145 lines · plain
1# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 42# RUN: llc --verify-machineinstrs -mtriple=aarch64 -o - %s -run-pass pipeliner -aarch64-enable-pipeliner -pipeliner-mve-cg -pipeliner-force-ii=3 -mcpu=neoverse-n1 2>&1 | FileCheck %s3 4# test pipeliner code genearation by MVE algorithm5# #stages: 2, unroll count: 36 7# the calculation result can be checked as follows (driver code written by C):8# for (i=2; i<N; i++)9# func_noswp(i, 1) == func(i, 1)10 11...12---13name: func14tracksRegLiveness: true15body: |16 ; CHECK-LABEL: name: func17 ; CHECK: bb.0.entry:18 ; CHECK-NEXT: successors: %bb.3(0x80000000)19 ; CHECK-NEXT: liveins: $x0, $x120 ; CHECK-NEXT: {{ $}}21 ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr64 = COPY $x022 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr64 = COPY $x123 ; CHECK-NEXT: [[FMOVDi:%[0-9]+]]:fpr64 = FMOVDi 124 ; CHECK-NEXT: B %bb.325 ; CHECK-NEXT: {{ $}}26 ; CHECK-NEXT: bb.3:27 ; CHECK-NEXT: successors: %bb.4(0x40000000), %bb.7(0x40000000)28 ; CHECK-NEXT: {{ $}}29 ; CHECK-NEXT: [[ADDXrr:%[0-9]+]]:gpr64 = ADDXrr [[COPY1]], [[COPY1]]30 ; CHECK-NEXT: dead $xzr = SUBSXrr [[COPY]], [[ADDXrr]], implicit-def $nzcv31 ; CHECK-NEXT: [[CSINCXr:%[0-9]+]]:gpr64common = CSINCXr $xzr, $xzr, 1, implicit $nzcv32 ; CHECK-NEXT: [[ADDXrr1:%[0-9]+]]:gpr64 = ADDXrr [[ADDXrr]], [[COPY1]]33 ; CHECK-NEXT: dead $xzr = SUBSXrr [[COPY]], [[ADDXrr1]], implicit-def $nzcv34 ; CHECK-NEXT: [[CSINCXr1:%[0-9]+]]:gpr64common = CSINCXr [[CSINCXr]], [[CSINCXr]], 1, implicit $nzcv35 ; CHECK-NEXT: [[ADDXrr2:%[0-9]+]]:gpr64 = ADDXrr [[ADDXrr1]], [[COPY1]]36 ; CHECK-NEXT: dead $xzr = SUBSXrr [[COPY]], [[ADDXrr2]], implicit-def $nzcv37 ; CHECK-NEXT: [[CSINCXr2:%[0-9]+]]:gpr64common = CSINCXr [[CSINCXr1]], [[CSINCXr1]], 1, implicit $nzcv38 ; CHECK-NEXT: [[ADDXrr3:%[0-9]+]]:gpr64 = ADDXrr [[ADDXrr2]], [[COPY1]]39 ; CHECK-NEXT: dead $xzr = SUBSXrr [[COPY]], [[ADDXrr3]], implicit-def $nzcv40 ; CHECK-NEXT: [[CSINCXr3:%[0-9]+]]:gpr64common = CSINCXr [[CSINCXr2]], [[CSINCXr2]], 1, implicit $nzcv41 ; CHECK-NEXT: dead $xzr = SUBSXri [[CSINCXr3]], 0, 0, implicit-def $nzcv42 ; CHECK-NEXT: Bcc 0, %bb.4, implicit $nzcv43 ; CHECK-NEXT: B %bb.744 ; CHECK-NEXT: {{ $}}45 ; CHECK-NEXT: bb.4:46 ; CHECK-NEXT: successors: %bb.5(0x80000000)47 ; CHECK-NEXT: {{ $}}48 ; CHECK-NEXT: [[FADDDrr:%[0-9]+]]:fpr64 = FADDDrr [[FMOVDi]], [[FMOVDi]], implicit $fpcr49 ; CHECK-NEXT: [[ADDXrr4:%[0-9]+]]:gpr64 = ADDXrr [[COPY1]], [[COPY1]]50 ; CHECK-NEXT: dead $xzr = SUBSXrr [[COPY]], [[ADDXrr4]], implicit-def $nzcv51 ; CHECK-NEXT: [[FADDDrr1:%[0-9]+]]:fpr64 = FADDDrr [[FADDDrr]], [[FMOVDi]], implicit $fpcr52 ; CHECK-NEXT: {{ $}}53 ; CHECK-NEXT: bb.5:54 ; CHECK-NEXT: successors: %bb.5(0x40000000), %bb.6(0x40000000)55 ; CHECK-NEXT: {{ $}}56 ; CHECK-NEXT: [[PHI:%[0-9]+]]:fpr64 = PHI [[FADDDrr5:%[0-9]+]], %bb.5, [[FMOVDi]], %bb.457 ; CHECK-NEXT: [[PHI1:%[0-9]+]]:gpr64 = PHI [[ADDXrr6:%[0-9]+]], %bb.5, [[COPY1]], %bb.458 ; CHECK-NEXT: [[PHI2:%[0-9]+]]:fpr64 = PHI [[FADDDrr8:%[0-9]+]], %bb.5, [[FADDDrr]], %bb.459 ; CHECK-NEXT: [[PHI3:%[0-9]+]]:gpr64 = PHI [[ADDXrr7:%[0-9]+]], %bb.5, [[ADDXrr4]], %bb.460 ; CHECK-NEXT: [[PHI4:%[0-9]+]]:fpr64 = PHI [[FADDDrr10:%[0-9]+]], %bb.5, [[FADDDrr1]], %bb.461 ; CHECK-NEXT: [[FADDDrr2:%[0-9]+]]:fpr64 = FADDDrr [[FMOVDi]], [[PHI2]], implicit $fpcr62 ; CHECK-NEXT: [[ADDXrr5:%[0-9]+]]:gpr64 = ADDXrr [[PHI3]], [[COPY1]]63 ; CHECK-NEXT: [[FADDDrr3:%[0-9]+]]:fpr64 = FADDDrr [[PHI4]], [[PHI]], implicit $fpcr64 ; CHECK-NEXT: dead $xzr = SUBSXrr [[COPY]], [[ADDXrr5]], implicit-def $nzcv65 ; CHECK-NEXT: [[FADDDrr4:%[0-9]+]]:fpr64 = FADDDrr [[FADDDrr2]], [[PHI2]], implicit $fpcr66 ; CHECK-NEXT: [[FADDDrr5]]:fpr64 = FADDDrr [[FMOVDi]], [[FADDDrr2]], implicit $fpcr67 ; CHECK-NEXT: [[ADDXrr6]]:gpr64 = ADDXrr [[ADDXrr5]], [[COPY1]]68 ; CHECK-NEXT: [[FADDDrr6:%[0-9]+]]:fpr64 = FADDDrr [[FADDDrr4]], [[PHI2]], implicit $fpcr69 ; CHECK-NEXT: dead $xzr = SUBSXrr [[COPY]], [[ADDXrr6]], implicit-def $nzcv70 ; CHECK-NEXT: [[FADDDrr7:%[0-9]+]]:fpr64 = FADDDrr [[FADDDrr5]], [[FADDDrr2]], implicit $fpcr71 ; CHECK-NEXT: [[FADDDrr8]]:fpr64 = FADDDrr [[FMOVDi]], [[FADDDrr5]], implicit $fpcr72 ; CHECK-NEXT: [[ADDXrr7]]:gpr64 = ADDXrr [[ADDXrr6]], [[COPY1]]73 ; CHECK-NEXT: [[FADDDrr9:%[0-9]+]]:fpr64 = FADDDrr [[FADDDrr7]], [[FADDDrr2]], implicit $fpcr74 ; CHECK-NEXT: dead $xzr = SUBSXrr [[COPY]], [[ADDXrr7]], implicit-def $nzcv75 ; CHECK-NEXT: [[FADDDrr10]]:fpr64 = FADDDrr [[FADDDrr8]], [[FADDDrr5]], implicit $fpcr76 ; CHECK-NEXT: [[CSINCXr4:%[0-9]+]]:gpr64common = CSINCXr $xzr, $xzr, 1, implicit $nzcv77 ; CHECK-NEXT: [[ADDXrr8:%[0-9]+]]:gpr64 = ADDXrr [[ADDXrr7]], [[COPY1]]78 ; CHECK-NEXT: dead $xzr = SUBSXrr [[COPY]], [[ADDXrr8]], implicit-def $nzcv79 ; CHECK-NEXT: [[CSINCXr5:%[0-9]+]]:gpr64common = CSINCXr [[CSINCXr4]], [[CSINCXr4]], 1, implicit $nzcv80 ; CHECK-NEXT: [[ADDXrr9:%[0-9]+]]:gpr64 = ADDXrr [[ADDXrr8]], [[COPY1]]81 ; CHECK-NEXT: dead $xzr = SUBSXrr [[COPY]], [[ADDXrr9]], implicit-def $nzcv82 ; CHECK-NEXT: [[CSINCXr6:%[0-9]+]]:gpr64common = CSINCXr [[CSINCXr5]], [[CSINCXr5]], 1, implicit $nzcv83 ; CHECK-NEXT: dead $xzr = SUBSXri [[CSINCXr6]], 0, 0, implicit-def $nzcv84 ; CHECK-NEXT: Bcc 0, %bb.5, implicit $nzcv85 ; CHECK-NEXT: B %bb.686 ; CHECK-NEXT: {{ $}}87 ; CHECK-NEXT: bb.6:88 ; CHECK-NEXT: successors: %bb.7(0x40000000), %bb.2(0x40000000)89 ; CHECK-NEXT: {{ $}}90 ; CHECK-NEXT: [[FADDDrr11:%[0-9]+]]:fpr64 = FADDDrr [[FADDDrr10]], [[FADDDrr5]], implicit $fpcr91 ; CHECK-NEXT: dead $xzr = SUBSXrr [[COPY]], [[ADDXrr7]], implicit-def $nzcv92 ; CHECK-NEXT: [[CSINCXr7:%[0-9]+]]:gpr64common = CSINCXr $xzr, $xzr, 1, implicit $nzcv93 ; CHECK-NEXT: dead $xzr = SUBSXri [[CSINCXr7]], 0, 0, implicit-def $nzcv94 ; CHECK-NEXT: Bcc 0, %bb.7, implicit $nzcv95 ; CHECK-NEXT: B %bb.296 ; CHECK-NEXT: {{ $}}97 ; CHECK-NEXT: bb.7:98 ; CHECK-NEXT: successors: %bb.1(0x80000000)99 ; CHECK-NEXT: {{ $}}100 ; CHECK-NEXT: [[PHI5:%[0-9]+]]:fpr64 = PHI [[FMOVDi]], %bb.3, [[FADDDrr8]], %bb.6101 ; CHECK-NEXT: [[PHI6:%[0-9]+]]:gpr64 = PHI [[COPY1]], %bb.3, [[ADDXrr7]], %bb.6102 ; CHECK-NEXT: B %bb.1103 ; CHECK-NEXT: {{ $}}104 ; CHECK-NEXT: bb.1:105 ; CHECK-NEXT: successors: %bb.1(0x40000000), %bb.2(0x40000000)106 ; CHECK-NEXT: {{ $}}107 ; CHECK-NEXT: [[PHI7:%[0-9]+]]:gpr64 = PHI [[PHI6]], %bb.7, [[ADDXrr10:%[0-9]+]], %bb.1108 ; CHECK-NEXT: [[PHI8:%[0-9]+]]:fpr64 = PHI [[PHI5]], %bb.7, [[FADDDrr12:%[0-9]+]], %bb.1109 ; CHECK-NEXT: [[ADDXrr10]]:gpr64 = ADDXrr [[PHI7]], [[COPY1]]110 ; CHECK-NEXT: dead $xzr = SUBSXrr [[COPY]], [[ADDXrr10]], implicit-def $nzcv111 ; CHECK-NEXT: [[FADDDrr12]]:fpr64 = FADDDrr [[FMOVDi]], [[PHI8]], implicit $fpcr112 ; CHECK-NEXT: [[FADDDrr13:%[0-9]+]]:fpr64 = FADDDrr [[FADDDrr12]], [[PHI8]], implicit $fpcr113 ; CHECK-NEXT: [[FADDDrr14:%[0-9]+]]:fpr64 = FADDDrr [[FADDDrr13]], [[PHI8]], implicit $fpcr114 ; CHECK-NEXT: Bcc 1, %bb.1, implicit $nzcv115 ; CHECK-NEXT: B %bb.2116 ; CHECK-NEXT: {{ $}}117 ; CHECK-NEXT: bb.2:118 ; CHECK-NEXT: [[PHI9:%[0-9]+]]:fpr64 = PHI [[FADDDrr12]], %bb.1, [[FADDDrr8]], %bb.6119 ; CHECK-NEXT: [[PHI10:%[0-9]+]]:fpr64 = PHI [[FADDDrr14]], %bb.1, [[FADDDrr11]], %bb.6120 ; CHECK-NEXT: [[FMULDrr:%[0-9]+]]:fpr64 = FMULDrr [[PHI9]], [[PHI10]], implicit $fpcr121 ; CHECK-NEXT: $d0 = COPY [[FMULDrr]]122 ; CHECK-NEXT: RET_ReallyLR implicit $d0123 bb.0.entry:124 liveins: $x0, $x1125 %10:gpr64 = COPY $x0126 %11:gpr64 = COPY $x1127 %20:fpr64 = FMOVDi 1128 129 bb.1:130 %12:gpr64 = PHI %11, %bb.0, %13, %bb.1131 %24:fpr64 = PHI %20, %bb.0, %21, %bb.1132 %13:gpr64 = ADDXrr %12, %11133 dead $xzr = SUBSXrr %10, %13, implicit-def $nzcv134 %21:fpr64 = FADDDrr %20, %24, implicit $fpcr135 %22:fpr64 = FADDDrr %21, %24, implicit $fpcr136 %23:fpr64 = FADDDrr %22, %24, implicit $fpcr137 Bcc 1, %bb.1, implicit $nzcv138 B %bb.2139 140 bb.2:141 %25:fpr64 = FMULDrr %21, %23, implicit $fpcr142 $d0 = COPY %25143 RET_ReallyLR implicit $d0144...145