1194 lines · plain
1//=- AArch64SMEInstrInfo.td - AArch64 SME Instructions -*- tablegen -*-----=//2//3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.4// See https://llvm.org/LICENSE.txt for license information.5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception6//7//===----------------------------------------------------------------------===//8//9// AArch64 Scalable Matrix Extension (SME) Instruction definitions.10//11//===----------------------------------------------------------------------===//12 13def AArch64_smstart : SDNode<"AArch64ISD::SMSTART", SDTypeProfile<0, 1,14 [SDTCisInt<0>]>,15 [SDNPHasChain, SDNPSideEffect, SDNPVariadic,16 SDNPOptInGlue, SDNPOutGlue]>;17def AArch64_smstop : SDNode<"AArch64ISD::SMSTOP", SDTypeProfile<0, 1,18 [SDTCisInt<0>]>,19 [SDNPHasChain, SDNPSideEffect, SDNPVariadic,20 SDNPOptInGlue, SDNPOutGlue]>;21def AArch64_cond_smstart : SDNode<"AArch64ISD::COND_SMSTART", SDTypeProfile<0, 3,22 [SDTCisInt<0>, SDTCisInt<1>, SDTCisInt<2>]>,23 [SDNPHasChain, SDNPSideEffect, SDNPVariadic,24 SDNPOptInGlue, SDNPOutGlue]>;25def AArch64_cond_smstop : SDNode<"AArch64ISD::COND_SMSTOP", SDTypeProfile<0, 3,26 [SDTCisInt<0>, SDTCisInt<1>, SDTCisInt<2>]>,27 [SDNPHasChain, SDNPSideEffect, SDNPVariadic,28 SDNPOptInGlue, SDNPOutGlue]>;29def AArch64_restore_za : SDNode<"AArch64ISD::RESTORE_ZA", SDTypeProfile<0, 3,30 [SDTCisInt<0>, SDTCisPtrTy<1>]>,31 [SDNPHasChain, SDNPSideEffect, SDNPVariadic,32 SDNPOptInGlue]>;33def AArch64_restore_zt : SDNode<"AArch64ISD::RESTORE_ZT", SDTypeProfile<0, 2,34 [SDTCisInt<0>, SDTCisPtrTy<1>]>,35 [SDNPHasChain, SDNPSideEffect, SDNPMayLoad]>;36def AArch64_save_zt : SDNode<"AArch64ISD::SAVE_ZT", SDTypeProfile<0, 2,37 [SDTCisInt<0>, SDTCisPtrTy<1>]>,38 [SDNPHasChain, SDNPSideEffect, SDNPMayStore]>;39def AArch64CoalescerBarrier40 : SDNode<"AArch64ISD::COALESCER_BARRIER", SDTypeProfile<1, 1, []>, [SDNPOptInGlue, SDNPOutGlue]>;41 42def AArch64EntryPStateSM43 : SDNode<"AArch64ISD::ENTRY_PSTATE_SM", SDTypeProfile<1, 0,44 [SDTCisInt<0>]>, [SDNPHasChain, SDNPSideEffect]>;45 46let usesCustomInserter = 1 in {47 def EntryPStateSM : Pseudo<(outs GPR64:$is_streaming), (ins), []>, Sched<[]> {}48}49def : Pat<(i64 (AArch64EntryPStateSM)), (EntryPStateSM)>;50 51// Pseudo-instruction that compares the current SVE vector length (VL) with the52// streaming vector length (SVL). If the two lengths do not match, the check53// lowers to a `brk`, causing a trap.54let hasSideEffects = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in55def CHECK_MATCHING_VL_PSEUDO : Pseudo<(outs), (ins), []>, Sched<[]>;56 57def AArch64_check_matching_vl58 : SDNode<"AArch64ISD::CHECK_MATCHING_VL", SDTypeProfile<0, 0,[]>,59 [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;60def : Pat<(AArch64_check_matching_vl), (CHECK_MATCHING_VL_PSEUDO)>;61 62//===----------------------------------------------------------------------===//63// Old SME ABI lowering ISD nodes/pseudos (deprecated)64//===----------------------------------------------------------------------===//65 66def AArch64AllocateZABuffer : SDNode<"AArch64ISD::ALLOCATE_ZA_BUFFER", SDTypeProfile<1, 1,67 [SDTCisInt<0>, SDTCisInt<1>]>,68 [SDNPHasChain, SDNPSideEffect]>;69let usesCustomInserter = 1, Defs = [SP], Uses = [SP] in {70 def AllocateZABuffer : Pseudo<(outs GPR64sp:$dst), (ins GPR64:$size), []>, Sched<[WriteI]> {}71}72def : Pat<(i64 (AArch64AllocateZABuffer GPR64:$size)),73 (AllocateZABuffer $size)>;74 75def AArch64InitTPIDR2Obj : SDNode<"AArch64ISD::INIT_TPIDR2OBJ", SDTypeProfile<0, 2,76 [SDTCisInt<0>, SDTCisInt<1>]>, [SDNPHasChain, SDNPMayStore]>;77let usesCustomInserter = 1 in {78 def InitTPIDR2Obj : Pseudo<(outs), (ins GPR64:$buffer, GPR64:$save_slices), [(AArch64InitTPIDR2Obj GPR64:$buffer, GPR64:$save_slices)]>, Sched<[WriteI]> {}79}80 81// Nodes to allocate a save buffer for SME.82// Needed for __arm_agnostic("sme_za_state").83def AArch64SMESaveSize : SDNode<"AArch64ISD::GET_SME_SAVE_SIZE", SDTypeProfile<1, 0,84 [SDTCisInt<0>]>, [SDNPHasChain]>;85let usesCustomInserter = 1, Defs = [X0] in {86 def GetSMESaveSize : Pseudo<(outs GPR64:$dst), (ins), []>, Sched<[]> {}87}88def : Pat<(i64 AArch64SMESaveSize), (GetSMESaveSize)>;89 90def AArch64AllocateSMESaveBuffer : SDNode<"AArch64ISD::ALLOC_SME_SAVE_BUFFER", SDTypeProfile<1, 1,91 [SDTCisInt<0>, SDTCisInt<1>]>, [SDNPHasChain]>;92let usesCustomInserter = 1, Defs = [SP] in {93 def AllocateSMESaveBuffer : Pseudo<(outs GPR64sp:$dst), (ins GPR64:$size), []>, Sched<[WriteI]> {}94}95def : Pat<(i64 (AArch64AllocateSMESaveBuffer GPR64:$size)),96 (AllocateSMESaveBuffer $size)>;97 98//===----------------------------------------------------------------------===//99// New SME ABI lowering ISD nodes/pseudos (-aarch64-new-sme-abi)100//===----------------------------------------------------------------------===//101 102let hasSideEffects = 1, isMeta = 1 in {103 def InOutZAUsePseudo : Pseudo<(outs), (ins), []>, Sched<[]>;104 def RequiresZASavePseudo : Pseudo<(outs), (ins), []>, Sched<[]>;105 def RequiresZT0SavePseudo : Pseudo<(outs), (ins), []>, Sched<[]>;106}107 108def SMEStateAllocPseudo : Pseudo<(outs), (ins), []>, Sched<[]>;109 110def CommitZASavePseudo111 : Pseudo<(outs),112 (ins GPR64:$tpidr2_el0, i1imm:$zero_za, i1imm:$zero_zt0,113 i64imm:$commit_routine, variable_ops), []>,114 Sched<[]>;115 116def AArch64_inout_za_use117 : SDNode<"AArch64ISD::INOUT_ZA_USE", SDTypeProfile<0, 0,[]>,118 [SDNPHasChain, SDNPInGlue, SDNPOutGlue]>;119def : Pat<(AArch64_inout_za_use), (InOutZAUsePseudo)>;120 121def AArch64_requires_za_save122 : SDNode<"AArch64ISD::REQUIRES_ZA_SAVE", SDTypeProfile<0, 0,[]>,123 [SDNPHasChain, SDNPInGlue, SDNPOutGlue]>;124def : Pat<(AArch64_requires_za_save), (RequiresZASavePseudo)>;125 126def AArch64_requires_zt0_save127 : SDNode<"AArch64ISD::REQUIRES_ZT0_SAVE", SDTypeProfile<0, 0, []>,128 [SDNPHasChain, SDNPInGlue, SDNPOutGlue]>;129def : Pat<(AArch64_requires_zt0_save), (RequiresZT0SavePseudo)>;130 131def AArch64_sme_state_alloc132 : SDNode<"AArch64ISD::SME_STATE_ALLOC", SDTypeProfile<0, 0,[]>,133 [SDNPHasChain]>;134def : Pat<(AArch64_sme_state_alloc), (SMEStateAllocPseudo)>;135 136//===----------------------------------------------------------------------===//137// Instruction naming conventions.138//===----------------------------------------------------------------------===//139 140// M = SME array register (ZA)141// P = Predicate register142// C = Predicate-as-counter register143// I = immediate144// Z = SVE vector register145// T = ZT0 register146//147 148//===----------------------------------------------------------------------===//149// Add vector elements horizontally or vertically to ZA tile.150//===----------------------------------------------------------------------===//151 152def SDT_AArch64RDSVL : SDTypeProfile<1, 1, [SDTCisInt<0>, SDTCisInt<1>]>;153def AArch64rdsvl : SDNode<"AArch64ISD::RDSVL", SDT_AArch64RDSVL>;154 155def sme_rdsvl_shl_imm : ComplexPattern<i64, 1, "SelectRDSVLShiftImm<1, 31>">;156 157let Predicates = [HasSMEandIsNonStreamingSafe] in {158def RDSVLI_XI : sve_int_read_vl_a<0b0, 0b11111, "rdsvl", /*streaming_sve=*/0b1>;159def ADDSPL_XXI : sve_int_arith_vl<0b1, "addspl", /*streaming_sve=*/0b1>;160def ADDSVL_XXI : sve_int_arith_vl<0b0, "addsvl", /*streaming_sve=*/0b1>;161 162def : Pat<(i64 (shl (AArch64rdsvl (i32 1)), (sme_rdsvl_shl_imm i64:$imm))),163 (RDSVLI_XI (!cast<SDNodeXForm>("trunc_imm") $imm))>;164 165def : Pat<(AArch64rdsvl (i32 simm6_32b:$imm)), (RDSVLI_XI simm6_32b:$imm)>;166}167 168let Predicates = [HasSME] in {169defm ADDHA_MPPZ_S : sme_add_vector_to_tile_u32<0b0, "addha", int_aarch64_sme_addha>;170defm ADDVA_MPPZ_S : sme_add_vector_to_tile_u32<0b1, "addva", int_aarch64_sme_addva>;171}172 