494 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py2; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx908 -stop-after=instruction-select -o - %s | FileCheck %s -check-prefix=GFX9083; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx90a -stop-after=instruction-select -o - %s | FileCheck %s -check-prefix=GFX90A4 5; Natural mapping6define amdgpu_ps void @raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(float %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) {7 ; GFX908-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset8 ; GFX908: bb.1 (%ir-block.0):9 ; GFX908-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr110 ; GFX908-NEXT: {{ $}}11 ; GFX908-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr012 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr213 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr314 ; GFX908-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr415 ; GFX908-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr516 ; GFX908-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr117 ; GFX908-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr618 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub319 ; GFX908-NEXT: BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)20 ; GFX908-NEXT: S_ENDPGM 021 ;22 ; GFX90A-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset23 ; GFX90A: bb.1 (%ir-block.0):24 ; GFX90A-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr125 ; GFX90A-NEXT: {{ $}}26 ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr027 ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr228 ; GFX90A-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr329 ; GFX90A-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr430 ; GFX90A-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr531 ; GFX90A-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr132 ; GFX90A-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr633 ; GFX90A-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub334 ; GFX90A-NEXT: BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)35 ; GFX90A-NEXT: S_ENDPGM 036 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fadd.f32(float %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0)37 ret void38}39 40define amdgpu_ps void @raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset_plus4095__sgpr_soffset(float %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) {41 ; GFX908-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset_plus4095__sgpr_soffset42 ; GFX908: bb.1 (%ir-block.0):43 ; GFX908-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr144 ; GFX908-NEXT: {{ $}}45 ; GFX908-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr046 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr247 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr348 ; GFX908-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr449 ; GFX908-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr550 ; GFX908-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr151 ; GFX908-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr652 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub353 ; GFX908-NEXT: BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 4095, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)54 ; GFX908-NEXT: S_ENDPGM 055 ;56 ; GFX90A-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset_plus4095__sgpr_soffset57 ; GFX90A: bb.1 (%ir-block.0):58 ; GFX90A-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr159 ; GFX90A-NEXT: {{ $}}60 ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr061 ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr262 ; GFX90A-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr363 ; GFX90A-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr464 ; GFX90A-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr565 ; GFX90A-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr166 ; GFX90A-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr667 ; GFX90A-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub368 ; GFX90A-NEXT: BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 4095, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)69 ; GFX90A-NEXT: S_ENDPGM 070 %voffset.add = add i32 %voffset, 409571 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fadd.f32(float %val, ptr addrspace(8) %rsrc, i32 %voffset.add, i32 %soffset, i32 0)72 ret void73}74 75define amdgpu_ps void @raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset_4095__sgpr_soffset(float %val, ptr addrspace(8) inreg %rsrc, i32 inreg %soffset) {76 ; GFX908-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset_4095__sgpr_soffset77 ; GFX908: bb.1 (%ir-block.0):78 ; GFX908-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr079 ; GFX908-NEXT: {{ $}}80 ; GFX908-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr081 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr282 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr383 ; GFX908-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr484 ; GFX908-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr585 ; GFX908-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr686 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub387 ; GFX908-NEXT: BUFFER_ATOMIC_ADD_F32_OFFSET [[COPY]], [[REG_SEQUENCE]], [[COPY5]], 4095, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)88 ; GFX908-NEXT: S_ENDPGM 089 ;90 ; GFX90A-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset_4095__sgpr_soffset91 ; GFX90A: bb.1 (%ir-block.0):92 ; GFX90A-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr093 ; GFX90A-NEXT: {{ $}}94 ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr095 ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr296 ; GFX90A-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr397 ; GFX90A-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr498 ; GFX90A-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr599 ; GFX90A-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6100 ; GFX90A-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3101 ; GFX90A-NEXT: BUFFER_ATOMIC_ADD_F32_OFFSET [[COPY]], [[REG_SEQUENCE]], [[COPY5]], 4095, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)102 ; GFX90A-NEXT: S_ENDPGM 0103 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fadd.f32(float %val, ptr addrspace(8) %rsrc, i32 4095, i32 %soffset, i32 0)104 ret void105}106 107; Natural mapping, no voffset108define amdgpu_ps void @raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__0_voffset__sgpr_soffset(float %val, ptr addrspace(8) inreg %rsrc, i32 inreg %soffset) {109 ; GFX908-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__0_voffset__sgpr_soffset110 ; GFX908: bb.1 (%ir-block.0):111 ; GFX908-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0112 ; GFX908-NEXT: {{ $}}113 ; GFX908-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0114 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2115 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3116 ; GFX908-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4117 ; GFX908-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5118 ; GFX908-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6119 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3120 ; GFX908-NEXT: BUFFER_ATOMIC_ADD_F32_OFFSET [[COPY]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)121 ; GFX908-NEXT: S_ENDPGM 0122 ;123 ; GFX90A-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__0_voffset__sgpr_soffset124 ; GFX90A: bb.1 (%ir-block.0):125 ; GFX90A-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0126 ; GFX90A-NEXT: {{ $}}127 ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0128 ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2129 ; GFX90A-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3130 ; GFX90A-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4131 ; GFX90A-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5132 ; GFX90A-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6133 ; GFX90A-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3134 ; GFX90A-NEXT: BUFFER_ATOMIC_ADD_F32_OFFSET [[COPY]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)135 ; GFX90A-NEXT: S_ENDPGM 0136 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fadd.f32(float %val, ptr addrspace(8) %rsrc, i32 0, i32 %soffset, i32 0)137 ret void138}139 140; All operands need regbank legalization141define amdgpu_ps void @raw_ptr_buffer_atomic_add_f32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset(float inreg %val, ptr addrspace(8) %rsrc, i32 inreg %voffset, i32 %soffset) {142 ; GFX908-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset143 ; GFX908: bb.1 (%ir-block.0):144 ; GFX908-NEXT: successors: %bb.2(0x80000000)145 ; GFX908-NEXT: liveins: $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4146 ; GFX908-NEXT: {{ $}}147 ; GFX908-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2148 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0149 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1150 ; GFX908-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2151 ; GFX908-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3152 ; GFX908-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3153 ; GFX908-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr4154 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3155 ; GFX908-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY]]156 ; GFX908-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY5]]157 ; GFX908-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec158 ; GFX908-NEXT: {{ $}}159 ; GFX908-NEXT: bb.2:160 ; GFX908-NEXT: successors: %bb.3(0x80000000)161 ; GFX908-NEXT: {{ $}}162 ; GFX908-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec163 ; GFX908-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec164 ; GFX908-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec165 ; GFX908-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec166 ; GFX908-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3167 ; GFX908-NEXT: [[COPY9:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1168 ; GFX908-NEXT: [[COPY10:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3169 ; GFX908-NEXT: [[COPY11:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1170 ; GFX908-NEXT: [[COPY12:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3171 ; GFX908-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY11]], [[COPY9]], implicit $exec172 ; GFX908-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY12]], [[COPY10]], implicit $exec173 ; GFX908-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc174 ; GFX908-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]], implicit $exec175 ; GFX908-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY6]], implicit $exec176 ; GFX908-NEXT: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc177 ; GFX908-NEXT: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_1]], implicit-def $exec, implicit-def $scc, implicit $exec178 ; GFX908-NEXT: {{ $}}179 ; GFX908-NEXT: bb.3:180 ; GFX908-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000)181 ; GFX908-NEXT: {{ $}}182 ; GFX908-NEXT: BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY7]], [[COPY8]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)183 ; GFX908-NEXT: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc184 ; GFX908-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec185 ; GFX908-NEXT: {{ $}}186 ; GFX908-NEXT: bb.4:187 ; GFX908-NEXT: successors: %bb.5(0x80000000)188 ; GFX908-NEXT: {{ $}}189 ; GFX908-NEXT: $exec = S_MOV_B64_term [[S_MOV_B64_]]190 ; GFX908-NEXT: {{ $}}191 ; GFX908-NEXT: bb.5:192 ; GFX908-NEXT: S_ENDPGM 0193 ;194 ; GFX90A-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__sgpr_val__vgpr_rsrc__sgpr_voffset__vgpr_soffset195 ; GFX90A: bb.1 (%ir-block.0):196 ; GFX90A-NEXT: successors: %bb.2(0x80000000)197 ; GFX90A-NEXT: liveins: $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4198 ; GFX90A-NEXT: {{ $}}199 ; GFX90A-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2200 ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0201 ; GFX90A-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1202 ; GFX90A-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2203 ; GFX90A-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3204 ; GFX90A-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3205 ; GFX90A-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr4206 ; GFX90A-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128_align2 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3207 ; GFX90A-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY]]208 ; GFX90A-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY5]]209 ; GFX90A-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec210 ; GFX90A-NEXT: {{ $}}211 ; GFX90A-NEXT: bb.2:212 ; GFX90A-NEXT: successors: %bb.3(0x80000000)213 ; GFX90A-NEXT: {{ $}}214 ; GFX90A-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec215 ; GFX90A-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec216 ; GFX90A-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec217 ; GFX90A-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec218 ; GFX90A-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3219 ; GFX90A-NEXT: [[COPY9:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]].sub0_sub1220 ; GFX90A-NEXT: [[COPY10:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]].sub2_sub3221 ; GFX90A-NEXT: [[COPY11:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1222 ; GFX90A-NEXT: [[COPY12:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3223 ; GFX90A-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY11]], [[COPY9]], implicit $exec224 ; GFX90A-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY12]], [[COPY10]], implicit $exec225 ; GFX90A-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc226 ; GFX90A-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY6]], implicit $exec227 ; GFX90A-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY6]], implicit $exec228 ; GFX90A-NEXT: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc229 ; GFX90A-NEXT: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_1]], implicit-def $exec, implicit-def $scc, implicit $exec230 ; GFX90A-NEXT: {{ $}}231 ; GFX90A-NEXT: bb.3:232 ; GFX90A-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000)233 ; GFX90A-NEXT: {{ $}}234 ; GFX90A-NEXT: BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY7]], [[COPY8]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)235 ; GFX90A-NEXT: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc236 ; GFX90A-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec237 ; GFX90A-NEXT: {{ $}}238 ; GFX90A-NEXT: bb.4:239 ; GFX90A-NEXT: successors: %bb.5(0x80000000)240 ; GFX90A-NEXT: {{ $}}241 ; GFX90A-NEXT: $exec = S_MOV_B64_term [[S_MOV_B64_]]242 ; GFX90A-NEXT: {{ $}}243 ; GFX90A-NEXT: bb.5:244 ; GFX90A-NEXT: S_ENDPGM 0245 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fadd.f32(float %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0)246 ret void247}248 249; All operands need regbank legalization, no voffset250define amdgpu_ps void @raw_ptr_buffer_atomic_add_f32_noret__sgpr_val__vgpr_rsrc__0_voffset__vgpr_soffset(float