141 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py2; RUN: llc -mtriple=amdgcn -mcpu=gfx90a -stop-after=amdgpu-isel -amdgpu-atomic-optimizer-strategy=None < %s | FileCheck -check-prefixes=GFX90A_GFX942,GFX90A %s3; RUN: llc -mtriple=amdgcn -mcpu=gfx942 -stop-after=amdgpu-isel -amdgpu-atomic-optimizer-strategy=None < %s | FileCheck -check-prefixes=GFX90A_GFX942,GFX942 %s4 5define amdgpu_ps void @global_atomic_fadd_f64_no_rtn_atomicrmw(ptr addrspace(1) %ptr, double %data) {6 ; GFX90A_GFX942-LABEL: name: global_atomic_fadd_f64_no_rtn_atomicrmw7 ; GFX90A_GFX942: bb.0 (%ir-block.0):8 ; GFX90A_GFX942-NEXT: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr39 ; GFX90A_GFX942-NEXT: {{ $}}10 ; GFX90A_GFX942-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr311 ; GFX90A_GFX942-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr212 ; GFX90A_GFX942-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr113 ; GFX90A_GFX942-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr014 ; GFX90A_GFX942-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub115 ; GFX90A_GFX942-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub116 ; GFX90A_GFX942-NEXT: [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]17 ; GFX90A_GFX942-NEXT: [[COPY5:%[0-9]+]]:av_64_align2 = COPY [[REG_SEQUENCE]]18 ; GFX90A_GFX942-NEXT: GLOBAL_ATOMIC_ADD_F64 killed [[COPY4]], killed [[COPY5]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)19 ; GFX90A_GFX942-NEXT: S_ENDPGM 020 %ret = atomicrmw fadd ptr addrspace(1) %ptr, double %data syncscope("wavefront") monotonic, !amdgpu.no.fine.grained.memory !021 ret void22}23 24define amdgpu_ps double @global_atomic_fadd_f64_rtn_atomicrmw(ptr addrspace(1) %ptr, double %data) {25 ; GFX90A-LABEL: name: global_atomic_fadd_f64_rtn_atomicrmw26 ; GFX90A: bb.0 (%ir-block.0):27 ; GFX90A-NEXT: successors: %bb.1(0x80000000)28 ; GFX90A-NEXT: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr329 ; GFX90A-NEXT: {{ $}}30 ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr331 ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr232 ; GFX90A-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr133 ; GFX90A-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr034 ; GFX90A-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub135 ; GFX90A-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub136 ; GFX90A-NEXT: [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]37 ; GFX90A-NEXT: [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]38 ; GFX90A-NEXT: [[COPY6:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]39 ; GFX90A-NEXT: [[GLOBAL_LOAD_DWORDX2_:%[0-9]+]]:av_64_align2 = GLOBAL_LOAD_DWORDX2 [[COPY6]], 0, 0, implicit $exec :: (load (s64) from %ir.ptr, addrspace 1)40 ; GFX90A-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64 = S_MOV_B64 041 ; GFX90A-NEXT: {{ $}}42 ; GFX90A-NEXT: bb.1.atomicrmw.start:43 ; GFX90A-NEXT: successors: %bb.2(0x04000000), %bb.1(0x7c000000)44 ; GFX90A-NEXT: {{ $}}45 ; GFX90A-NEXT: [[PHI:%[0-9]+]]:sreg_64 = PHI [[S_MOV_B64_]], %bb.0, %4, %bb.146 ; GFX90A-NEXT: [[PHI1:%[0-9]+]]:vreg_64_align2 = PHI [[GLOBAL_LOAD_DWORDX2_]], %bb.0, %3, %bb.147 ; GFX90A-NEXT: [[V_ADD_F64_e64_:%[0-9]+]]:vreg_64_align2 = nofpexcept V_ADD_F64_e64 0, [[PHI1]], 0, [[COPY4]], 0, 0, implicit $mode, implicit $exec48 ; GFX90A-NEXT: [[COPY7:%[0-9]+]]:av_32 = COPY [[V_ADD_F64_e64_]].sub149 ; GFX90A-NEXT: [[COPY8:%[0-9]+]]:av_32 = COPY [[V_ADD_F64_e64_]].sub050 ; GFX90A-NEXT: [[COPY9:%[0-9]+]]:av_32 = COPY [[PHI1]].sub151 ; GFX90A-NEXT: [[COPY10:%[0-9]+]]:av_32 = COPY [[PHI1]].sub052 ; GFX90A-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_128_align2 = REG_SEQUENCE killed [[COPY8]], %subreg.sub0, killed [[COPY7]], %subreg.sub1, killed [[COPY10]], %subreg.sub2, killed [[COPY9]], %subreg.sub353 ; GFX90A-NEXT: [[GLOBAL_ATOMIC_CMPSWAP_X2_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_CMPSWAP_X2_RTN [[COPY5]], killed [[REG_SEQUENCE2]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic monotonic (s64) on %ir.ptr, addrspace 1)54 ; GFX90A-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64 = V_CMP_EQ_U64_e64 [[GLOBAL_ATOMIC_CMPSWAP_X2_RTN]], [[PHI1]], implicit $exec55 ; GFX90A-NEXT: [[COPY11:%[0-9]+]]:av_64_align2 = COPY [[GLOBAL_ATOMIC_CMPSWAP_X2_RTN]]56 ; GFX90A-NEXT: [[SI_IF_BREAK:%[0-9]+]]:sreg_64 = SI_IF_BREAK killed [[V_CMP_EQ_U64_e64_]], [[PHI]], implicit-def dead $scc57 ; GFX90A-NEXT: SI_LOOP [[SI_IF_BREAK]], %bb.1, implicit-def dead $exec, implicit-def dead $scc, implicit $exec58 ; GFX90A-NEXT: S_BRANCH %bb.259 ; GFX90A-NEXT: {{ $}}60 ; GFX90A-NEXT: bb.2.atomicrmw.end:61 ; GFX90A-NEXT: [[PHI2:%[0-9]+]]:av_64_align2 = PHI [[COPY11]], %bb.162 ; GFX90A-NEXT: [[PHI3:%[0-9]+]]:sreg_64 = PHI [[SI_IF_BREAK]], %bb.163 ; GFX90A-NEXT: SI_END_CF [[PHI3]], implicit-def dead $exec, implicit-def dead $scc, implicit $exec64 ; GFX90A-NEXT: [[COPY12:%[0-9]+]]:vgpr_32 = COPY [[PHI2]].sub065 ; GFX90A-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 killed [[COPY12]], implicit $exec66 ; GFX90A-NEXT: [[COPY13:%[0-9]+]]:vgpr_32 = COPY [[PHI2]].sub167 ; GFX90A-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 killed [[COPY13]], implicit $exec68 ; GFX90A-NEXT: $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]69 ; GFX90A-NEXT: $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]70 ; GFX90A-NEXT: SI_RETURN_TO_EPILOG $sgpr0, $sgpr171 ;72 ; GFX942-LABEL: name: global_atomic_fadd_f64_rtn_atomicrmw73 ; GFX942: bb.0 (%ir-block.0):74 ; GFX942-NEXT: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr375 ; GFX942-NEXT: {{ $}}76 ; GFX942-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr377 ; GFX942-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr278 ; GFX942-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr179 ; GFX942-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr080 ; GFX942-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub181 ; GFX942-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub182 ; GFX942-NEXT: [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]83 ; GFX942-NEXT: [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]84 ; GFX942-NEXT: [[GLOBAL_ATOMIC_ADD_F64_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_RTN killed [[COPY4]], killed [[COPY5]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)85 ; GFX942-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub086 ; GFX942-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 killed [[COPY6]], implicit $exec87 ; GFX942-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub188 ; GFX942-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 killed [[COPY7]], implicit $exec89 ; GFX942-NEXT: $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]90 ; GFX942-NEXT: $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]91 ; GFX942-NEXT: SI_RETURN_TO_EPILOG $sgpr0, $sgpr192 %ret = atomicrmw fadd ptr addrspace(1) %ptr, double %data syncscope("wavefront") monotonic93 ret double %ret94}95 96define amdgpu_ps void @global_atomic_fadd_f64_saddr_no_rtn_atomicrmw(ptr addrspace(1) inreg %ptr, double %data) {97 ; GFX90A_GFX942-LABEL: name: global_atomic_fadd_f64_saddr_no_rtn_atomicrmw98 ; GFX90A_GFX942: bb.0 (%ir-block.0):99 ; GFX90A_GFX942-NEXT: liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1100 ; GFX90A_GFX942-NEXT: {{ $}}101 ; GFX90A_GFX942-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr1102 ; GFX90A_GFX942-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0103 ; GFX90A_GFX942-NEXT: [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr1104 ; GFX90A_GFX942-NEXT: [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr0105 ; GFX90A_GFX942-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1106 ; GFX90A_GFX942-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1107 ; GFX90A_GFX942-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec108 ; GFX90A_GFX942-NEXT: [[COPY4:%[0-9]+]]:av_64_align2 = COPY [[REG_SEQUENCE1]]109 ; GFX90A_GFX942-NEXT: GLOBAL_ATOMIC_ADD_F64_SADDR killed [[V_MOV_B32_e32_]], killed [[COPY4]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)110 ; GFX90A_GFX942-NEXT: S_ENDPGM 0111 %ret = atomicrmw fadd ptr addrspace(1) %ptr, double %data syncscope("wavefront") monotonic, !amdgpu.no.fine.grained.memory !0112 ret void113}114 115define amdgpu_ps double @global_atomic_fadd_f64_saddr_rtn_atomicrmw(ptr addrspace(1) inreg %ptr, double %data) {116 ; GFX90A_GFX942-LABEL: name: global_atomic_fadd_f64_saddr_rtn_atomicrmw117 ; GFX90A_GFX942: bb.0 (%ir-block.0):118 ; GFX90A_GFX942-NEXT: liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1119 ; GFX90A_GFX942-NEXT: {{ $}}120 ; GFX90A_GFX942-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr1121 ; GFX90A_GFX942-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0122 ; GFX90A_GFX942-NEXT: [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr1123 ; GFX90A_GFX942-NEXT: [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr0124 ; GFX90A_GFX942-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1125 ; GFX90A_GFX942-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1126 ; GFX90A_GFX942-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec127 ; GFX90A_GFX942-NEXT: [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]128 ; GFX90A_GFX942-NEXT: [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_SADDR_RTN killed [[V_MOV_B32_e32_]], killed [[COPY4]], killed [[REG_SEQUENCE]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)129 ; GFX90A_GFX942-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub0130 ; GFX90A_GFX942-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 killed [[COPY5]], implicit $exec131 ; GFX90A_GFX942-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub1132 ; GFX90A_GFX942-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 killed [[COPY6]], implicit $exec133 ; GFX90A_GFX942-NEXT: $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]134 ; GFX90A_GFX942-NEXT: $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]135 ; GFX90A_GFX942-NEXT: SI_RETURN_TO_EPILOG $sgpr0, $sgpr1136 %ret = atomicrmw fadd ptr addrspace(1) %ptr, double %data syncscope("wavefront") monotonic, !amdgpu.no.fine.grained.memory !0137 ret double %ret138}139 140!0 = !{}141