173let Predicates = [HasSMEI16I64] in {174defm ADDHA_MPPZ_D : sme_add_vector_to_tile_u64<0b0, "addha", int_aarch64_sme_addha>;175defm ADDVA_MPPZ_D : sme_add_vector_to_tile_u64<0b1, "addva", int_aarch64_sme_addva>;176}177 178let Predicates = [HasSME] in {179//===----------------------------------------------------------------------===//180// Outer products181//===----------------------------------------------------------------------===//182 183defm BFMOPA_MPPZZ : sme_bf16_outer_product<0b000, "bfmopa", int_aarch64_sme_mopa_wide>;184defm BFMOPS_MPPZZ : sme_bf16_outer_product<0b001, "bfmops", int_aarch64_sme_mops_wide>;185 186defm FMOPA_MPPZZ_S : sme_outer_product_fp32<0b0, 0b00, ZPR32, "fmopa", int_aarch64_sme_mopa>;187defm FMOPS_MPPZZ_S : sme_outer_product_fp32<0b1, 0b00, ZPR32, "fmops", int_aarch64_sme_mops>;188}189 190let Predicates = [HasSMEF64F64] in {191defm FMOPA_MPPZZ_D : sme_outer_product_fp64<0b0, "fmopa", int_aarch64_sme_mopa>;192defm FMOPS_MPPZZ_D : sme_outer_product_fp64<0b1, "fmops", int_aarch64_sme_mops>;193}194 195let Predicates = [HasSME] in {196defm FMOPAL_MPPZZ : sme_f16_outer_product<0b010, "fmopa", int_aarch64_sme_mopa_wide>;197defm FMOPSL_MPPZZ : sme_f16_outer_product<0b011, "fmops", int_aarch64_sme_mops_wide>;198 199defm SMOPA_MPPZZ_S : sme_int_outer_product_i32<0b000, "smopa", int_aarch64_sme_smopa_wide>;200defm SMOPS_MPPZZ_S : sme_int_outer_product_i32<0b001, "smops", int_aarch64_sme_smops_wide>;201defm UMOPA_MPPZZ_S : sme_int_outer_product_i32<0b110, "umopa", int_aarch64_sme_umopa_wide>;202defm UMOPS_MPPZZ_S : sme_int_outer_product_i32<0b111, "umops", int_aarch64_sme_umops_wide>;203defm SUMOPA_MPPZZ_S : sme_int_outer_product_i32<0b010, "sumopa", int_aarch64_sme_sumopa_wide>;204defm SUMOPS_MPPZZ_S : sme_int_outer_product_i32<0b011, "sumops", int_aarch64_sme_sumops_wide>;205defm USMOPA_MPPZZ_S : sme_int_outer_product_i32<0b100, "usmopa", int_aarch64_sme_usmopa_wide>;206defm USMOPS_MPPZZ_S : sme_int_outer_product_i32<0b101, "usmops", int_aarch64_sme_usmops_wide>;207}208 209let Predicates = [HasSMEI16I64] in {210defm SMOPA_MPPZZ_D : sme_int_outer_product_i64<0b000, "smopa", int_aarch64_sme_smopa_wide>;211defm SMOPS_MPPZZ_D : sme_int_outer_product_i64<0b001, "smops", int_aarch64_sme_smops_wide>;212defm UMOPA_MPPZZ_D : sme_int_outer_product_i64<0b110, "umopa", int_aarch64_sme_umopa_wide>;213defm UMOPS_MPPZZ_D : sme_int_outer_product_i64<0b111, "umops", int_aarch64_sme_umops_wide>;214defm SUMOPA_MPPZZ_D : sme_int_outer_product_i64<0b010, "sumopa", int_aarch64_sme_sumopa_wide>;215defm SUMOPS_MPPZZ_D : sme_int_outer_product_i64<0b011, "sumops", int_aarch64_sme_sumops_wide>;216defm USMOPA_MPPZZ_D : sme_int_outer_product_i64<0b100, "usmopa", int_aarch64_sme_usmopa_wide>;217defm USMOPS_MPPZZ_D : sme_int_outer_product_i64<0b101, "usmops", int_aarch64_sme_usmops_wide>;218}219 220let Predicates = [HasSME_MOP4] in {221 defm SMOP4A : sme_quarter_outer_product_i8_i32<0b0, 0b0, 0b0, "smop4a", "int_aarch64_sme_smop4a_wide">;222 defm SMOP4S : sme_quarter_outer_product_i8_i32<0b0, 0b0, 0b1, "smop4s", "int_aarch64_sme_smop4s_wide">;223 defm SUMOP4A : sme_quarter_outer_product_i8_i32<0b0, 0b1, 0b0, "sumop4a", "int_aarch64_sme_sumop4a_wide">;224 defm SUMOP4S : sme_quarter_outer_product_i8_i32<0b0, 0b1, 0b1, "sumop4s", "int_aarch64_sme_sumop4s_wide">;225 defm USMOP4A : sme_quarter_outer_product_i8_i32<0b1, 0b0, 0b0, "usmop4a", "int_aarch64_sme_usmop4a_wide">;226 defm USMOP4S : sme_quarter_outer_product_i8_i32<0b1, 0b0, 0b1, "usmop4s", "int_aarch64_sme_usmop4s_wide">;227 defm UMOP4A : sme_quarter_outer_product_i8_i32<0b1, 0b1, 0b0, "umop4a", "int_aarch64_sme_umop4a_wide">;228 defm UMOP4S : sme_quarter_outer_product_i8_i32<0b1, 0b1, 0b1, "umop4s", "int_aarch64_sme_umop4s_wide">;229 230 defm SMOP4A : sme_quarter_outer_product_i16_i32<0b0, 0b0, "smop4a", "int_aarch64_sme_smop4a_wide">;231 