATURE-NEXT: v_cmp_lt_u32_e64 s[0:1], 14, v0
1072
1072
; FEATURE-NEXT: v_cmp_gt_u32_e32 vcc, 15, v0
1073
- ; FEATURE-NEXT: ; implicit-def: $vgpr1_vgpr2
1073
+ ; FEATURE-NEXT: ; implicit-def: $vgpr3_vgpr4
1074
1074
; FEATURE-NEXT: s_waitcnt lgkmcnt(0)
1075
- ; FEATURE-NEXT: global_load_dwordx2 v[3:4 ], v5, s[8:9]
1075
+ ; FEATURE-NEXT: global_load_dwordx2 v[1:2 ], v5, s[8:9]
1076
1076
; FEATURE-NEXT: s_and_saveexec_b64 s[2:3], vcc
1077
1077
; FEATURE-NEXT: s_cbranch_execz .LBB9_2
1078
1078
; FEATURE-NEXT: ; %bb.1: ; %bb.1
1079
- ; FEATURE-NEXT: global_load_dwordx2 v[1:2], v5, s[10:11]
1080
- ; FEATURE-NEXT: s_mov_b32 s4, 0
1079
+ ; FEATURE-NEXT: global_load_dwordx2 v[3:4], v5, s[10:11]
1081
1080
; FEATURE-NEXT: v_cmp_gt_u32_e32 vcc, 7, v0
1082
- ; FEATURE-NEXT: s_mov_b32 s5, s4
1083
1081
; FEATURE-NEXT: s_waitcnt vmcnt(1)
1084
- ; FEATURE-NEXT: v_mov_b32_e32 v3, s4
1085
- ; FEATURE-NEXT: v_mov_b32_e32 v4, s5
1082
+ ; FEATURE-NEXT: v_mov_b32_e32 v1, 0
1086
1083
; FEATURE-NEXT: s_andn2_b64 s[0:1], s[0:1], exec
1087
1084
; FEATURE-NEXT: s_and_b64 s[4:5], vcc, exec
1085
+ ; FEATURE-NEXT: v_mov_b32_e32 v2, v1
1088
1086
; FEATURE-NEXT: s_or_b64 s[0:1], s[0:1], s[4:5]
1089
1087
; FEATURE-NEXT: .LBB9_2: ; %Flow
1090
1088
; FEATURE-NEXT: s_or_b64 exec, exec, s[2:3]
1091
1089
; FEATURE-NEXT: s_and_saveexec_b64 s[2:3], s[0:1]
1092
1090
; FEATURE-NEXT: s_cbranch_execz .LBB9_4
1093
1091
; FEATURE-NEXT: ; %bb.3: ; %bb.2
1094
1092
; FEATURE-NEXT: s_waitcnt vmcnt(0)
1095
- ; FEATURE-NEXT: v_mov_b32_e32 v1, v3
1093
+ ; FEATURE-NEXT: v_mov_b32_e32 v4, v2
1096
1094
; FEATURE-NEXT: v_mov_b32_e32 v0, 0
1097
- ; FEATURE-NEXT: v_mov_b32_e32 v2, v4
1098
- ; FEATURE-NEXT: global_store_dwordx2 v0, v[3:4 ], s[12:13]
1095
+ ; FEATURE-NEXT: v_mov_b32_e32 v3, v1
1096
+ ; FEATURE-NEXT: global_store_dwordx2 v0, v[1:2 ], s[12:13]
1099
1097
; FEATURE-NEXT: .LBB9_4: ; %bb.3
1100
1098
; FEATURE-NEXT: s_or_b64 exec, exec, s[2:3]
1101
1099
; FEATURE-NEXT: v_mov_b32_e32 v0, 0
1102
1100
; FEATURE-NEXT: s_waitcnt vmcnt(0)
1103
- ; FEATURE-NEXT: global_store_dwordx2 v0, v[1:2 ], s[14:15]
1101
+ ; FEATURE-NEXT: global_store_dwordx2 v0, v[3:4 ], s[14:15]
1104
1102
; FEATURE-NEXT: s_endpgm
1105
1103
;
1106
1104
; DEFAULT-LABEL: v8i8_phi_zeroinit:
@@ -1109,37 +1107,35 @@ define amdgpu_kernel void @v8i8_phi_zeroinit(ptr addrspace(1) %src1, ptr addrspa
1109
1107
; DEFAULT-NEXT: v_lshlrev_b32_e32 v5, 3, v0
1110
1108
; DEFAULT-NEXT: v_cmp_lt_u32_e64 s[0:1], 14, v0
1111
1109
; DEFAULT-NEXT: v_cmp_gt_u32_e32 vcc, 15, v0
1112
- ; DEFAULT-NEXT: ; implicit-def: $vgpr1_vgpr2
1110
+ ; DEFAULT-NEXT: ; implicit-def: $vgpr3_vgpr4
1113
1111
; DEFAULT-NEXT: s_waitcnt lgkmcnt(0)
1114
- ; DEFAULT-NEXT: global_load_dwordx2 v[3:4 ], v5, s[8:9]
1112
+ ; DEFAULT-NEXT: global_load_dwordx2 v[1:2 ], v5, s[8:9]
1115
1113
; DEFAULT-NEXT: s_and_saveexec_b64 s[2:3], vcc
1116
1114
; DEFAULT-NEXT: s_cbranch_execz .LBB9_2
1117
1115
; DEFAULT-NEXT: ; %bb.1: ; %bb.1
1118
- ; DEFAULT-NEXT: global_load_dwordx2 v[1:2], v5, s[10:11]
1119
- ; DEFAULT-NEXT: s_mov_b32 s4, 0
1116
+ ; DEFAULT-NEXT: global_load_dwordx2 v[3:4], v5, s[10:11]
1120
1117
; DEFAULT-NEXT: v_cmp_gt_u32_e32 vcc, 7, v0
1121
- ; DEFAULT-NEXT: s_mov_b32 s5, s4
1122
1118
; DEFAULT-NEXT: s_waitcnt vmcnt(1)
1123
- ; DEFAULT-NEXT: v_mov_b32_e32 v3, s4
1124
- ; DEFAULT-NEXT: v_mov_b32_e32 v4, s5
1119
+ ; DEFAULT-NEXT: v_mov_b32_e32 v1, 0
1125
1120
; DEFAULT-NEXT: s_andn2_b64 s[0:1], s[0:1], exec
1126
1121
; DEFAULT-NEXT: s_and_b64 s[4:5], vcc, exec
1122
+ ; DEFAULT-NEXT: v_mov_b32_e32 v2, v1
1127
1123
; DEFAULT-NEXT: s_or_b64 s[0:1], s[0:1], s[4:5]
1128
1124
; DEFAULT-NEXT: .LBB9_2: ; %Flow
1129
1125
; DEFAULT-NEXT: s_or_b64 exec, exec, s[2:3]
1130
1126
; DEFAULT-NEXT: s_and_saveexec_b64 s[2:3], s[0:1]
1131
1127
; DEFAULT-NEXT: s_cbranch_execz .LBB9_4
1132
1128
; DEFAULT-NEXT: ; %bb.3: ; %bb.2
1133
1129
; DEFAULT-NEXT: s_waitcnt vmcnt(0)
1134
- ; DEFAULT-NEXT: v_mov_b32_e32 v1, v3
1130
+ ; DEFAULT-NEXT: v_mov_b32_e32 v4, v2
1135
1131
; DEFAULT-NEXT: v_mov_b32_e32 v0, 0
1136
- ; DEFAULT-NEXT: v_mov_b32_e32 v2, v4
1137
- ; DEFAULT-NEXT: global_store_dwordx2 v0, v[3:4 ], s[12:13]
1132
+ ; DEFAULT-NEXT: v_mov_b32_e32 v3, v1
1133
+ ; DEFAULT-NEXT: global_store_dwordx2 v0, v[1:2 ], s[12:13]
1138
1134
; DEFAULT-NEXT: .LBB9_4: ; %bb.3
1139
1135
; DEFAULT-NEXT: s_or_b64 exec, exec, s[2:3]
1140
1136
; DEFAULT-NEXT: v_mov_b32_e32 v0, 0
1141
1137
; DEFAULT-NEXT: s_waitcnt vmcnt(0)
1142
- ; DEFAULT-NEXT: global_store_dwordx2 v0, v[1:2 ], s[14:15]
1138
+ ; DEFAULT-NEXT: global_store_dwordx2 v0, v[3:4 ], s[14:15]
1143
