204 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 52; RUN: llc < %s -mtriple=arm -mattr=+vfp4d16sp,+fullfp16,-bf16 -stop-after=finalize-isel | FileCheck %s --check-prefixes=CHECK-NOBF163 4; Check that the output instructions have the same fast math flags as the input5; fadd, even when bf16 is legalized to f32.6; FIXME: We should also test with +bf16, but it currently fails in instruction7; selection.8 9define bfloat @normal_fadd(bfloat %x, bfloat %y) {10 ; CHECK-NOBF16-LABEL: name: normal_fadd11 ; CHECK-NOBF16: bb.0.entry:12 ; CHECK-NOBF16-NEXT: liveins: $r0, $r113 ; CHECK-NOBF16-NEXT: {{ $}}14 ; CHECK-NOBF16-NEXT: [[COPY:%[0-9]+]]:gpr = COPY $r115 ; CHECK-NOBF16-NEXT: [[COPY1:%[0-9]+]]:gpr = COPY $r016 ; CHECK-NOBF16-NEXT: [[MOVsi:%[0-9]+]]:gpr = MOVsi [[COPY]], 130, 14 /* CC::al */, $noreg, $noreg17 ; CHECK-NOBF16-NEXT: [[VMOVSR:%[0-9]+]]:spr = VMOVSR killed [[MOVsi]], 14 /* CC::al */, $noreg18 ; CHECK-NOBF16-NEXT: [[MOVsi1:%[0-9]+]]:gpr = MOVsi [[COPY1]], 130, 14 /* CC::al */, $noreg, $noreg19 ; CHECK-NOBF16-NEXT: [[VMOVSR1:%[0-9]+]]:spr = VMOVSR killed [[MOVsi1]], 14 /* CC::al */, $noreg20 ; CHECK-NOBF16-NEXT: [[VADDS:%[0-9]+]]:spr = nofpexcept VADDS killed [[VMOVSR1]], killed [[VMOVSR]], 14 /* CC::al */, $noreg, implicit $fpscr21 ; CHECK-NOBF16-NEXT: [[VMOVRS:%[0-9]+]]:gpr = VMOVRS killed [[VADDS]], 14 /* CC::al */, $noreg22 ; CHECK-NOBF16-NEXT: ADJCALLSTACKDOWN 0, 0, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp23 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRS]]24 ; CHECK-NOBF16-NEXT: BL &__truncsfbf2, csr_aapcs, implicit-def dead $lr, implicit $sp, implicit $r0, implicit-def $sp, implicit-def $r025 ; CHECK-NOBF16-NEXT: ADJCALLSTACKUP 0, -1, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp26 ; CHECK-NOBF16-NEXT: [[COPY2:%[0-9]+]]:rgpr = COPY $r027 ; CHECK-NOBF16-NEXT: [[VMOVHR:%[0-9]+]]:hpr = VMOVHR [[COPY2]], 14, $noreg28 ; CHECK-NOBF16-NEXT: [[VMOVRH:%[0-9]+]]:rgpr = VMOVRH killed [[VMOVHR]], 14, $noreg29 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRH]]30 ; CHECK-NOBF16-NEXT: MOVPCLR 14 /* CC::al */, $noreg, implicit $r031entry:32 %add = fadd bfloat %x, %y33 ret bfloat %add34}35 36define bfloat @fast_fadd(bfloat %x, bfloat %y) {37 ; CHECK-NOBF16-LABEL: name: fast_fadd38 ; CHECK-NOBF16: bb.0.entry:39 ; CHECK-NOBF16-NEXT: liveins: $r0, $r140 ; CHECK-NOBF16-NEXT: {{ $}}41 ; CHECK-NOBF16-NEXT: [[COPY:%[0-9]+]]:gpr = COPY $r142 ; CHECK-NOBF16-NEXT: [[COPY1:%[0-9]+]]:gpr = COPY $r043 ; CHECK-NOBF16-NEXT: [[MOVsi:%[0-9]+]]:gpr = MOVsi [[COPY]], 130, 14 /* CC::al */, $noreg, $noreg44 ; CHECK-NOBF16-NEXT: [[VMOVSR:%[0-9]+]]:spr = VMOVSR killed [[MOVsi]], 14 /* CC::al */, $noreg45 ; CHECK-NOBF16-NEXT: [[MOVsi1:%[0-9]+]]:gpr = MOVsi [[COPY1]], 130, 14 /* CC::al */, $noreg, $noreg46 ; CHECK-NOBF16-NEXT: [[VMOVSR1:%[0-9]+]]:spr = VMOVSR killed [[MOVsi1]], 