defm SMOP4S : sme_quarter_outer_product_i16_i32<0b0, 0b1, "smop4s", "int_aarch64_sme_smop4s_wide">;232 defm UMOP4A : sme_quarter_outer_product_i16_i32<0b1, 0b0, "umop4a", "int_aarch64_sme_umop4a_wide">;233 defm UMOP4S : sme_quarter_outer_product_i16_i32<0b1, 0b1, "umop4s", "int_aarch64_sme_umop4s_wide">;234}235 236let Predicates = [HasSME_MOP4, HasSMEI16I64] in {237 defm SMOP4A : sme_quarter_outer_product_i64<0b0, 0b0, 0b0, "smop4a", "int_aarch64_sme_smop4a_za64_wide">;238 defm SMOP4S : sme_quarter_outer_product_i64<0b0, 0b0, 0b1, "smop4s", "int_aarch64_sme_smop4s_za64_wide">;239 defm SUMOP4A : sme_quarter_outer_product_i64<0b0, 0b1, 0b0, "sumop4a", "int_aarch64_sme_sumop4a_za64_wide">;240 defm SUMOP4S : sme_quarter_outer_product_i64<0b0, 0b1, 0b1, "sumop4s", "int_aarch64_sme_sumop4s_za64_wide">;241 defm UMOP4A : sme_quarter_outer_product_i64<0b1, 0b1, 0b0, "umop4a", "int_aarch64_sme_umop4a_za64_wide">;242 defm UMOP4S : sme_quarter_outer_product_i64<0b1, 0b1, 0b1, "umop4s", "int_aarch64_sme_umop4s_za64_wide">;243 defm USMOP4A : sme_quarter_outer_product_i64<0b1, 0b0, 0b0, "usmop4a", "int_aarch64_sme_usmop4a_za64_wide">;244 defm USMOP4S : sme_quarter_outer_product_i64<0b1, 0b0, 0b1, "usmop4s", "int_aarch64_sme_usmop4s_za64_wide">;245}246 247let Predicates = [HasSME_TMOP] in {248 defm STMOPA_M2ZZZI_BtoS : sme_tmopa_32b<0b00100, ZZ_b_mul_r, ZPR8, nxv16i8, "stmopa", int_aarch64_sme_stmopa_za32>;249 defm STMOPA_M2ZZZI_HtoS : sme_tmopa_32b<0b00101, ZZ_h_mul_r, ZPR16, nxv8i16, "stmopa", int_aarch64_sme_stmopa_za32>;250 defm UTMOPA_M2ZZZI_BtoS : sme_tmopa_32b<0b11100, ZZ_b_mul_r, ZPR8, nxv16i8, "utmopa", int_aarch64_sme_utmopa_za32>;251 defm UTMOPA_M2ZZZI_HtoS : sme_tmopa_32b<0b10101, ZZ_h_mul_r, ZPR16, nxv8i16, "utmopa", int_aarch64_sme_utmopa_za32>;252 defm SUTMOPA_M2ZZZI_BtoS : sme_tmopa_32b<0b01100, ZZ_b_mul_r, ZPR8, nxv16i8, "sutmopa", int_aarch64_sme_sutmopa_za32>;253 defm USTMOPA_M2ZZZI_BtoS : sme_tmopa_32b<0b10100, ZZ_b_mul_r, ZPR8, nxv16i8, "ustmopa", int_aarch64_sme_ustmopa_za32>;254 defm FTMOPA_M2ZZZI_HtoS : sme_tmopa_32b<0b11000, ZZ_h_mul_r, ZPR16, nxv8f16, "ftmopa", int_aarch64_sme_ftmopa_za32, [FPCR]>;255 defm FTMOPA_M2ZZZI_StoS : sme_tmopa_32b<0b00000, ZZ_s_mul_r, ZPR32, nxv4f32, "ftmopa", int_aarch64_sme_ftmopa_za32, [FPCR]>;256 defm BFTMOPA_M2ZZZI_HtoS : sme_tmopa_32b<0b10000, ZZ_h_mul_r, ZPR16, nxv8bf16, "bftmopa", int_aarch64_sme_ftmopa_za32, [FPCR]>;257}258 259let Predicates = [HasSME_TMOP, HasSMEF16F16] in {260 defm FTMOPA_M2ZZZI_HtoH : sme_tmopa_16b<0b10001, ZZ_h_mul_r, ZPR16, nxv8f16, "ftmopa", int_aarch64_sme_ftmopa_za16, [FPCR]>;261}262 263let Predicates = [HasSME_TMOP, HasSMEB16B16] in {264 defm BFTMOPA_M2ZZZI_HtoH : sme_tmopa_16b<0b11001, ZZ_h_mul_r, ZPR16, nxv8bf16, "bftmopa", int_aarch64_sme_ftmopa_za16, [FPCR]>;265}266 267let Predicates = [HasSME_TMOP, HasSMEF8F16] in {268 defm FTMOPA_M2ZZZI_BtoH : sme_tmopa_16b<0b01001, ZZ_b_mul_r, ZPR8, nxv16i8, "ftmopa", int_aarch64_sme_ftmopa_za16, [FPMR, FPCR]>;269}270 271let Predicates = [HasSME_TMOP, HasSMEF8F32] in {272 defm FTMOPA_M2ZZZI_BtoS : sme_tmopa_32b<0b01000, ZZ_b_mul_r, ZPR8, nxv16i8, "ftmopa", int_aarch64_sme_ftmopa_za32, [FPMR, FPCR]>;273}274 275let Predicates = [HasSME] in {276//===----------------------------------------------------------------------===//277// Loads and stores278//===----------------------------------------------------------------------===//279 280defm LD1_MXIPXX : sme_mem_ld_ss<"ld1">;281defm ST1_MXIPXX : sme_mem_st_ss<"st1">;282 283//===----------------------------------------------------------------------===//284// Move instructions285//===----------------------------------------------------------------------===//286 287defm INSERT_MXIPZ : sme_vector_to_tile<"mova">;288defm EXTRACT_ZPMXI : sme_tile_to_vector<"mova">;289} // End let Predicates = [HasSME]290 291let Predicates = [HasSMEandIsNonStreamingSafe] in {292//===----------------------------------------------------------------------===//293// Spill + fill294//===----------------------------------------------------------------------===//295 296defm LDR_ZA : sme_fill<"ldr">;297defm STR_ZA : sme_spill<"str">;298 299//===----------------------------------------------------------------------===//300// Zero instruction301//===----------------------------------------------------------------------===//302 303defm ZERO_M : sme_zero<"zero">;304 305//===----------------------------------------------------------------------===//306// Mode selection and state access instructions307//===----------------------------------------------------------------------===//308 309// Pseudo to conditionally restore ZA state. This expands:310//311// pseudonode tpidr2_el0, tpidr2obj, restore_routine312//313// Into:314//315// if (tpidr2_el0 == 0)316// BL restore_routine, implicit-use tpidr2obj317//318def RestoreZAPseudo :319 Pseudo<(outs),320 (ins GPR64:$tpidr2_el0, GPR64sp:$tpidr2obj, i64imm:$restore_routine, variable_ops), []>,321 Sched<[]>;322 323def : Pat<(AArch64_restore_za324 (i64 GPR64:$tpidr2_el0), (i64 GPR64sp:$tpidr2obj), (i64 texternalsym:$restore_routine)),325 (RestoreZAPseudo GPR64:$tpidr2_el0, GPR64sp:$tpidr2obj, texternalsym:$restore_routine)>;326 327// Read and write TPIDR2_EL0328def : Pat<(int_aarch64_sme_set_tpidr2 i64:$val),329 (MSR 0xde85, GPR64:$val)>;330def : Pat<(i64 (int_aarch64_sme_get_tpidr2)),331 (MRS 0xde85)>;332 333} // End let Predicates = [HasSMEandIsNonStreamingSafe]334 335multiclass CoalescerBarrierPseudo<RegisterClass rc, list<ValueType> vts> {336 def NAME : Pseudo<(outs rc:$dst), (ins rc:$src), []>, Sched<[]> {337 let Constraints = "$dst = $src";338 }339 foreach vt = vts in {340 def : Pat<(vt (AArch64CoalescerBarrier (vt rc:$src))),341 (!cast<Instruction>(NAME) rc:$src)>;342 }343}344 345multiclass CoalescerBarriers {346 defm _FPR16 : CoalescerBarrierPseudo<FPR16, [bf16, f16]>;347 defm _FPR32 : CoalescerBarrierPseudo<FPR32, [f32]>;348 defm _FPR64 : CoalescerBarrierPseudo<FPR64, [f64, v8i8, v4i16, v2i32, v1i64, v4f16, v2f32, v1f64, v4bf16]>;349 defm _FPR128 : CoalescerBarrierPseudo<FPR128, [f128, v16i8, v8i16, v4i32, v2i64, v8f16, v4f32, v2f64, v8bf16]>;350}351 352defm COALESCER_BARRIER : CoalescerBarriers;353 354// Pseudo to match to smstart/smstop. This expands:355//356// pseudonode (pstate_za|pstate_sm), before_call, expected_value357//358// Into:359//360// if (before_call != expected_value)361// node (pstate_za|pstate_sm)362//363// where node can be either 'smstart' or 'smstop'.364//365// This pseudo and corresponding patterns don't need to be predicated by SME,366// because when they're emitted for streaming-compatible functions and run367// in a non-SME context the generated code-paths will never execute any368// SME instructions.369def MSRpstatePseudo :370 Pseudo<(outs),371 (ins svcr_op:$pstatefield, timm0_1:$imm, timm0_31:$condition, variable_ops), []>,372 Sched<[WriteSys]> {373 let hasPostISelHook = 1;374 let Uses = [VG];375 let Defs = [VG];376}377 378def : Pat<(AArch64_cond_smstart (i32 svcr_op:$pstate), (i64 timm0_31:$condition), (i64 GPR64:$pstatesm)),379 (MSRpstatePseudo svcr_op:$pstate, 0b1, timm0_31:$condition, GPR64:$pstatesm)>;380def : Pat<(AArch64_cond_smstop (i32 svcr_op:$pstate), (i64 timm0_31:$condition), (i64 GPR64:$pstatesm)),381 (MSRpstatePseudo svcr_op:$pstate, 0b0, timm0_31:$condition, GPR64:$pstatesm)>;382 383// Unconditional start/stop384def : Pat<(AArch64_smstart (i32 svcr_op:$pstate)),385 (MSRpstatesvcrImm1 svcr_op:$pstate, 0b1)>;386def : Pat<(AArch64_smstop (i32 svcr_op:$pstate)),387 (MSRpstatesvcrImm1 svcr_op:$pstate, 0b0)>;388 389//===----------------------------------------------------------------------===//390// SME2 