322 lines · plain
1# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 52# RUN: llc %s -mtriple=amdgcn -mcpu=gfx1030 -run-pass=si-peephole-sdwa -o - | FileCheck %s3 4---5name: cndmask_b32 # can be directly converted to SDWA without a copy to VCC6tracksRegLiveness: true7body: |8 bb.0:9 liveins: $vgpr0, $vgpr1, $vcc10 11 ; CHECK-LABEL: name: cndmask_b3212 ; CHECK: liveins: $vgpr0, $vgpr1, $vcc13 ; CHECK-NEXT: {{ $}}14 ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr115 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr016 ; CHECK-NEXT: [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, [[COPY]], implicit $exec17 ; CHECK-NEXT: [[V_LSHRREV_B32_e64_1:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, [[COPY1]], implicit $exec18 ; CHECK-NEXT: [[V_CNDMASK_B32_sdwa:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_sdwa 0, [[COPY]], 0, [[COPY1]], 0, 6, 0, 5, 5, implicit $vcc_lo, implicit $exec19 ; CHECK-NEXT: $vgpr0 = COPY [[V_CNDMASK_B32_sdwa]]20 ; CHECK-NEXT: SI_RETURN implicit $vgpr021 %0:vgpr_32 = COPY $vgpr122 %1:vgpr_32 = COPY $vgpr023 %2:vgpr_32 = V_LSHRREV_B32_e64 16, %0, implicit $exec24 %3:vgpr_32 = V_LSHRREV_B32_e64 16, %1, implicit $exec25 %4:vgpr_32 = V_CNDMASK_B32_e32 killed %2, killed %3, implicit $exec, implicit $vcc26 $vgpr0 = COPY %427 SI_RETURN implicit $vgpr028...29 30# For SDWA conversion of V_CNDMASK, the carry-in operand must be31# available in VCC_LO. This is achieved by introducing a COPY32# instruction. Comparison instructions could be changed to VOP2 form33# intead, but we prefer to use a COPY.34 35---36name: carry-compare37tracksRegLiveness: true38body: |39 bb.0:40 liveins: $vgpr041 42 ; CHECK-LABEL: name: carry-compare43 ; CHECK: liveins: $vgpr044 ; CHECK-NEXT: {{ $}}45 ; CHECK-NEXT: [[DEF:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF46 ; CHECK-NEXT: [[DEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF47 ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 undef [[DEF]], 1, implicit $exec48 ; CHECK-NEXT: [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, undef [[DEF1]], implicit $exec49 ; CHECK-NEXT: $vcc_lo = COPY killed [[V_CMP_EQ_U32_e64_]]50 ; CHECK-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec51 ; CHECK-NEXT: [[V_CNDMASK_B32_sdwa:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_sdwa 0, [[V_MOV_B32_e32_]], 0, undef [[DEF1]], 0, 6, 0, 6, 5, implicit $vcc_lo, implicit $exec52 ; CHECK-NEXT: $vgpr0 = COPY [[V_CNDMASK_B32_sdwa]]53 ; CHECK-NEXT: SI_RETURN implicit $vgpr054 %0:sreg_32_xm0_xexec = IMPLICIT_DEF55 %1:vgpr_32 = IMPLICIT_DEF56 %2:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 undef %0, 1, implicit $exec57 %3:vgpr_32 = V_LSHRREV_B32_e64 16, undef %1, implicit $exec58 %4:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, %3, killed %2, implicit $exec59 $vgpr0 = COPY %460 SI_RETURN implicit $vgpr061...62 63---64name: carry-compare-class65tracksRegLiveness: true66body: |67 bb.0:68 liveins: $vgpr069 70 ; CHECK-LABEL: name: carry-compare-class71 ; CHECK: liveins: $vgpr072 ; CHECK-NEXT: {{ $}}73 ; CHECK-NEXT: [[DEF:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF74 ; CHECK-NEXT: [[DEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF75 ; CHECK-NEXT: [[V_CMP_CLASS_F32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_CLASS_F32_e64 2, undef [[DEF]], 1, implicit $exec76 ; CHECK-NEXT: [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, undef [[DEF1]], implicit $exec77 ; CHECK-NEXT: $vcc_lo = COPY killed [[V_CMP_CLASS_F32_e64_]]78 ; CHECK-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec79 ; CHECK-NEXT: [[V_CNDMASK_B32_sdwa:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_sdwa 0, [[V_MOV_B32_e32_]], 0, undef [[DEF1]], 0, 6, 0, 6, 5, implicit $vcc_lo, implicit $exec80 ; CHECK-NEXT: $vgpr0 = COPY [[V_CNDMASK_B32_sdwa]]81 ; CHECK-NEXT: SI_RETURN implicit $vgpr082 %0:sreg_32_xm0_xexec = IMPLICIT_DEF83 %1:vgpr_32 = IMPLICIT_DEF84 %2:sreg_32_xm0_xexec = V_CMP_CLASS_F32_e64 2, undef %0, 1, implicit $exec85 %3:vgpr_32 = V_LSHRREV_B32_e64 16, undef %1, implicit $exec86 %4:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, %3, killed %2, implicit $exec87 $vgpr0 = COPY %488 SI_RETURN implicit $vgpr089...90 91...92---93name: carry-non-compare94tracksRegLiveness: true95body: |96 bb.0:97 liveins: $vgpr098 99 ; CHECK-LABEL: name: carry-non-compare100 ; CHECK: liveins: $vgpr0101 ; CHECK-NEXT: {{ $}}102 ; CHECK-NEXT: [[DEF:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF103 ; CHECK-NEXT: [[DEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF104 ; CHECK-NEXT: [[V_PK_MAX_F16_:%[0-9]+]]:vgpr_32 = V_PK_MAX_F16 8, [[DEF1]], 8, [[DEF1]], 0, 0, 0, 0, 0, implicit $mode, implicit $exec105 ; CHECK-NEXT: [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, [[V_PK_MAX_F16_]], implicit $exec106 ; CHECK-NEXT: $vcc_lo = COPY killed undef [[DEF]]107 ; CHECK-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec108 ; CHECK-NEXT: [[V_CNDMASK_B32_sdwa:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_sdwa 0, [[V_MOV_B32_e32_]], 0, [[V_PK_MAX_F16_]], 0, 6, 0, 6, 5, implicit $vcc_lo, implicit $exec109 ; CHECK-NEXT: $vgpr0 = COPY [[V_CNDMASK_B32_sdwa]]110 ; CHECK-NEXT: SI_RETURN implicit $vgpr0111 %0:sreg_32_xm0_xexec = IMPLICIT_DEF112 %1:vgpr_32 = IMPLICIT_DEF113 %2:vgpr_32 = V_PK_MAX_F16 8, %1, 8, %1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec114 %3:vgpr_32 = V_LSHRREV_B32_e64 16, %2, implicit $exec115 %4:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, %3, killed undef %0, implicit $exec116 $vgpr0 = COPY %4117 SI_RETURN implicit $vgpr0118...119 120...121---122name: carry-multiuse123tracksRegLiveness: true124body: |125 bb.0:126 liveins: $vgpr0, $vgpr1127 128 ; CHECK-LABEL: name: carry-multiuse129 ; CHECK: liveins: $vgpr0, $vgpr1130 ; CHECK-NEXT: {{ $}}131 ; CHECK-NEXT: [[DEF:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF132 ; CHECK-NEXT: [[DEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF133 ; CHECK-NEXT: [[V_PK_MAX_F16_:%[0-9]+]]:vgpr_32 = V_PK_MAX_F16 8, [[DEF1]], 8, [[DEF1]], 0, 0, 0, 0, 0, implicit $mode, implicit $exec134 ; CHECK-NEXT: [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, [[V_PK_MAX_F16_]], implicit $exec135 ; CHECK-NEXT: $vcc_lo = COPY undef [[DEF]]136 ; CHECK-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec137 ; CHECK-NEXT: [[V_CNDMASK_B32_sdwa:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_sdwa 0, [[V_MOV_B32_e32_]], 