14 /* CC::al */, $noreg47 ; CHECK-NOBF16-NEXT: [[VADDS:%[0-9]+]]:spr = nnan ninf nsz arcp contract afn reassoc nofpexcept VADDS killed [[VMOVSR1]], killed [[VMOVSR]], 14 /* CC::al */, $noreg, implicit $fpscr48 ; CHECK-NOBF16-NEXT: [[VMOVRS:%[0-9]+]]:gpr = VMOVRS killed [[VADDS]], 14 /* CC::al */, $noreg49 ; CHECK-NOBF16-NEXT: ADJCALLSTACKDOWN 0, 0, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp50 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRS]]51 ; CHECK-NOBF16-NEXT: BL &__truncsfbf2, csr_aapcs, implicit-def dead $lr, implicit $sp, implicit $r0, implicit-def $sp, implicit-def $r052 ; CHECK-NOBF16-NEXT: ADJCALLSTACKUP 0, -1, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp53 ; CHECK-NOBF16-NEXT: [[COPY2:%[0-9]+]]:rgpr = COPY $r054 ; CHECK-NOBF16-NEXT: [[VMOVHR:%[0-9]+]]:hpr = VMOVHR [[COPY2]], 14, $noreg55 ; CHECK-NOBF16-NEXT: [[VMOVRH:%[0-9]+]]:rgpr = VMOVRH killed [[VMOVHR]], 14, $noreg56 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRH]]57 ; CHECK-NOBF16-NEXT: MOVPCLR 14 /* CC::al */, $noreg, implicit $r058entry:59 %add = fadd fast bfloat %x, %y60 ret bfloat %add61}62 63define bfloat @ninf_fadd(bfloat %x, bfloat %y) {64 ; CHECK-NOBF16-LABEL: name: ninf_fadd65 ; CHECK-NOBF16: bb.0.entry:66 ; CHECK-NOBF16-NEXT: liveins: $r0, $r167 ; CHECK-NOBF16-NEXT: {{ $}}68 ; CHECK-NOBF16-NEXT: [[COPY:%[0-9]+]]:gpr = COPY $r169 ; CHECK-NOBF16-NEXT: [[COPY1:%[0-9]+]]:gpr = COPY $r070 ; CHECK-NOBF16-NEXT: [[MOVsi:%[0-9]+]]:gpr = MOVsi [[COPY]], 130, 14 /* CC::al */, $noreg, $noreg71 ; CHECK-NOBF16-NEXT: [[VMOVSR:%[0-9]+]]:spr = VMOVSR killed [[MOVsi]], 14 /* CC::al */, $noreg72 ; CHECK-NOBF16-NEXT: [[MOVsi1:%[0-9]+]]:gpr = MOVsi [[COPY1]], 130, 14 /* CC::al */, $noreg, $noreg73 ; CHECK-NOBF16-NEXT: [[VMOVSR1:%[0-9]+]]:spr = VMOVSR killed [[MOVsi1]], 14 /* CC::al */, $noreg74 ; CHECK-NOBF16-NEXT: [[VADDS:%[0-9]+]]:spr = ninf nofpexcept VADDS killed [[VMOVSR1]], killed [[VMOVSR]], 14 /* CC::al */, $noreg, implicit $fpscr75 ; CHECK-NOBF16-NEXT: [[VMOVRS:%[0-9]+]]:gpr = VMOVRS killed [[VADDS]], 14 /* CC::al */, $noreg76 ; CHECK-NOBF16-NEXT: ADJCALLSTACKDOWN 0, 0, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp77 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRS]]78 ; CHECK-NOBF16-NEXT: BL &__truncsfbf2, csr_aapcs, implicit-def dead $lr, implicit $sp, implicit $r0, implicit-def $sp, implicit-def $r079 ; CHECK-NOBF16-NEXT: ADJCALLSTACKUP 0, -1, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp80 ; CHECK-NOBF16-NEXT: [[COPY2:%[0-9]+]]:rgpr = COPY $r081 ; CHECK-NOBF16-NEXT: [[VMOVHR:%[0-9]+]]:hpr = VMOVHR [[COPY2]], 14, $noreg82 ; CHECK-NOBF16-NEXT: [[VMOVRH:%[0-9]+]]:rgpr = VMOVRH killed [[VMOVHR]], 14, $noreg83 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRH]]84 ; CHECK-NOBF16-NEXT: MOVPCLR 14 /* CC::al */, $noreg, implicit $r085entry:86 %add = fadd ninf bfloat %x, %y87 ret bfloat %add88}89 90; Check that when we have the right fast math flags the converts in between the91; two fadds are removed.92 93define bfloat @normal_fadd_sequence(bfloat %x, bfloat %y, bfloat %z) {94 ; CHECK-NOBF16-LABEL: name: normal_fadd_sequence95 ; CHECK-NOBF16: bb.0.entry:96 ; CHECK-NOBF16-NEXT: liveins: $r0, $r1, $r297 ; CHECK-NOBF16-NEXT: {{ $}}98 ; CHECK-NOBF16-NEXT: [[COPY:%[0-9]+]]:gpr = COPY $r299 ; CHECK-NOBF16-NEXT: [[COPY1:%[0-9]+]]:gpr = COPY $r1100 ; CHECK-NOBF16-NEXT: [[COPY2:%[0-9]+]]:gpr = COPY $r0101 ; CHECK-NOBF16-NEXT: [[MOVsi:%[0-9]+]]:gpr = MOVsi [[COPY1]], 130, 14 /* CC::al */, $noreg, $noreg102 ; CHECK-NOBF16-NEXT: [[VMOVSR:%[0-9]+]]:spr = VMOVSR killed [[MOVsi]], 14 /* CC::al */, $noreg103 ; CHECK-NOBF16-NEXT: [[MOVsi1:%[0-9]+]]:gpr = MOVsi [[COPY2]], 130, 14 /* CC::al */, $noreg, $noreg104 ; CHECK-NOBF16-NEXT: [[VMOVSR1:%[0-9]+]]:spr = VMOVSR killed [[MOVsi1]], 14 /* CC::al */, $noreg105 ; CHECK-NOBF16-NEXT: [[VADDS:%[0-9]+]]:spr = nofpexcept VADDS killed [[VMOVSR1]], killed [[VMOVSR]], 14 /* CC::al */, $noreg, implicit $fpscr106 ; CHECK-NOBF16-NEXT: [[VMOVRS:%[0-9]+]]:gpr = VMOVRS killed [[VADDS]], 14 /* CC::al */, $noreg107 ; CHECK-NOBF16-NEXT: ADJCALLSTACKDOWN 0, 0, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp108 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRS]]109 ; CHECK-NOBF16-NEXT: BL &__truncsfbf2, csr_aapcs, implicit-def dead $lr, implicit $sp, implicit $r0, implicit-def $sp, implicit-def $r0110 ; CHECK-NOBF16-NEXT: ADJCALLSTACKUP 0, -1, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp111 ; CHECK-NOBF16-NEXT: [[COPY3:%[0-9]+]]:gpr = COPY $r0112 ; CHECK-NOBF16-NEXT: [[MOVsi2:%[0-9]+]]:gpr = MOVsi [[COPY]], 130, 14 /* CC::al */, $noreg, $noreg113 ; CHECK-NOBF16-NEXT: [[VMOVSR2:%[0-9]+]]:spr = VMOVSR killed [[MOVsi2]], 14 /* CC::al */, $noreg114 ; CHECK-NOBF16-NEXT: [[MOVsi3:%[0-9]+]]:gpr = MOVsi [[COPY3]], 130, 14 /* CC::al */, $noreg, $noreg115 ; CHECK-NOBF16-NEXT: [[VMOVSR3:%[0-9]+]]:spr = VMOVSR killed [[MOVsi3]], 14 /* CC::al */, $noreg116 ; CHECK-NOBF16-NEXT: [[VADDS1:%[0-9]+]]:spr = nofpexcept VADDS killed [[VMOVSR3]], killed [[VMOVSR2]], 14 /* CC::al */, $noreg, implicit $fpscr117 ; CHECK-NOBF16-NEXT: [[VMOVRS1:%[0-9]+]]:gpr = VMOVRS killed [[VADDS1]], 14 /* CC::al */, $noreg118 ; CHECK-NOBF16-NEXT: ADJCALLSTACKDOWN 0, 0, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp119 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRS1]]120 ; CHECK-NOBF16-NEXT: BL &__truncsfbf2, csr_aapcs, implicit-def dead $lr, implicit $sp, implicit $r0, implicit-def $sp, implicit-def $r0121 ; CHECK-NOBF16-NEXT: ADJCALLSTACKUP 0, -1, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp122 ; CHECK-NOBF16-NEXT: [[COPY4:%[0-9]+]]:rgpr = COPY $r0123 ; CHECK-NOBF16-NEXT: [[VMOVHR:%[0-9]+]]:hpr = VMOVHR [[COPY4]], 14, $noreg124 ; CHECK-NOBF16-NEXT: [[VMOVRH:%[0-9]+]]:rgpr = VMOVRH killed [[VMOVHR]], 14, $noreg125 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRH]]126 ; CHECK-NOBF16-NEXT: MOVPCLR 14 /* CC::al */, $noreg, implicit $r0127entry:128 %add1 = fadd bfloat %x, %y129 %add2 = fadd bfloat %add1, %z130 ret bfloat %add2131}132 133define bfloat @nnan_ninf_contract_fadd_sequence(bfloat %x, bfloat %y, bfloat %z) {134 ; CHECK-NOBF16-LABEL: name: nnan_ninf_contract_fadd_sequence135 ; CHECK-NOBF16: bb.0.entry:136 ; CHECK-NOBF16-NEXT: liveins: $r0, $r1, $r2137 ; CHECK-NOBF16-NEXT: {{ $}}138 ; CHECK-NOBF16-NEXT: [[COPY:%[0-9]+]]:gpr = COPY $r2139 ; CHECK-NOBF16-NEXT: [[COPY1:%[0-9]+]]:gpr = COPY $r1140 ; CHECK-NOBF16-NEXT: [[COPY2:%[0-9]+]]:gpr = COPY $r0141 ; CHECK-NOBF16-NEXT: [[MOVsi:%[0-9]+]]:gpr = MOVsi [[COPY1]], 130, 14 /* CC::al */, $noreg, $noreg142 ; CHECK-NOBF16-NEXT: [[VMOVSR:%[0-9]+]]:spr = VMOVSR killed [[MOVsi]], 14 /* CC::al */, $noreg143 ; CHECK-NOBF16-NEXT: [[MOVsi1:%[0-9]+]]:gpr = MOVsi [[COPY2]], 130, 14 /* CC::al */, $noreg, $noreg144 ; CHECK-NOBF16-NEXT: [[VMOVSR1:%[0-9]+]]:spr = VMOVSR killed [[MOVsi1]], 14 /* CC::al */, $noreg145 ; CHECK-NOBF16-NEXT: [[VADDS:%[0-9]+]]:spr = nnan ninf contract nofpexcept VADDS killed [[VMOVSR1]], killed [[VMOVSR]], 14 /* CC::al */, $noreg, implicit $fpscr146 ; CHECK-NOBF16-NEXT: [[MOVsi2:%[0-9]+]]:gpr = MOVsi [[COPY]], 130, 14 /* CC::al */, $noreg, $noreg147 ; CHECK-NOBF16-NEXT: [[VMOVSR2:%[0-9]+]]:spr = VMOVSR killed [[MOVsi2]], 14 /* CC::al */, $noreg148 ; CHECK-NOBF16-NEXT: [[VADDS1:%[0-9]+]]:spr = nnan ninf contract nofpexcept VADDS killed [[VADDS]], killed [[VMOVSR2]], 14 /* CC::al */, $noreg, implicit $fpscr149 ; CHECK-NOBF16-NEXT: [[VMOVRS:%[0-9]+]]:gpr = VMOVRS killed [[VADDS1]], 14 /* CC::al */, $noreg150 ; CHECK-NOBF16-NEXT: ADJCALLSTACKDOWN 0, 0, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp151 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRS]]152 ; CHECK-NOBF16-NEXT: BL &__truncsfbf2, csr_aapcs, implicit-def dead $lr, implicit $sp, implicit $r0, implicit-def $sp, implicit-def $r0153 ; CHECK-NOBF16-NEXT: ADJCALLSTACKUP 0, -1, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp154 ; CHECK-NOBF16-NEXT: [[COPY3:%[0-9]+]]:rgpr = COPY $r0155 ; CHECK-NOBF16-NEXT: [[VMOVHR:%[0-9]+]]:hpr = VMOVHR [[COPY3]], 14, $noreg156 ; CHECK-NOBF16-NEXT: [[VMOVRH:%[0-9]+]]:rgpr = VMOVRH killed [[VMOVHR]], 14, $noreg157 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRH]]158 ; CHECK-NOBF16-NEXT: MOVPCLR 14 /* CC::al */, $noreg, implicit $r0159entry:160 %add1 = fadd nnan ninf contract bfloat %x, %y161 %add2 = fadd nnan ninf contract bfloat %add1, %z162 ret bfloat %add2163}164 165define bfloat @ninf_fadd_sequence(bfloat %x, bfloat %y, bfloat %z) {166 ; CHECK-NOBF16-LABEL: name: ninf_fadd_sequence167 ; CHECK-NOBF16: bb.0.entry:168 ; CHECK-NOBF16-NEXT: liveins: $r0, $r1, $r2169 ; CHECK-NOBF16-NEXT: {{ $}}170 ; CHECK-NOBF16-NEXT: [[COPY:%[0-9]+]]:gpr = COPY $r2171 ; CHECK-NOBF16-NEXT: [[COPY1:%[0-9]+]]:gpr = COPY $r1172 ; CHECK-NOBF16-NEXT: [[COPY2:%[0-9]+]]:gpr = COPY $r0173 ; CHECK-NOBF16-NEXT: [[MOVsi:%[0-9]+]]:gpr = MOVsi [[COPY1]], 130, 14 /* CC::al */, $noreg, $noreg174 ; CHECK-NOBF16-NEXT: [[VMOVSR:%[0-9]+]]:spr = VMOVSR killed [[MOVsi]], 14 /* CC::al */, $noreg175 ; CHECK-NOBF16-NEXT: [[MOVsi1:%[0-9]+]]:gpr = MOVsi [[COPY2]], 130, 14 /* CC::al */, $noreg, $noreg176 ; CHECK-NOBF16-NEXT: [[VMOVSR1:%[0-9]+]]:spr = VMOVSR killed [[MOVsi1]], 14 /* CC::al */, $noreg177 ; CHECK-NOBF16-NEXT: [[VADDS:%[0-9]+]]:spr = ninf nofpexcept VADDS killed [[VMOVSR1]], killed [[VMOVSR]], 14 /* CC::al */, $noreg, implicit $fpscr178 ; CHECK-NOBF16-NEXT: [[VMOVRS:%[0-9]+]]:gpr = VMOVRS killed [[VADDS]], 14 /* CC::al */, $noreg179 ; CHECK-NOBF16-NEXT: ADJCALLSTACKDOWN 0, 0, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp180 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRS]]181 ; CHECK-NOBF16-NEXT: BL &__truncsfbf2, csr_aapcs, implicit-def dead $lr, implicit $sp, implicit $r0, implicit-def $sp, implicit-def $r0182 ; CHECK-NOBF16-NEXT: ADJCALLSTACKUP 0, -1, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp183 ; CHECK-NOBF16-NEXT: [[COPY3:%[0-9]+]]:gpr = COPY $r0184 ; CHECK-NOBF16-NEXT: [[MOVsi2:%[0-9]+]]:gpr = MOVsi [[COPY]], 130, 14 /* CC::al */, $noreg, $noreg185 ; CHECK-NOBF16-NEXT: [[VMOVSR2:%[0-9]+]]:spr = VMOVSR killed [[MOVsi2]], 14 /* CC::al */, $noreg186 ; CHECK-NOBF16-NEXT: [[MOVsi3:%[0-9]+]]:gpr = MOVsi [[COPY3]], 130, 14 /* CC::al */, $noreg, $noreg187 ; CHECK-NOBF16-NEXT: [[VMOVSR3:%[0-9]+]]:spr = VMOVSR killed [[MOVsi3]], 14 /* CC::al */, $noreg188 ; CHECK-NOBF16-NEXT: [[VADDS1:%[0-9]+]]:spr = ninf nofpexcept VADDS killed [[VMOVSR3]], killed [[VMOVSR2]], 14 /* CC::al */, $noreg, implicit $fpscr189 ; CHECK-NOBF16-NEXT: [[VMOVRS1:%[0-9]+]]:gpr = VMOVRS killed [[VADDS1]], 14 /* CC::al */, $noreg190 ; CHECK-NOBF16-NEXT: ADJCALLSTACKDOWN 0, 0, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp191 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRS1]]192 ; CHECK-NOBF16-NEXT: BL &__truncsfbf2, csr_aapcs, implicit-def dead $lr, implicit $sp, implicit $r0, implicit-def $sp, implicit-def $r0193 ; CHECK-NOBF16-NEXT: ADJCALLSTACKUP 0, -1, 14 /* CC::al */, $noreg, implicit-def dead $sp, implicit $sp194 ; CHECK-NOBF16-NEXT: [[COPY4:%[0-9]+]]:rgpr = COPY $r0195 ; CHECK-NOBF16-NEXT: [[VMOVHR:%[0-9]+]]:hpr = VMOVHR [[COPY4]], 14, $noreg196 ; CHECK-NOBF16-NEXT: [[VMOVRH:%[0-9]+]]:rgpr = VMOVRH killed [[VMOVHR]], 14, $noreg197 ; CHECK-NOBF16-NEXT: $r0 = COPY [[VMOVRH]]198 ; CHECK-NOBF16-NEXT: MOVPCLR 14 /* CC::al */, $noreg, implicit $r0199entry:200 %add1 = fadd ninf bfloat %x, %y201 %add2 = fadd ninf bfloat %add1, %z202 ret bfloat %add2203}204