Instructions391//===----------------------------------------------------------------------===//392let Predicates = [HasSME2] in {393defm ADD_VG2_M2ZZ_S : sme2_dot_mla_add_sub_array_vg2_single<"add", 0b0011010, MatrixOp32, ZZ_s, ZPR4b32, nxv4i32, int_aarch64_sme_add_write_single_za_vg1x2>;394defm ADD_VG4_M4ZZ_S : sme2_dot_mla_add_sub_array_vg4_single<"add", 0b0111010, MatrixOp32, ZZZZ_s, ZPR4b32, nxv4i32, int_aarch64_sme_add_write_single_za_vg1x4>;395defm ADD_VG2_M2Z2Z_S : sme2_dot_mla_add_sub_array_vg2_multi<"add", 0b0110010, MatrixOp32, ZZ_s_mul_r, nxv4i32, int_aarch64_sme_add_write_za_vg1x2>;396defm ADD_VG4_M4Z4Z_S : sme2_dot_mla_add_sub_array_vg4_multi<"add", 0b0110010, MatrixOp32, ZZZZ_s_mul_r, nxv4i32, int_aarch64_sme_add_write_za_vg1x4>;397 398defm ADD_VG2_2ZZ : sme2_int_sve_destructive_vector_vg2_single<"add", 0b0110000>;399defm ADD_VG4_4ZZ : sme2_int_sve_destructive_vector_vg4_single<"add", 0b0110000>;400 401defm SUB_VG2_M2ZZ_S : sme2_dot_mla_add_sub_array_vg2_single<"sub", 0b0011011, MatrixOp32, ZZ_s, ZPR4b32, nxv4i32, int_aarch64_sme_sub_write_single_za_vg1x2>;402defm SUB_VG4_M4ZZ_S : sme2_dot_mla_add_sub_array_vg4_single<"sub", 0b0111011, MatrixOp32, ZZZZ_s, ZPR4b32, nxv4i32, int_aarch64_sme_sub_write_single_za_vg1x4>;403defm SUB_VG2_M2Z2Z_S : sme2_dot_mla_add_sub_array_vg2_multi<"sub", 0b0110011, MatrixOp32, ZZ_s_mul_r, nxv4i32, int_aarch64_sme_sub_write_za_vg1x2>;404defm SUB_VG4_M4Z4Z_S : sme2_dot_mla_add_sub_array_vg4_multi<"sub", 0b0110011, MatrixOp32, ZZZZ_s_mul_r, nxv4i32, int_aarch64_sme_sub_write_za_vg1x4>;405 406defm FMLA_VG2_M2ZZ_S : sme2_dot_mla_add_sub_array_vg2_single<"fmla", 0b0011000, MatrixOp32, ZZ_s, ZPR4b32, nxv4f32, int_aarch64_sme_fmla_single_vg1x2>;407defm FMLA_VG4_M4ZZ_S : sme2_dot_mla_add_sub_array_vg4_single<"fmla", 0b0111000, MatrixOp32, ZZZZ_s, ZPR4b32, nxv4f32, int_aarch64_sme_fmla_single_vg1x4>;408defm FMLA_VG2_M2Z2Z_S : sme2_dot_mla_add_sub_array_vg2_multi<"fmla", 0b0110000, MatrixOp32, ZZ_s_mul_r, nxv4f32, int_aarch64_sme_fmla_vg1x2>;409defm FMLA_VG4_M4Z4Z_S : sme2_dot_mla_add_sub_array_vg4_multi<"fmla", 0b0110000, MatrixOp32, ZZZZ_s_mul_r, nxv4f32, int_aarch64_sme_fmla_vg1x4>;410defm FMLA_VG2_M2ZZI_S : sme2_multi_vec_array_vg2_index_32b<"fmla", 0b01, 0b0000, ZZ_s_mul_r, ZPR4b32, nxv4f32, int_aarch64_sme_fmla_lane_vg1x2>;411defm FMLA_VG4_M4ZZI_S : sme2_multi_vec_array_vg4_index_32b<"fmla", 0b0000, ZZZZ_s_mul_r, ZPR4b32, nxv4f32, int_aarch64_sme_fmla_lane_vg1x4>;412 413defm FMLS_VG2_M2ZZ_S : sme2_dot_mla_add_sub_array_vg2_single<"fmls", 0b0011001, MatrixOp32, ZZ_s, ZPR4b32, nxv4f32, int_aarch64_sme_fmls_single_vg1x2>;414defm FMLS_VG4_M4ZZ_S : sme2_dot_mla_add_sub_array_vg4_single<"fmls", 0b0111001, MatrixOp32, ZZZZ_s, ZPR4b32, nxv4f32, int_aarch64_sme_fmls_single_vg1x4>;415defm FMLS_VG2_M2Z2Z_S : sme2_dot_mla_add_sub_array_vg2_multi<"fmls", 0b0110001, MatrixOp32, ZZ_s_mul_r, nxv4f32, int_aarch64_sme_fmls_vg1x2>;416defm FMLS_VG4_M4Z4Z_S : sme2_dot_mla_add_sub_array_vg4_multi<"fmls", 0b0110001, MatrixOp32, ZZZZ_s_mul_r, nxv4f32, int_aarch64_sme_fmls_vg1x4>;417defm FMLS_VG2_M2ZZI_S : sme2_multi_vec_array_vg2_index_32b<"fmls", 0b01, 0b0010, ZZ_s_mul_r, ZPR4b32, nxv4f32, int_aarch64_sme_fmls_lane_vg1x2>;418defm FMLS_VG4_M4ZZI_S : sme2_multi_vec_array_vg4_index_32b<"fmls", 0b0010, ZZZZ_s_mul_r, ZPR4b32, nxv4f32, int_aarch64_sme_fmls_lane_vg1x4>;419 420defm ADD_VG2_M2Z_S : sme2_multivec_accum_add_sub_vg2<"add", 0b0010, MatrixOp32, ZZ_s_mul_r, nxv4i32, int_aarch64_sme_add_za32_vg1x2>;421defm ADD_VG4_M4Z_S : sme2_multivec_accum_add_sub_vg4<"add", 0b0010, MatrixOp32, ZZZZ_s_mul_r, nxv4i32, int_aarch64_sme_add_za32_vg1x4>;422 423defm SUB_VG2_M2Z_S : sme2_multivec_accum_add_sub_vg2<"sub", 0b0011, MatrixOp32, ZZ_s_mul_r, nxv4i32, int_aarch64_sme_sub_za32_vg1x2>;424defm SUB_VG4_M4Z_S : sme2_multivec_accum_add_sub_vg4<"sub", 0b0011, MatrixOp32, ZZZZ_s_mul_r, nxv4i32, int_aarch64_sme_sub_za32_vg1x4>;425 