0, [[V_PK_MAX_F16_]], 0, 6, 0, 6, 5, implicit $vcc_lo, implicit $exec138 ; CHECK-NEXT: $vgpr0 = COPY [[V_CNDMASK_B32_sdwa]]139 ; CHECK-NEXT: $vgpr1 = COPY [[DEF]]140 ; CHECK-NEXT: SI_RETURN implicit $vgpr0141 %0:sreg_32_xm0_xexec = IMPLICIT_DEF142 %1:vgpr_32 = IMPLICIT_DEF143 %2:vgpr_32 = V_PK_MAX_F16 8, %1, 8, %1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec144 %3:vgpr_32 = V_LSHRREV_B32_e64 16, %2, implicit $exec145 %4:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, %3, undef %0, implicit $exec146 $vgpr0 = COPY %4147 $vgpr1 = COPY %0148 SI_RETURN implicit $vgpr0149...150 151...152---153name: live-vcc # cannot convert because of live VCC154tracksRegLiveness: true155body: |156 bb.0:157 liveins: $vgpr0158 159 ; CHECK-LABEL: name: live-vcc160 ; CHECK: liveins: $vgpr0161 ; CHECK-NEXT: {{ $}}162 ; CHECK-NEXT: [[DEF:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF163 ; CHECK-NEXT: [[DEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF164 ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 [[DEF]], 1, implicit $exec165 ; CHECK-NEXT: [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, [[DEF1]], implicit $exec166 ; CHECK-NEXT: V_CMP_EQ_U32_e32 1, undef [[DEF1]], implicit-def $vcc_lo, implicit $exec167 ; CHECK-NEXT: [[V_CNDMASK_B32_e64_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, [[V_LSHRREV_B32_e64_]], [[V_CMP_EQ_U32_e64_]], implicit $exec168 ; CHECK-NEXT: [[V_CNDMASK_B32_e32_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e32 killed [[V_LSHRREV_B32_e64_]], killed [[V_LSHRREV_B32_e64_]], implicit $vcc_lo, implicit $exec169 ; CHECK-NEXT: $vgpr0 = COPY [[V_CNDMASK_B32_e64_]]170 ; CHECK-NEXT: SI_RETURN implicit $vgpr0171 %0:sreg_32_xm0_xexec = IMPLICIT_DEF172 %2:vgpr_32 = IMPLICIT_DEF173 %3:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 %0, 1, implicit $exec174 %5:vgpr_32 = V_LSHRREV_B32_e64 16, %2, implicit $exec175 V_CMP_EQ_U32_e32 1, undef %2, implicit-def $vcc, implicit $exec176 %6:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, %5, %3, implicit $exec177 %7:vgpr_32 = V_CNDMASK_B32_e32 killed %5, killed %5, implicit $vcc, implicit $exec178 $vgpr0 = COPY %6179 SI_RETURN implicit $vgpr0180...181 182...183---184name: cannot-shrink-with-source-mods185tracksRegLiveness: true186body: |187 bb.0:188 liveins: $vgpr0189 190 ; CHECK-LABEL: name: cannot-shrink-with-source-mods191 ; CHECK: liveins: $vgpr0192 ; CHECK-NEXT: {{ $}}193 ; CHECK-NEXT: [[DEF:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF194 ; CHECK-NEXT: [[DEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF195 ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 [[DEF]], 0, implicit $exec196 ; CHECK-NEXT: [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, [[DEF1]], implicit $exec197 ; CHECK-NEXT: [[V_CNDMASK_B32_e64_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 1, 0, 0, [[V_LSHRREV_B32_e64_]], killed [[V_CMP_EQ_U32_e64_]], implicit $exec198 ; CHECK-NEXT: $vgpr0 = COPY [[V_CNDMASK_B32_e64_]]199 ; CHECK-NEXT: SI_RETURN implicit $vgpr0200 %0:sreg_32_xm0_xexec = IMPLICIT_DEF201 %1:vgpr_32 = IMPLICIT_DEF202 %2:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 %0, 0, implicit $exec203 %3:vgpr_32 = V_LSHRREV_B32_e64 16, %1, implicit $exec204 %4:vgpr_32 = V_CNDMASK_B32_e64 1, 0, 0, %3, killed %2, implicit $exec205 $vgpr0 = COPY %4206 SI_RETURN implicit $vgpr0207...208 209...210---211name: missing-carry-def212tracksRegLiveness: true213registers:214 - { id: 0, class: sreg_32_xm0_xexec }215body: |216 bb.0:217 liveins: $vgpr0218 219 ; CHECK-LABEL: name: missing-carry-def220 ; CHECK: liveins: $vgpr0221 ; CHECK-NEXT: {{ $}}222 ; CHECK-NEXT: [[DEF:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF223 ; CHECK-NEXT: [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, [[DEF]], implicit $exec224 ; CHECK-NEXT: [[V_CNDMASK_B32_e64_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, [[V_LSHRREV_B32_e64_]], undef %0:sreg_32_xm0_xexec, implicit $exec225 ; CHECK-NEXT: $vgpr0 = COPY [[V_CNDMASK_B32_e64_]]226 ; CHECK-NEXT: SI_RETURN implicit $vgpr0227 %1:sreg_32_xm0_xexec = IMPLICIT_DEF228 %2:vgpr_32 = V_LSHRREV_B32_e64 16, %1:sreg_32_xm0_xexec, implicit $exec229 %3:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, %2, undef %0:sreg_32_xm0_xexec, implicit $exec230 $vgpr0 = COPY %3231 SI_RETURN implicit $vgpr0232...233 234---235name: cndmask-not-converted236tracksRegLiveness: true237body: |238 ; CHECK-LABEL: name: cndmask-not-converted239 ; CHECK: bb.0:240 ; CHECK-NEXT: successors: %bb.1(0x40000000), %bb.2(0x40000000)241 ; CHECK-NEXT: liveins: $vgpr0, $sgpr8_sgpr9242 ; CHECK-NEXT: {{ $}}243 ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_64 = COPY $sgpr8_sgpr9244 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32(s32) = COPY $vgpr0245 ; CHECK-NEXT: [[S_LOAD_DWORDX2_IMM:%[0-9]+]]:sreg_64_xexec = S_LOAD_DWORDX2_IMM [[COPY]], 0, 0246 ; CHECK-NEXT: S_BITCMP1_B32 [[S_LOAD_DWORDX2_IMM]].sub1, 0, implicit-def $scc247 ; CHECK-NEXT: [[S_CSELECT_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_CSELECT_B32 -1, 0, implicit $scc248 ; CHECK-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec249 ; CHECK-NEXT: [[S_AND_B32_:%[0-9]+]]:sreg_32 = S_AND_B32 $exec_lo, [[S_CSELECT_B32_]], implicit-def dead $scc250 ; CHECK-NEXT: $vcc_lo = COPY [[S_AND_B32_]]251 ; CHECK-NEXT: S_CBRANCH_VCCNZ %bb.2, implicit $vcc_lo252 ; CHECK-NEXT: S_BRANCH %bb.1253 ; CHECK-NEXT: {{ $}}254 ; CHECK-NEXT: bb.1:255 ; CHECK-NEXT: successors: %bb.2(0x80000000)256 ; CHECK-NEXT: {{ $}}257 ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_64 = COPY [[S_LOAD_DWORDX2_IMM]]258 ; CHECK-NEXT: [[V_MUL_U32_U24_e64_:%[0-9]+]]:vgpr_32 = V_MUL_U32_U24_e64 [[COPY1]](s32), 5, 0, implicit $exec259 ; CHECK-NEXT: [[V_MOV_B32_e32_1:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec260 ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[V_MUL_U32_U24_e64_]], %subreg.sub0, killed [[V_MOV_B32_e32_1]], %subreg.sub1261 ; CHECK-NEXT: [[GLOBAL_LOAD_USHORT:%[0-9]+]]:vgpr_32 = GLOBAL_LOAD_USHORT [[REG_SEQUENCE]], 3, 0, implicit $exec262 ; CHECK-NEXT: [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[GLOBAL_LOAD_USHORT]], 255, implicit $exec263 ; CHECK-NEXT: [[V_MOV_B32_e32_2:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 65535, implicit $exec264 ; CHECK-NEXT: [[V_AND_B32_sdwa:%[0-9]+]]:vgpr_32 = V_AND_B32_sdwa 0, [[V_MOV_B32_e32_2]], 0, [[GLOBAL_LOAD_USHORT]], 0, 6, 0, 6, 0, implicit $exec265 ; CHECK-NEXT: S_CMP_EQ_U32 [[COPY2]].sub0, 0, implicit-def $scc266 ; CHECK-NEXT: [[S_CSELECT_B32_1:%[0-9]+]]:sreg_32_xm0_xexec = S_CSELECT_B32 -1, 0, implicit $scc267 ; CHECK-NEXT: $vcc_lo = COPY [[S_CSELECT_B32_1]]268 ; CHECK-NEXT: [[V_CNDMASK_B32_e32_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e32 0, killed [[V_AND_B32_sdwa]], implicit $vcc_lo, implicit $exec269 ; CHECK-NEXT: [[V_MOV_B32_e32_3:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 24, implicit $exec270 ; CHECK-NEXT: [[V_LSHRREV_B32_sdwa:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_sdwa 0, [[V_MOV_B32_e32_3]], 0, [[V_CNDMASK_B32_e32_]], 0, 1, 0, 6, 6, implicit $exec271 ; CHECK-NEXT: [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 16, [[V_CNDMASK_B32_e32_]], implicit $exec272 ; CHECK-NEXT: [[V_MOV_B32_e32_4:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 255, implicit $exec273 ; CHECK-NEXT: [[V_AND_B32_sdwa1:%[0-9]+]]:vgpr_32 = V_AND_B32_sdwa 0, [[V_CNDMASK_B32_e32_]], 0, [[V_MOV_B32_e32_4]], 0, 6, 0, 5, 6, implicit $exec274 ; CHECK-NEXT: [[V_OR_B32_sdwa:%[0-9]+]]:vgpr_32 = V_OR_B32_sdwa 0, [[V_AND_B32_sdwa1]], 0, [[V_LSHRREV_B32_sdwa]], 0, 5, 0, 6, 6, implicit $exec275 ; CHECK-NEXT: {{ $}}276 ; CHECK-NEXT: bb.2:277 ; CHECK-NEXT: [[PHI:%[0-9]+]]:vgpr_32 = PHI [[V_MOV_B32_e32_]], %bb.0, [[V_OR_B32_sdwa]], %bb.1278 ; CHECK-NEXT: [[V_MOV_B:%[0-9]+]]:vreg_64 = V_MOV_B64_PSEUDO 0, implicit $exec279 ; CHECK-NEXT: GLOBAL_STORE_BYTE killed [[V_MOV_B]], [[PHI]], 0, 0, implicit $exec280 ; CHECK-NEXT: S_ENDPGM 0281 bb.0:282 successors: %bb.1(0x40000000), %bb.2(0x40000000)283 liveins: $vgpr0, $sgpr8_sgpr9284 285 %0:sgpr_64 = COPY $sgpr8_sgpr9286 %1:vgpr_32 = COPY $vgpr0287 %2:sreg_64_xexec = S_LOAD_DWORDX2_IMM %0, 0, 0288 S_BITCMP1_B32 %2.sub1, 0, implicit-def $scc289 %3:sreg_32_xm0_xexec = S_CSELECT_B32 -1, 0, implicit $scc290 %4:vgpr_32 = V_MOV_B32_e32 0, implicit $exec291 %5:sreg_32 = S_AND_B32 $exec_lo, %3, implicit-def dead $scc292 $vcc_lo = COPY %5293 S_CBRANCH_VCCNZ %bb.2, implicit $vcc294 S_BRANCH %bb.1295 296 bb.1:297 successors: %bb.2(0x80000000)298 299 %6:sreg_64 = COPY %2300 %7:vgpr_32 = V_MUL_U32_U24_e64 %1(s32), 5, 0, implicit $exec301 %8:vgpr_32 = V_MOV_B32_e32 0, implicit $exec302 %9:vreg_64 = REG_SEQUENCE %7, %subreg.sub0, killed %8, %subreg.sub1303 %10:vgpr_32 = GLOBAL_LOAD_USHORT %9, 3, 0, implicit $exec304 %11:vgpr_32 = V_AND_B32_e64 %10, 255, implicit $exec305 %12:vgpr_32 = V_AND_B32_e64 65535, killed %11, implicit $exec306 S_CMP_EQ_U32 %6.sub0, 0, implicit-def $scc307 %13:sreg_32_xm0_xexec = S_CSELECT_B32 -1, 0, implicit $scc308 %14:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, killed %12, %13, implicit $exec309 %15:vgpr_32 = V_LSHRREV_B32_e64 24, %14, implicit $exec310 %16:vgpr_32 = V_LSHLREV_B16_e64 8, %15, implicit $exec311 %17:vgpr_32 = V_LSHRREV_B32_e64 16, %14, implicit $exec312 %18:vgpr_32 = V_AND_B32_e64 %17, 255, implicit $exec313 %19:vgpr_32 = V_OR_B32_e64 killed %18, killed %16, implicit $exec314 %20:vgpr_32 = V_LSHLREV_B32_e64 16, killed %19, implicit $exec315 316 bb.2:317 %21:vgpr_32 = PHI %4, %bb.0, %20, %bb.1318 %22:vreg_64 = V_MOV_B64_PSEUDO 0, implicit $exec319 GLOBAL_STORE_BYTE killed %22, %21, 0, 0, implicit $exec320 S_ENDPGM 0321...322