1139
; DEFAULT-NEXT: s_endpgm
1144
1140
entry:
1145
1141
%idx = call i32 @llvm.amdgcn.workitem.id.x ()
@@ -1279,14 +1275,14 @@ define amdgpu_kernel void @v8i8_phi_const(ptr addrspace(1) %src1, ptr addrspace(
1279
1275
; FEATURE-NEXT: v_cmp_gt_u32_e32 vcc, 7, v0
1280
1276
; FEATURE-NEXT: s_andn2_b64 s[0:1], s[0:1], exec
1281
1277
; FEATURE-NEXT: s_and_b64 s[4:5], vcc, exec
1282
- ; FEATURE-NEXT: v_mov_b32_e32 v1, 1
1283
- ; FEATURE-NEXT: v_mov_b32_e32 v10, 2
1284
- ; FEATURE-NEXT: v_mov_b32_e32 v9, 3
1285
- ; FEATURE-NEXT: v_mov_b32_e32 v8, 4
1286
- ; FEATURE-NEXT: v_mov_b32_e32 v2, 5
1287
- ; FEATURE-NEXT: v_mov_b32_e32 v7, 6
1288
- ; FEATURE-NEXT: v_mov_b32_e32 v6, 7
1289
1278
; FEATURE-NEXT: v_mov_b32_e32 v5, 8
1279
+ ; FEATURE-NEXT: v_mov_b32_e32 v6, 7
1280
+ ; FEATURE-NEXT: v_mov_b32_e32 v7, 6
1281
+ ; FEATURE-NEXT: v_mov_b32_e32 v2, 5
1282
+ ; FEATURE-NEXT: v_mov_b32_e32 v8, 4
1283
+ ; FEATURE-NEXT: v_mov_b32_e32 v9, 3
1284
+ ; FEATURE-NEXT: v_mov_b32_e32 v10, 2
1285
+ ; FEATURE-NEXT: v_mov_b32_e32 v1, 1
1290
1286
; FEATURE-NEXT: s_or_b64 s[0:1], s[0:1], s[4:5]
1291
1287
; FEATURE-NEXT: s_waitcnt vmcnt(0)
1292
1288
; FEATURE-NEXT: v_lshrrev_b32_e32 v16, 24, v4
@@ -1365,14 +1361,14 @@ define amdgpu_kernel void @v8i8_phi_const(ptr addrspace(1) %src1, ptr addrspace(
1365
1361
; DEFAULT-NEXT: v_cmp_gt_u32_e32 vcc, 7, v0
1366
1362
; DEFAULT-NEXT: s_andn2_b64 s[0:1], s[0:1], exec
1367
1363
; DEFAULT-NEXT: s_and_b64 s[4:5], vcc, exec
1368
- ; DEFAULT-NEXT: v_mov_b32_e32 v1, 1
1369
- ; DEFAULT-NEXT: v_mov_b32_e32 v10, 2
1370
- ; DEFAULT-NEXT: v_mov_b32_e32 v9, 3
1371
- ; DEFAULT-NEXT: v_mov_b32_e32 v8, 4
1372
- ; DEFAULT-NEXT: v_mov_b32_e32 v2, 5
1373
- ; DEFAULT-NEXT: v_mov_b32_e32 v7, 6
1374
- ; DEFAULT-NEXT: v_mov_b32_e32 v6, 7
1375
1364
; DEFAULT-NEXT: v_mov_b32_e32 v5, 8
1365
+ ; DEFAULT-NEXT: v_mov_b32_e32 v6, 7
1366
+ ; DEFAULT-NEXT: v_mov_b32_e32 v7, 6
1367
+ ; DEFAULT-NEXT: v_mov_b32_e32 v2, 5
1368
+ ; DEFAULT-NEXT: v_mov_b32_e32 v8, 4
1369
+ ; DEFAULT-NEXT: v_mov_b32_e32 v9, 3
1370
+ ; DEFAULT-NEXT: v_mov_b32_e32 v10, 2
1371
+ ; DEFAULT-NEXT: v_mov_b32_e32 v1, 1
1376
1372
; DEFAULT-NEXT: s_or_b64 s[0:1], s[0:1], s[4:5]
1377
1373
; DEFAULT-NEXT: s_waitcnt vmcnt(0)
1378
1374
; DEFAULT-NEXT: v_lshrrev_b32_e32 v16, 24, v4
0 commit comments