426defm FADD_VG2_M2Z_S : sme2_multivec_accum_add_sub_vg2<"fadd", 0b0000, MatrixOp32, ZZ_s_mul_r, nxv4f32, int_aarch64_sme_add_za32_vg1x2>;427defm FADD_VG4_M4Z_S : sme2_multivec_accum_add_sub_vg4<"fadd", 0b0000, MatrixOp32, ZZZZ_s_mul_r, nxv4f32, int_aarch64_sme_add_za32_vg1x4>;428 429defm FSUB_VG2_M2Z_S : sme2_multivec_accum_add_sub_vg2<"fsub", 0b0001, MatrixOp32, ZZ_s_mul_r, nxv4f32, int_aarch64_sme_sub_za32_vg1x2>;430defm FSUB_VG4_M4Z_S : sme2_multivec_accum_add_sub_vg4<"fsub", 0b0001, MatrixOp32, ZZZZ_s_mul_r, nxv4f32, int_aarch64_sme_sub_za32_vg1x4>;431 432defm SQDMULH_VG2_2ZZ : sme2_int_sve_destructive_vector_vg2_single<"sqdmulh", 0b1000000>;433defm SQDMULH_VG4_4ZZ : sme2_int_sve_destructive_vector_vg4_single<"sqdmulh", 0b1000000>;434defm SQDMULH_VG2_2Z2Z : sme2_int_sve_destructive_vector_vg2_multi<"sqdmulh", 0b1000000>;435defm SQDMULH_VG4_4Z4Z : sme2_int_sve_destructive_vector_vg4_multi<"sqdmulh", 0b1000000>;436 437defm FMLAL_MZZI : sme2_mla_long_array_index<"fmlal", 0b10, 0b00, nxv8f16, int_aarch64_sme_fmlal_lane_vg2x1>;438defm FMLAL_VG2_M2ZZI : sme2_fp_mla_long_array_vg2_index<"fmlal", 0b00, nxv8f16, int_aarch64_sme_fmlal_lane_vg2x2>;439defm FMLAL_VG4_M4ZZI : sme2_fp_mla_long_array_vg4_index<"fmlal", 0b00, nxv8f16, int_aarch64_sme_fmlal_lane_vg2x4>;440defm FMLAL_MZZ : sme2_mla_long_array_single<"fmlal", 0b00, 0b00, nxv8f16, int_aarch64_sme_fmlal_single_vg2x1>;441defm FMLAL_VG2_M2ZZ_HtoS : sme2_fp_mla_long_array_vg2_single<"fmlal", 0b000, MatrixOp32, ZZ_h, ZPR4b16, nxv8f16, int_aarch64_sme_fmlal_single_vg2x2>;442defm FMLAL_VG4_M4ZZ_HtoS : sme2_fp_mla_long_array_vg4_single<"fmlal", 0b000, MatrixOp32, ZZZZ_h, ZPR4b16, nxv8f16, int_aarch64_sme_fmlal_single_vg2x4>;443defm FMLAL_VG2_M2Z2Z_HtoS : sme2_fp_mla_long_array_vg2_multi<"fmlal", 0b000, MatrixOp32, ZZ_h_mul_r, nxv8f16, int_aarch64_sme_fmlal_vg2x2>;444defm FMLAL_VG4_M4Z4Z_HtoS : sme2_fp_mla_long_array_vg4_multi<"fmlal", 0b000, MatrixOp32, ZZZZ_h_mul_r, nxv8f16, int_aarch64_sme_fmlal_vg2x4>;445 446defm FMLSL_MZZI : sme2_mla_long_array_index<"fmlsl", 0b10, 0b01, nxv8f16, int_aarch64_sme_fmlsl_lane_vg2x1>;447defm FMLSL_VG2_M2ZZI : sme2_fp_mla_long_array_vg2_index<"fmlsl", 0b01, nxv8f16, int_aarch64_sme_fmlsl_lane_vg2x2>;448defm FMLSL_VG4_M4ZZI : sme2_fp_mla_long_array_vg4_index<"fmlsl", 0b01, nxv8f16, int_aarch64_sme_fmlsl_lane_vg2x4>;449defm FMLSL_MZZ : sme2_mla_long_array_single<"fmlsl", 0b00, 0b01, nxv8f16, int_aarch64_sme_fmlsl_single_vg2x1>;450defm FMLSL_VG2_M2ZZ_HtoS : sme2_fp_mla_long_array_vg2_single<"fmlsl", 0b010, MatrixOp32, ZZ_h, ZPR4b16, nxv8f16, int_aarch64_sme_fmlsl_single_vg2x2>;451defm FMLSL_VG4_M4ZZ_HtoS : sme2_fp_mla_long_array_vg4_single<"fmlsl", 0b010, MatrixOp32, ZZZZ_h, ZPR4b16, nxv8f16, int_aarch64_sme_fmlsl_single_vg2x4>;452defm FMLSL_VG2_M2Z2Z_HtoS : sme2_fp_mla_long_array_vg2_multi<"fmlsl", 0b001, MatrixOp32, ZZ_h_mul_r, nxv8f16, int_aarch64_sme_fmlsl_vg2x2>;453defm FMLSL_VG4_M4Z4Z_HtoS : sme2_fp_mla_long_array_vg4_multi<"fmlsl", 0b001, MatrixOp32, ZZZZ_h_mul_r, nxv8f16, int_aarch64_sme_fmlsl_vg2x4>;454 455defm BFMLAL_MZZI : sme2_mla_long_array_index<"bfmlal", 0b10, 0b10, nxv8bf16, int_aarch64_sme_fmlal_lane_vg2x1>;456defm BFMLAL_VG2_M2ZZI : sme2_fp_mla_long_array_vg2_index<"bfmlal", 0b10, nxv8bf16, int_aarch64_sme_fmlal_lane_vg2x2>;457defm BFMLAL_VG4_M4ZZI : sme2_fp_mla_long_array_vg4_index<"bfmlal", 0b10, nxv8bf16, int_aarch64_sme_fmlal_lane_vg2x4>;458defm BFMLAL_MZZ : sme2_mla_long_array_single<"bfmlal", 0b00, 0b10, nxv8bf16, int_aarch64_sme_fmlal_single_vg2x1>;459defm BFMLAL_VG2_M2ZZ_HtoS : sme2_fp_mla_long_array_vg2_single<"bfmlal", 0b100, MatrixOp32, ZZ_h, ZPR4b16, nxv8bf16, int_aarch64_sme_fmlal_single_vg2x2>;460defm BFMLAL_VG4_M4ZZ_HtoS : sme2_fp_mla_long_array_vg4_single<"bfmlal", 0b100, MatrixOp32, ZZZZ_h, ZPR4b16, nxv8bf16, int_aarch64_sme_fmlal_single_vg2x4>;461defm