inreg %val, ptr addrspace(8) %rsrc, i32 %soffset) {251 ; GFX908-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__sgpr_val__vgpr_rsrc__0_voffset__vgpr_soffset252 ; GFX908: bb.1 (%ir-block.0):253 ; GFX908-NEXT: successors: %bb.2(0x80000000)254 ; GFX908-NEXT: liveins: $sgpr2, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4255 ; GFX908-NEXT: {{ $}}256 ; GFX908-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2257 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0258 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1259 ; GFX908-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2260 ; GFX908-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3261 ; GFX908-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr4262 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3263 ; GFX908-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY]]264 ; GFX908-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec265 ; GFX908-NEXT: {{ $}}266 ; GFX908-NEXT: bb.2:267 ; GFX908-NEXT: successors: %bb.3(0x80000000)268 ; GFX908-NEXT: {{ $}}269 ; GFX908-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec270 ; GFX908-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec271 ; GFX908-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec272 ; GFX908-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec273 ; GFX908-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3274 ; GFX908-NEXT: [[COPY7:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1275 ; GFX908-NEXT: [[COPY8:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3276 ; GFX908-NEXT: [[COPY9:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1277 ; GFX908-NEXT: [[COPY10:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3278 ; GFX908-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY9]], [[COPY7]], implicit $exec279 ; GFX908-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY10]], [[COPY8]], implicit $exec280 ; GFX908-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc281 ; GFX908-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY5]], implicit $exec282 ; GFX908-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY5]], implicit $exec283 ; GFX908-NEXT: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc284 ; GFX908-NEXT: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_1]], implicit-def $exec, implicit-def $scc, implicit $exec285 ; GFX908-NEXT: {{ $}}286 ; GFX908-NEXT: bb.3:287 ; GFX908-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000)288 ; GFX908-NEXT: {{ $}}289 ; GFX908-NEXT: BUFFER_ATOMIC_ADD_F32_OFFSET [[COPY6]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)290 ; GFX908-NEXT: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc291 ; GFX908-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec292 ; GFX908-NEXT: {{ $}}293 ; GFX908-NEXT: bb.4:294 ; GFX908-NEXT: successors: %bb.5(0x80000000)295 ; GFX908-NEXT: {{ $}}296 ; GFX908-NEXT: $exec = S_MOV_B64_term [[S_MOV_B64_]]297 ; GFX908-NEXT: {{ $}}298 ; GFX908-NEXT: bb.5:299 ; GFX908-NEXT: S_ENDPGM 0300 ;301 ; GFX90A-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__sgpr_val__vgpr_rsrc__0_voffset__vgpr_soffset302 ; GFX90A: bb.1 (%ir-block.0):303 ; GFX90A-NEXT: successors: %bb.2(0x80000000)304 ; GFX90A-NEXT: liveins: $sgpr2, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4305 ; GFX90A-NEXT: {{ $}}306 ; GFX90A-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr2307 ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0308 ; GFX90A-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1309 ; GFX90A-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2310 ; GFX90A-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr3311 ; GFX90A-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr4312 ; GFX90A-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128_align2 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3313 ; GFX90A-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY]]314 ; GFX90A-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec315 ; GFX90A-NEXT: {{ $}}316 ; GFX90A-NEXT: bb.2:317 ; GFX90A-NEXT: successors: %bb.3(0x80000000)318 ; GFX90A-NEXT: {{ $}}319 ; GFX90A-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY1]], implicit $exec320 ; GFX90A-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY2]], implicit $exec321 ; GFX90A-NEXT: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY3]], implicit $exec322 ; GFX90A-NEXT: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec323 ; GFX90A-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3324 ; GFX90A-NEXT: [[COPY7:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]].sub0_sub1325 ; GFX90A-NEXT: [[COPY8:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]].sub2_sub3326 ; GFX90A-NEXT: [[COPY9:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub0_sub1327 ; GFX90A-NEXT: [[COPY10:%[0-9]+]]:sreg_64 = COPY [[REG_SEQUENCE1]].sub2_sub3328 ; GFX90A-NEXT: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY9]], [[COPY7]], implicit $exec329 ; GFX90A-NEXT: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[COPY10]], [[COPY8]], implicit $exec330 ; GFX90A-NEXT: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def dead $scc331 ; GFX90A-NEXT: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY5]], implicit $exec332 ; GFX90A-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY5]], implicit $exec333 ; GFX90A-NEXT: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def dead $scc334 ; GFX90A-NEXT: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_1]], implicit-def $exec, implicit-def $scc, implicit $exec335 ; GFX90A-NEXT: {{ $}}336 ; GFX90A-NEXT: bb.3:337 ; GFX90A-NEXT: successors: %bb.4(0x40000000), %bb.2(0x40000000)338 ; GFX90A-NEXT: {{ $}}339 ; GFX90A-NEXT: BUFFER_ATOMIC_ADD_F32_OFFSET [[COPY6]], [[REG_SEQUENCE1]], [[V_READFIRSTLANE_B32_4]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)340 ; GFX90A-NEXT: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc341 ; GFX90A-NEXT: SI_WATERFALL_LOOP %bb.2, implicit $exec342 ; GFX90A-NEXT: {{ $}}343 ; GFX90A-NEXT: bb.4:344 ; GFX90A-NEXT: successors: %bb.5(0x80000000)345 ; GFX90A-NEXT: {{ $}}346 ; GFX90A-NEXT: $exec = S_MOV_B64_term [[S_MOV_B64_]]347 ; GFX90A-NEXT: {{ $}}348 ; GFX90A-NEXT: bb.5:349 ; GFX90A-NEXT: S_ENDPGM 0350 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fadd.f32(float %val, ptr addrspace(8) %rsrc, i32 0, i32 %soffset, i32 0)351 ret void352}353 354define amdgpu_ps void @raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add4095(float %val, ptr addrspace(8) inreg %rsrc, i32 %voffset.base, i32 inreg %soffset) {355 ; GFX908-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add4095356 ; GFX908: bb.1 (%ir-block.0):357 ; GFX908-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1358 ; GFX908-NEXT: {{ $}}359 ; GFX908-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0360 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2361 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3362 ; GFX908-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4363 ; GFX908-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5364 ; GFX908-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1365 ; GFX908-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6366 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3367 ; GFX908-NEXT: BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 4095, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)368 ; GFX908-NEXT: S_ENDPGM 0369 ;370 ; GFX90A-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add4095371 ; GFX90A: bb.1 (%ir-block.0):372 ; GFX90A-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1373 ; GFX90A-NEXT: {{ $}}374 ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0375 ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2376 ; GFX90A-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3377 ; GFX90A-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4378 ; GFX90A-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5379 ; GFX90A-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1380 ; GFX90A-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6381 ; GFX90A-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3382 ; GFX90A-NEXT: BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 4095, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)383 ; GFX90A-NEXT: S_ENDPGM 0384 %voffset = add i32 %voffset.base, 4095385 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fadd.f32(float %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0)386 ret void387}388 389; Natural mapping + slc390define amdgpu_ps void @raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc(float %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) {391 ; GFX908-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc392 ; GFX908: bb.1 (%ir-block.0):393 ; GFX908-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1394 ; GFX908-NEXT: {{ $}}395 ; GFX908-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0396 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2397 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3398 ; GFX908-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4399 ; GFX908-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5400 ; GFX908-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1401 ; GFX908-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6402 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3403 ; GFX908-NEXT: BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 2, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)404 ; GFX908-NEXT: S_ENDPGM 0405 ;406 ; GFX90A-LABEL: name: raw_ptr_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc407 ; GFX90A: bb.1 (%ir-block.0):408 ; GFX90A-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1409 ; GFX90A-NEXT: {{ $}}410 ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0411 ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2412 ; GFX90A-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3413 ; GFX90A-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4414 ; GFX90A-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5415 ; GFX90A-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1416 ; GFX90A-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6417 ; GFX90A-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3418 ; GFX90A-NEXT: BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 2, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.rsrc, align 1, addrspace 8)419 ; GFX90A-NEXT: S_ENDPGM 0420 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fadd.f32(float %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 2)421 ret void422}423 424define amdgpu_ps void @raw_ptr_buffer_atomic_add_v2f16_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset(<2 x half> %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) {425 ; GFX908-LABEL: name: raw_ptr_buffer_atomic_add_v2f16_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset426 ; GFX908: bb.1 (%ir-block.0):427 ; GFX908-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1428 ; GFX908-NEXT: {{ $}}429 ; GFX908-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0430 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2431 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3432 ; GFX908-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4433 ; GFX908-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5434 ; GFX908-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1435 ; GFX908-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6436 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3437 ; GFX908-NEXT: BUFFER_ATOMIC_PK_ADD_F16_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.rsrc, align 1, addrspace 8)438 ; GFX908-NEXT: S_ENDPGM 0439 ;440 ; GFX90A-LABEL: name: raw_ptr_buffer_atomic_add_v2f16_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset441 ; GFX90A: bb.1 (%ir-block.0):442 ; GFX90A-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1443 ; GFX90A-NEXT: {{ $}}444 ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0445 ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2446 ; GFX90A-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3447 ; GFX90A-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4448 ; GFX90A-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5449 ; GFX90A-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1450 ; GFX90A-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6451 ; GFX90A-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3452 ; GFX90A-NEXT: BUFFER_ATOMIC_PK_ADD_F16_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.rsrc, align 1, addrspace 8)453 ; GFX90A-NEXT: S_ENDPGM 0454 %ret = call <2 x half> @llvm.amdgcn.raw.ptr.buffer.atomic.fadd.v2f16(<2 x half> %val, ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0)455 ret void456}457 458define amdgpu_ps void @raw_ptr_buffer_atomic_add_v2f16_noret__vgpr_val__sgpr_rsrc__0_voffset__sgpr_soffset(<2 x half> %val, ptr addrspace(8) inreg %rsrc, i32 %voffset, i32 inreg %soffset) {459 ; GFX908-LABEL: name: raw_ptr_buffer_atomic_add_v2f16_noret__vgpr_val__sgpr_rsrc__0_voffset__sgpr_soffset460 ; GFX908: bb.1 (%ir-block.0):461 ; GFX908-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0462 ; GFX908-NEXT: {{ $}}463 ; GFX908-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0464 ; GFX908-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2465 ; GFX908-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3466 ; GFX908-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4467 ; GFX908-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5468 ; GFX908-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6469 ; GFX908-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3470 ; GFX908-NEXT: BUFFER_ATOMIC_PK_ADD_F16_OFFSET [[COPY]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.rsrc, align 1, addrspace 8)471 ; GFX908-NEXT: S_ENDPGM 0472 ;473 ; GFX90A-LABEL: name: raw_ptr_buffer_atomic_add_v2f16_noret__vgpr_val__sgpr_rsrc__0_voffset__sgpr_soffset474 ; GFX90A: bb.1 (%ir-block.0):475 ; GFX90A-NEXT: liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0476 ; GFX90A-NEXT: {{ $}}477 ; GFX90A-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0478 ; GFX90A-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr2479 ; GFX90A-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr3480 ; GFX90A-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr4481 ; GFX90A-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr5482 ; GFX90A-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr6483 ; GFX90A-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3484 ; GFX90A-NEXT: BUFFER_ATOMIC_PK_ADD_F16_OFFSET [[COPY]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.rsrc, align 1, addrspace 8)485 ; GFX90A-NEXT: S_ENDPGM 0486 %ret = call <2 x half> @llvm.amdgcn.raw.ptr.buffer.atomic.fadd.v2f16(<2 x half> %val, ptr addrspace(8) %rsrc, i32 0, i32 %soffset, i32 0)487 ret void488}489 490declare float @llvm.amdgcn.raw.ptr.buffer.atomic.fadd.f32(float, ptr addrspace(8), i32, i32, i32 immarg) #0491declare <2 x half> @llvm.amdgcn.raw.ptr.buffer.atomic.fadd.v2f16(<2 x half>, ptr addrspace(8), i32, i32, i32 immarg) #0492 493attributes #0 = { nounwind }494