brintos

brintos / llvm-project-archived public Read only

0
0
Text · 8.0 KiB · c4d0583 Raw
214 lines · plain
1# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py2# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 -verify-machineinstrs -run-pass=si-optimize-exec-masking-pre-ra -o - %s | FileCheck %s3 4 5# Cannot fold this without moving the def of %7 after the and.6---7name:            no_fold_andn2_select_condition_live_out_phi8tracksRegLiveness: true9body:             |10  ; CHECK-LABEL: name: no_fold_andn2_select_condition_live_out_phi11  ; CHECK: bb.0:12  ; CHECK-NEXT:   successors: %bb.2(0x80000000)13  ; CHECK-NEXT: {{  $}}14  ; CHECK-NEXT:   [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 -115  ; CHECK-NEXT:   undef [[V_MOV_B32_e32_:%[0-9]+]].sub0:vreg_64 = V_MOV_B32_e32 0, implicit $exec16  ; CHECK-NEXT:   S_BRANCH %bb.217  ; CHECK-NEXT: {{  $}}18  ; CHECK-NEXT: bb.1:19  ; CHECK-NEXT:   S_ENDPGM 020  ; CHECK-NEXT: {{  $}}21  ; CHECK-NEXT: bb.2:22  ; CHECK-NEXT:   successors: %bb.1(0x40000000), %bb.2(0x40000000)23  ; CHECK-NEXT: {{  $}}24  ; CHECK-NEXT:   [[V_CNDMASK_B32_e64_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, 1, [[S_MOV_B64_]], implicit $exec25  ; CHECK-NEXT:   V_CMP_NE_U32_e32 1, [[V_CNDMASK_B32_e64_]], implicit-def $vcc, implicit $exec26  ; CHECK-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]].sub1:vreg_64 = COPY [[V_MOV_B32_e32_]].sub027  ; CHECK-NEXT:   DS_WRITE_B64_gfx9 undef %3:vgpr_32, [[V_MOV_B32_e32_]], 0, 0, implicit $exec :: (store (s64), addrspace 3)28  ; CHECK-NEXT:   ATOMIC_FENCE 4, 229  ; CHECK-NEXT:   [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 030  ; CHECK-NEXT:   $vcc = S_AND_B64 $exec, $vcc, implicit-def dead $scc31  ; CHECK-NEXT:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc32  ; CHECK-NEXT:   S_BRANCH %bb.233  bb.0:34    successors: %bb.235 36    %7:sreg_64_xexec = S_MOV_B64 -137    undef %5.sub0:vreg_64 = V_MOV_B32_e32 0, implicit $exec38    S_BRANCH %bb.239 40  bb.1:41    S_ENDPGM 042 43  bb.2:44    successors: %bb.1, %bb.245 46    %4:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, 1, %7, implicit $exec47    V_CMP_NE_U32_e32 1, %4, implicit-def $vcc, implicit $exec48    %5.sub1:vreg_64 = COPY %5.sub049    DS_WRITE_B64_gfx9 undef %6:vgpr_32, %5, 0, 0, implicit $exec :: (store (s64), addrspace 3)50    ATOMIC_FENCE 4, 251    %7:sreg_64_xexec = S_MOV_B64 052    $vcc = S_AND_B64 $exec, killed $vcc, implicit-def dead $scc53    S_CBRANCH_VCCNZ %bb.1, implicit killed $vcc54    S_BRANCH %bb.255 56...57 58# It's OK to fold this, since the phi def is after the andn2 insert point.59---60name:            fold_andn2_select_condition_live_out_phi_reorder61tracksRegLiveness: true62body:             |63  ; CHECK-LABEL: name: fold_andn2_select_condition_live_out_phi_reorder64  ; CHECK: bb.0:65  ; CHECK-NEXT:   successors: %bb.2(0x80000000)66  ; CHECK-NEXT: {{  $}}67  ; CHECK-NEXT:   [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 -168  ; CHECK-NEXT:   undef [[V_MOV_B32_e32_:%[0-9]+]].sub0:vreg_64 = V_MOV_B32_e32 0, implicit $exec69  ; CHECK-NEXT:   S_BRANCH %bb.270  ; CHECK-NEXT: {{  $}}71  ; CHECK-NEXT: bb.1:72  ; CHECK-NEXT:   S_ENDPGM 073  ; CHECK-NEXT: {{  $}}74  ; CHECK-NEXT: bb.2:75  ; CHECK-NEXT:   successors: %bb.1(0x40000000), %bb.2(0x40000000)76  ; CHECK-NEXT: {{  $}}77  ; CHECK-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]].sub1:vreg_64 = COPY [[V_MOV_B32_e32_]].sub078  ; CHECK-NEXT:   DS_WRITE_B64_gfx9 undef %3:vgpr_32, [[V_MOV_B32_e32_]], 0, 0, implicit $exec :: (store (s64), addrspace 3)79  ; CHECK-NEXT:   ATOMIC_FENCE 4, 280  ; CHECK-NEXT:   $vcc = S_ANDN2_B64 $exec, [[S_MOV_B64_]], implicit-def dead $scc81  ; CHECK-NEXT:   [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 082  ; CHECK-NEXT:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc83  ; CHECK-NEXT:   S_BRANCH %bb.284  bb.0:85    successors: %bb.286 87    %7:sreg_64_xexec = S_MOV_B64 -188    undef %5.sub0:vreg_64 = V_MOV_B32_e32 0, implicit $exec89    S_BRANCH %bb.290 91  bb.1:92    S_ENDPGM 093 94  bb.2:95    successors: %bb.1, %bb.296 97    %4:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, 1, %7, implicit $exec98    V_CMP_NE_U32_e32 1, %4, implicit-def $vcc, implicit $exec99    %5.sub1:vreg_64 = COPY %5.sub0100    DS_WRITE_B64_gfx9 undef %6:vgpr_32, %5, 0, 0, implicit $exec :: (store (s64), addrspace 3)101    ATOMIC_FENCE 4, 2102    $vcc = S_AND_B64 $exec, killed $vcc, implicit-def dead $scc103    %7:sreg_64_xexec = S_MOV_B64 0104    S_CBRANCH_VCCNZ %bb.1, implicit killed $vcc105    S_BRANCH %bb.2106 107...108 109---110name:            no_fold_andn2_select_condition_live_out_phi_physreg111tracksRegLiveness: true112body:             |113  ; CHECK-LABEL: name: no_fold_andn2_select_condition_live_out_phi_physreg114  ; CHECK: bb.0:115  ; CHECK-NEXT:   successors: %bb.2(0x80000000)116  ; CHECK-NEXT: {{  $}}117  ; CHECK-NEXT:   $sgpr4_sgpr5 = S_MOV_B64 -1118  ; CHECK-NEXT:   undef [[V_MOV_B32_e32_:%[0-9]+]].sub0:vreg_64 = V_MOV_B32_e32 0, implicit $exec119  ; CHECK-NEXT:   S_BRANCH %bb.2120  ; CHECK-NEXT: {{  $}}121  ; CHECK-NEXT: bb.1:122  ; CHECK-NEXT:   S_ENDPGM 0123  ; CHECK-NEXT: {{  $}}124  ; CHECK-NEXT: bb.2:125  ; CHECK-NEXT:   successors: %bb.1(0x40000000), %bb.2(0x40000000)126  ; CHECK-NEXT:   liveins: $sgpr4_sgpr5127  ; CHECK-NEXT: {{  $}}128  ; CHECK-NEXT:   [[V_CNDMASK_B32_e64_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, 1, $sgpr4_sgpr5, implicit $exec129  ; CHECK-NEXT:   V_CMP_NE_U32_e32 1, [[V_CNDMASK_B32_e64_]], implicit-def $vcc, implicit $exec130  ; CHECK-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]].sub1:vreg_64 = COPY [[V_MOV_B32_e32_]].sub0131  ; CHECK-NEXT:   DS_WRITE_B64_gfx9 undef %2:vgpr_32, [[V_MOV_B32_e32_]], 0, 0, implicit $exec :: (store (s64), addrspace 3)132  ; CHECK-NEXT:   ATOMIC_FENCE 4, 2133  ; CHECK-NEXT:   $sgpr4_sgpr5 = S_MOV_B64 0134  ; CHECK-NEXT:   $vcc = S_AND_B64 $exec, $vcc, implicit-def dead $scc135  ; CHECK-NEXT:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc136  ; CHECK-NEXT:   S_BRANCH %bb.2137  bb.0:138    successors: %bb.2139 140    $sgpr4_sgpr5 = S_MOV_B64 -1141    undef %5.sub0:vreg_64 = V_MOV_B32_e32 0, implicit $exec142    S_BRANCH %bb.2143 144  bb.1:145    S_ENDPGM 0146 147  bb.2:148    successors: %bb.1, %bb.2149    liveins: $sgpr4_sgpr5150 151    %4:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, 1, $sgpr4_sgpr5, implicit $exec152    V_CMP_NE_U32_e32 1, %4, implicit-def $vcc, implicit $exec153    %5.sub1:vreg_64 = COPY %5.sub0154    DS_WRITE_B64_gfx9 undef %6:vgpr_32, %5, 0, 0, implicit $exec :: (store (s64), addrspace 3)155    ATOMIC_FENCE 4, 2156    $sgpr4_sgpr5 = S_MOV_B64 0157    $vcc = S_AND_B64 $exec, killed $vcc, implicit-def dead $scc158    S_CBRANCH_VCCNZ %bb.1, implicit killed $vcc159    S_BRANCH %bb.2160 161...162 163---164name:            fold_andn2_select_condition_live_out_phi_physreg_reorder165tracksRegLiveness: true166body:             |167  ; CHECK-LABEL: name: fold_andn2_select_condition_live_out_phi_physreg_reorder168  ; CHECK: bb.0:169  ; CHECK-NEXT:   successors: %bb.2(0x80000000)170  ; CHECK-NEXT: {{  $}}171  ; CHECK-NEXT:   $sgpr4_sgpr5 = S_MOV_B64 -1172  ; CHECK-NEXT:   undef [[V_MOV_B32_e32_:%[0-9]+]].sub0:vreg_64 = V_MOV_B32_e32 0, implicit $exec173  ; CHECK-NEXT:   S_BRANCH %bb.2174  ; CHECK-NEXT: {{  $}}175  ; CHECK-NEXT: bb.1:176  ; CHECK-NEXT:   S_ENDPGM 0177  ; CHECK-NEXT: {{  $}}178  ; CHECK-NEXT: bb.2:179  ; CHECK-NEXT:   successors: %bb.1(0x40000000), %bb.2(0x40000000)180  ; CHECK-NEXT:   liveins: $sgpr4_sgpr5181  ; CHECK-NEXT: {{  $}}182  ; CHECK-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]].sub1:vreg_64 = COPY [[V_MOV_B32_e32_]].sub0183  ; CHECK-NEXT:   DS_WRITE_B64_gfx9 undef %2:vgpr_32, [[V_MOV_B32_e32_]], 0, 0, implicit $exec :: (store (s64), addrspace 3)184  ; CHECK-NEXT:   ATOMIC_FENCE 4, 2185  ; CHECK-NEXT:   $vcc = S_ANDN2_B64 $exec, $sgpr4_sgpr5, implicit-def dead $scc186  ; CHECK-NEXT:   $sgpr4_sgpr5 = S_MOV_B64 0187  ; CHECK-NEXT:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc188  ; CHECK-NEXT:   S_BRANCH %bb.2189  bb.0:190    successors: %bb.2191 192    $sgpr4_sgpr5 = S_MOV_B64 -1193    undef %5.sub0:vreg_64 = V_MOV_B32_e32 0, implicit $exec194    S_BRANCH %bb.2195 196  bb.1:197    S_ENDPGM 0198 199  bb.2:200    successors: %bb.1, %bb.2201    liveins: $sgpr4_sgpr5202 203    %4:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, 1, $sgpr4_sgpr5, implicit $exec204    V_CMP_NE_U32_e32 1, %4, implicit-def $vcc, implicit $exec205    %5.sub1:vreg_64 = COPY %5.sub0206    DS_WRITE_B64_gfx9 undef %6:vgpr_32, %5, 0, 0, implicit $exec :: (store (s64), addrspace 3)207    ATOMIC_FENCE 4, 2208    $vcc = S_AND_B64 $exec, killed $vcc, implicit-def dead $scc209    $sgpr4_sgpr5 = S_MOV_B64 0210    S_CBRANCH_VCCNZ %bb.1, implicit killed $vcc211    S_BRANCH %bb.2212 213...214