273 lines · plain
1# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py2# RUN: llc -mtriple=amdgcn -mcpu=fiji -verify-machineinstrs -run-pass=si-lower-control-flow -o - %s | FileCheck %s3# RUN: llc -mtriple=amdgcn -mcpu=fiji -passes=si-lower-control-flow -o - %s | FileCheck %s4 5# Test si-lower-control-flow insertion points when other terminator6# instructions are present besides the control flow pseudo and a7# branch.8 9 10# There's another terminator instruction between SI_IF and11# S_BRANCH. The S_CBRANCH_EXECZ should be inserted immediately before12# S_BRANCH.13---14name: other_terminator_sbranch_after_si_if15tracksRegLiveness: true16body: |17 ; CHECK-LABEL: name: other_terminator_sbranch_after_si_if18 ; CHECK: bb.0:19 ; CHECK-NEXT: successors: %bb.2(0x40000000), %bb.1(0x40000000)20 ; CHECK-NEXT: liveins: $vgpr0, $vgpr1, $sgpr4_sgpr521 ; CHECK-NEXT: {{ $}}22 ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY killed $vgpr023 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_64_xexec = COPY $sgpr4_sgpr524 ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 0, [[COPY]], implicit $exec25 ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_64 = COPY $exec, implicit-def $exec26 ; CHECK-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64 = S_AND_B64 [[COPY2]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc27 ; CHECK-NEXT: [[S_XOR_B64_:%[0-9]+]]:sreg_64_xexec = S_XOR_B64 [[S_AND_B64_]], [[COPY2]], implicit-def dead $scc28 ; CHECK-NEXT: $exec = S_MOV_B64_term killed [[S_AND_B64_]]29 ; CHECK-NEXT: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term killed [[COPY1]], implicit $exec30 ; CHECK-NEXT: S_CBRANCH_EXECZ %bb.1, implicit $exec31 ; CHECK-NEXT: S_BRANCH %bb.232 ; CHECK-NEXT: {{ $}}33 ; CHECK-NEXT: bb.1:34 ; CHECK-NEXT: successors: %bb.2(0x80000000)35 ; CHECK-NEXT: {{ $}}36 ; CHECK-NEXT: S_BRANCH %bb.237 ; CHECK-NEXT: {{ $}}38 ; CHECK-NEXT: bb.2:39 ; CHECK-NEXT: S_ENDPGM 0, implicit [[S_MOV_B64_term]]40 bb.0:41 successors: %bb.2, %bb.142 liveins: $vgpr0, $vgpr1, $sgpr4_sgpr543 44 %0:vgpr_32 = COPY killed $vgpr045 %1:sreg_64_xexec = COPY $sgpr4_sgpr546 %2:sreg_64_xexec = V_CMP_EQ_U32_e64 0, %0, implicit $exec47 %3:sreg_64_xexec = SI_IF %2, %bb.1, implicit-def $exec, implicit-def dead $scc, implicit $exec48 %4:sreg_64_xexec = S_MOV_B64_term killed %1, implicit $exec49 S_BRANCH %bb.250 51 bb.1:52 S_BRANCH %bb.253 54 bb.2:55 S_ENDPGM 0, implicit %456 57...58 59# S_CBRANCH_EXECZ should be inserted after the other terminator60---61name: other_terminator_fallthrough_after_si_if62tracksRegLiveness: true63body: |64 ; CHECK-LABEL: name: other_terminator_fallthrough_after_si_if65 ; CHECK: bb.0:66 ; CHECK-NEXT: successors: %bb.2(0x40000000), %bb.1(0x40000000)67 ; CHECK-NEXT: liveins: $vgpr0, $vgpr1, $sgpr4_sgpr568 ; CHECK-NEXT: {{ $}}69 ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY killed $vgpr070 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_64_xexec = COPY $sgpr4_sgpr571 ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 0, [[COPY]], implicit $exec72 ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_64 = COPY $exec, implicit-def $exec73 ; CHECK-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64 = S_AND_B64 [[COPY2]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc74 ; CHECK-NEXT: [[S_XOR_B64_:%[0-9]+]]:sreg_64_xexec = S_XOR_B64 [[S_AND_B64_]], [[COPY2]], implicit-def dead $scc75 ; CHECK-NEXT: $exec = S_MOV_B64_term killed [[S_AND_B64_]]76 ; CHECK-NEXT: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term killed [[COPY1]], implicit $exec77 ; CHECK-NEXT: S_CBRANCH_EXECZ %bb.2, implicit $exec78 ; CHECK-NEXT: {{ $}}79 ; CHECK-NEXT: bb.1:80 ; CHECK-NEXT: successors: %bb.2(0x80000000)81 ; CHECK-NEXT: {{ $}}82 ; CHECK-NEXT: S_BRANCH %bb.283 ; CHECK-NEXT: {{ $}}84 ; CHECK-NEXT: bb.2:85 ; CHECK-NEXT: S_ENDPGM 0, implicit [[S_MOV_B64_term]]86 bb.0:87 successors: %bb.2, %bb.188 liveins: $vgpr0, $vgpr1, $sgpr4_sgpr589 90 %0:vgpr_32 = COPY killed $vgpr091 %1:sreg_64_xexec = COPY $sgpr4_sgpr592 %2:sreg_64_xexec = V_CMP_EQ_U32_e64 0, %0, implicit $exec93 %3:sreg_64_xexec = SI_IF %2, %bb.2, implicit-def $exec, implicit-def dead $scc, implicit $exec94 %4:sreg_64_xexec = S_MOV_B64_term killed %1, implicit $exec95 96 bb.1:97 S_BRANCH %bb.298 99 bb.2:100 S_ENDPGM 0, implicit %4101 102...103 104---105name: other_terminator_sbranch_after_si_else106tracksRegLiveness: true107body: |108 ; CHECK-LABEL: name: other_terminator_sbranch_after_si_else109 ; CHECK: bb.0:110 ; CHECK-NEXT: successors: %bb.2(0x40000000), %bb.1(0x40000000)111 ; CHECK-NEXT: liveins: $vgpr0, $vgpr1, $sgpr4_sgpr5112 ; CHECK-NEXT: {{ $}}113 ; CHECK-NEXT: [[S_OR_SAVEEXEC_B64_:%[0-9]+]]:sreg_64 = S_OR_SAVEEXEC_B64 %2, implicit-def $exec, implicit-def $scc, implicit $exec114 ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY killed $vgpr0115 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_64_xexec = COPY $sgpr4_sgpr5116 ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 0, [[COPY]], implicit $exec117 ; CHECK-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 $exec, [[S_OR_SAVEEXEC_B64_]], implicit-def $scc118 ; CHECK-NEXT: $exec = S_XOR_B64_term $exec, [[S_AND_B64_]], implicit-def $scc119 ; CHECK-NEXT: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term killed [[COPY1]], implicit $exec120 ; CHECK-NEXT: S_CBRANCH_EXECZ %bb.1, implicit $exec121 ; CHECK-NEXT: S_BRANCH %bb.2122 ; CHECK-NEXT: {{ $}}123 ; CHECK-NEXT: bb.1:124 ; CHECK-NEXT: successors: %bb.2(0x80000000)125 ; CHECK-NEXT: {{ $}}126 ; CHECK-NEXT: S_BRANCH %bb.2127 ; CHECK-NEXT: {{ $}}128 ; CHECK-NEXT: bb.2:129 ; CHECK-NEXT: S_ENDPGM 0, implicit [[S_MOV_B64_term]]130 bb.0:131 successors: %bb.2, %bb.1132 liveins: $vgpr0, $vgpr1, $sgpr4_sgpr5133 134 %0:vgpr_32 = COPY killed $vgpr0135 %1:sreg_64_xexec = COPY $sgpr4_sgpr5136 %2:sreg_64_xexec = V_CMP_EQ_U32_e64 0, %0, implicit $exec137 %3:sreg_64_xexec = SI_ELSE %2, %bb.1, implicit-def $exec, implicit-def dead $scc, implicit $exec138 %4:sreg_64_xexec = S_MOV_B64_term killed %1, implicit $exec139 S_BRANCH %bb.2140 141 bb.1:142 S_BRANCH %bb.2143 144 bb.2:145 S_ENDPGM 0, implicit %4146 147...148 149---150name: other_terminator_sbranch_after_si_loop151tracksRegLiveness: true152body: |153 ; CHECK-LABEL: name: other_terminator_sbranch_after_si_loop154 ; CHECK: bb.0:155 ; CHECK-NEXT: successors: %bb.2(0x40000000), %bb.1(0x40000000)156 ; CHECK-NEXT: liveins: $vgpr0, $vgpr1, $sgpr4_sgpr5157 ; CHECK-NEXT: {{ $}}158 ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY killed $vgpr0159 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sreg_64_xexec = COPY $sgpr4_sgpr5160 ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 0, [[COPY]], implicit $exec161 ; CHECK-NEXT: $exec = S_ANDN2_B64_term $exec, [[V_CMP_EQ_U32_e64_]], implicit-def $scc162 ; CHECK-NEXT: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term killed [[COPY1]], implicit $exec163 ; CHECK-NEXT: S_CBRANCH_EXECNZ %bb.1, implicit $exec164 ; CHECK-NEXT: S_BRANCH %bb.2165 ; CHECK-NEXT: {{ $}}166 ; CHECK-NEXT: bb.1:167 ; CHECK-NEXT: successors: %bb.2(0x80000000)168 ; CHECK-NEXT: {{ $}}169 ; CHECK-NEXT: S_BRANCH %bb.2170 ; CHECK-NEXT: {{ $}}171 ; CHECK-NEXT: bb.2:172 ; CHECK-NEXT: S_ENDPGM 0, implicit [[S_MOV_B64_term]]173 bb.0:174 successors: %bb.2, %bb.1175 liveins: $vgpr0, $vgpr1, $sgpr4_sgpr5176 177 %0:vgpr_32 = COPY killed $vgpr0178 %1:sreg_64_xexec = COPY $sgpr4_sgpr5179 %2:sreg_64_xexec = V_CMP_EQ_U32_e64 0, %0, implicit $exec180 SI_LOOP %2, %bb.1, implicit-def $exec, implicit-def dead $scc, implicit $exec181 %4:sreg_64_xexec = S_MOV_B64_term killed %1, implicit $exec182 S_BRANCH %bb.2183 184 bb.1:185 S_BRANCH %bb.2186 187 bb.2:188 S_ENDPGM 0, implicit %4189 190...191 192# The save exec result register of SI_IF is used by other terminators193# inserted to behave as a lowered phi. The output register of SI_IF194# was ignored, and the def was removed, so the S_MOV_B64_term uses195# would fail the verifier.196 197---198name: si_if_use199alignment: 1200legalized: true201regBankSelected: true202selected: true203tracksRegLiveness: true204body: |205 ; CHECK-LABEL: name: si_if_use206 ; CHECK: bb.0:207 ; CHECK-NEXT: successors: %bb.1(0x40000000), %bb.2(0x40000000)208 ; CHECK-NEXT: liveins: $vgpr0, $vgpr1, $sgpr30_sgpr31209 ; CHECK-NEXT: {{ $}}210 ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY killed $vgpr0211 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY killed $vgpr1212 ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 killed [[COPY]], killed [[COPY1]], implicit $exec213 ; CHECK-NEXT: [[COPY2:%[0-9]+]]:sreg_64 = COPY $exec, implicit-def $exec214 ; CHECK-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64 = S_AND_B64 [[COPY2]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc215 ; CHECK-NEXT: [[S_XOR_B64_:%[0-9]+]]:sreg_64_xexec = S_XOR_B64 [[S_AND_B64_]], [[COPY2]], implicit-def dead $scc216 ; CHECK-NEXT: $exec = S_MOV_B64_term killed [[S_AND_B64_]]217 ; CHECK-NEXT: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term [[S_XOR_B64_]], implicit $exec218 ; CHECK-NEXT: [[S_MOV_B64_term1:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term [[S_XOR_B64_]], implicit $exec219 ; CHECK-NEXT: S_CBRANCH_EXECZ %bb.1, implicit $exec220 ; CHECK-NEXT: S_BRANCH %bb.2221 ; CHECK-NEXT: {{ $}}222 ; CHECK-NEXT: bb.1:223 ; CHECK-NEXT: successors: %bb.2(0x80000000)224 ; CHECK-NEXT: {{ $}}225 ; CHECK-NEXT: [[COPY3:%[0-9]+]]:sreg_64_xexec = COPY [[S_MOV_B64_term1]]226 ; CHECK-NEXT: dead [[GLOBAL_LOAD_DWORD:%[0-9]+]]:vgpr_32 = GLOBAL_LOAD_DWORD undef %8:vreg_64, 0, 0, implicit $exec :: (volatile load (s32), addrspace 1)227 ; CHECK-NEXT: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = COPY [[COPY3]]228 ; CHECK-NEXT: {{ $}}229 ; CHECK-NEXT: bb.2:230 ; CHECK-NEXT: successors: %bb.3(0x80000000)231 ; CHECK-NEXT: {{ $}}232 ; CHECK-NEXT: [[COPY4:%[0-9]+]]:sreg_64_xexec = COPY [[S_MOV_B64_term]]233 ; CHECK-NEXT: $exec = S_OR_B64_term $exec, killed [[COPY4]], implicit-def $scc234 ; CHECK-NEXT: {{ $}}235 ; CHECK-NEXT: bb.3:236 ; CHECK-NEXT: successors: %bb.1(0x40000000), %bb.2(0x40000000)237 ; CHECK-NEXT: {{ $}}238 ; CHECK-NEXT: S_SLEEP 1239 ; CHECK-NEXT: [[COPY5:%[0-9]+]]:sreg_64 = COPY $exec, implicit-def $exec240 ; CHECK-NEXT: [[S_AND_B64_1:%[0-9]+]]:sreg_64 = S_AND_B64 [[COPY5]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc241 ; CHECK-NEXT: [[S_XOR_B64_1:%[0-9]+]]:sreg_64_xexec = S_XOR_B64 [[S_AND_B64_1]], [[COPY5]], implicit-def dead $scc242 ; CHECK-NEXT: $exec = S_MOV_B64_term killed [[S_AND_B64_1]]243 ; CHECK-NEXT: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term [[S_XOR_B64_1]], implicit $exec244 ; CHECK-NEXT: [[S_MOV_B64_term1:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term [[S_XOR_B64_1]], implicit $exec245 ; CHECK-NEXT: S_CBRANCH_EXECZ %bb.1, implicit $exec246 ; CHECK-NEXT: S_BRANCH %bb.2247 bb.0:248 liveins: $vgpr0, $vgpr1, $sgpr30_sgpr31249 250 %0:vgpr_32 = COPY killed $vgpr0251 %1:vgpr_32 = COPY killed $vgpr1252 %3:sreg_64_xexec = V_CMP_EQ_U32_e64 killed %0, killed %1, implicit $exec253 %10:sreg_64_xexec = SI_IF %3, %bb.1, implicit-def $exec, implicit-def dead $scc, implicit $exec254 %14:sreg_64_xexec = S_MOV_B64_term %10, implicit $exec255 %13:sreg_64_xexec = S_MOV_B64_term %10, implicit $exec256 S_BRANCH %bb.2257 258 bb.1:259 %11:sreg_64_xexec = COPY %13260 dead %6:vgpr_32 = GLOBAL_LOAD_DWORD undef %8:vreg_64, 0, 0, implicit $exec :: (volatile load (s32), addrspace 1)261 %14:sreg_64_xexec = COPY %11262 263 bb.2:264 %12:sreg_64_xexec = COPY %14265 SI_END_CF killed %12, implicit-def $exec, implicit-def dead $scc, implicit $exec266 S_SLEEP 1267 %9:sreg_64_xexec = SI_IF %3, %bb.1, implicit-def $exec, implicit-def dead $scc, implicit $exec268 %14:sreg_64_xexec = S_MOV_B64_term %9, implicit $exec269 %13:sreg_64_xexec = S_MOV_B64_term %9, implicit $exec270 S_BRANCH %bb.2271 272...273