BFMLAL_VG2_M2Z2Z_HtoS : sme2_fp_mla_long_array_vg2_multi<"bfmlal", 0b010, MatrixOp32, ZZ_h_mul_r, nxv8bf16, int_aarch64_sme_fmlal_vg2x2>;462defm BFMLAL_VG4_M4Z4Z_HtoS : sme2_fp_mla_long_array_vg4_multi<"bfmlal", 0b010, MatrixOp32, ZZZZ_h_mul_r, nxv8bf16, int_aarch64_sme_fmlal_vg2x4>;463 464defm BFMLSL_MZZI : sme2_mla_long_array_index<"bfmlsl", 0b10, 0b11, nxv8bf16, int_aarch64_sme_fmlsl_lane_vg2x1>;465defm BFMLSL_VG2_M2ZZI : sme2_fp_mla_long_array_vg2_index<"bfmlsl", 0b11, nxv8bf16, int_aarch64_sme_fmlsl_lane_vg2x2>;466defm BFMLSL_VG4_M4ZZI : sme2_fp_mla_long_array_vg4_index<"bfmlsl", 0b11, nxv8bf16, int_aarch64_sme_fmlsl_lane_vg2x4>;467defm BFMLSL_MZZ : sme2_mla_long_array_single<"bfmlsl", 0b00, 0b11, nxv8bf16, int_aarch64_sme_fmlsl_single_vg2x1>;468defm BFMLSL_VG2_M2ZZ_HtoS : sme2_fp_mla_long_array_vg2_single<"bfmlsl", 0b110, MatrixOp32, ZZ_h, ZPR4b16, nxv8bf16, int_aarch64_sme_fmlsl_single_vg2x2>;469defm BFMLSL_VG4_M4ZZ_HtoS : sme2_fp_mla_long_array_vg4_single<"bfmlsl", 0b110, MatrixOp32, ZZZZ_h, ZPR4b16, nxv8bf16, int_aarch64_sme_fmlsl_single_vg2x4>;470defm BFMLSL_VG2_M2Z2Z_HtoS : sme2_fp_mla_long_array_vg2_multi<"bfmlsl", 0b011, MatrixOp32, ZZ_h_mul_r, nxv8bf16, int_aarch64_sme_fmlsl_vg2x2>;471defm BFMLSL_VG4_M4Z4Z_HtoS : sme2_fp_mla_long_array_vg4_multi<"bfmlsl", 0b011, MatrixOp32, ZZZZ_h_mul_r, nxv8bf16, int_aarch64_sme_fmlsl_vg2x4>;472 473defm SMLAL_MZZI : sme2_mla_long_array_index<"smlal", 0b11, 0b00, nxv8i16, int_aarch64_sme_smlal_lane_vg2x1>;474defm SMLAL_VG2_M2ZZI : sme2_int_mla_long_array_vg2_index<"smlal", 0b00, int_aarch64_sme_smlal_lane_vg2x2>;475defm SMLAL_VG4_M4ZZI : sme2_int_mla_long_array_vg4_index<"smlal", 0b00, int_aarch64_sme_smlal_lane_vg2x4>;476defm SMLAL_MZZ : sme2_mla_long_array_single<"smlal",0b01, 0b00, nxv8i16, int_aarch64_sme_smlal_single_vg2x1>;477defm SMLAL_VG2_M2ZZ : sme2_int_mla_long_array_vg2_single<"smlal", 0b00, int_aarch64_sme_smlal_single_vg2x2>;478defm SMLAL_VG4_M4ZZ : sme2_int_mla_long_array_vg4_single<"smlal", 0b00, int_aarch64_sme_smlal_single_vg2x4>;479defm SMLAL_VG2_M2Z2Z : sme2_int_mla_long_array_vg2_multi<"smlal", 0b00, int_aarch64_sme_smlal_vg2x2>;480defm SMLAL_VG4_M4Z4Z : sme2_int_mla_long_array_vg4_multi<"smlal", 0b00, int_aarch64_sme_smlal_vg2x4>;481 482defm SMLSL_MZZI : sme2_mla_long_array_index<"smlsl", 0b11, 0b01, nxv8i16, int_aarch64_sme_smlsl_lane_vg2x1>;483defm SMLSL_VG2_M2ZZI : sme2_int_mla_long_array_vg2_index<"smlsl", 0b01, int_aarch64_sme_smlsl_lane_vg2x2>;484defm SMLSL_VG4_M4ZZI : sme2_int_mla_long_array_vg4_index<"smlsl", 0b01, int_aarch64_sme_smlsl_lane_vg2x4>;485defm SMLSL_MZZ : sme2_mla_long_array_single<"smlsl",0b01, 0b01, nxv8i16, int_aarch64_sme_smlsl_single_vg2x1>;486defm SMLSL_VG2_M2ZZ : sme2_int_mla_long_array_vg2_single<"smlsl", 0b01, int_aarch64_sme_smlsl_single_vg2x2>;487defm SMLSL_VG4_M4ZZ : sme2_int_mla_long_array_vg4_single<"smlsl", 0b01, int_aarch64_sme_smlsl_single_vg2x4>;488defm SMLSL_VG2_M2Z2Z : sme2_int_mla_long_array_vg2_multi<"smlsl", 0b01, int_aarch64_sme_smlsl_vg2x2>;489defm SMLSL_VG4_M4Z4Z : sme2_int_mla_long_array_vg4_multi<"smlsl", 0b01, int_aarch64_sme_smlsl_vg2x4>;490 491defm UMLAL_MZZI : sme2_mla_long_array_index<"umlal", 0b11, 0b10, nxv8i16, int_aarch64_sme_umlal_lane_vg2x1>;492defm UMLAL_VG2_M2ZZI : sme2_int_mla_long_array_vg2_index<"umlal", 0b10, int_aarch64_sme_umlal_lane_vg2x2>;493defm UMLAL_VG4_M4ZZI : sme2_int_mla_long_array_vg4_index<"umlal", 0b10, int_aarch64_sme_umlal_lane_vg2x4>;494defm UMLAL_MZZ : sme2_mla_long_array_single<"umlal",0b01, 0b10, nxv8i16, int_aarch64_sme_umlal_single_vg2x1>;495defm UMLAL_VG2_M2ZZ : sme2_int_mla_long_array_vg2_single<"umlal", 0b10, int_aarch64_sme_umlal_single_vg2x2>;496defm UMLAL_VG4_M4ZZ : sme2_int_mla_long_array_vg4_single<"umlal", 0b10, int_aarch64_sme_umlal_single_vg2x4>;497defm UMLAL_VG2_M2Z2Z : sme2_int_mla_long_array_vg2_multi<"umlal", 0b10, int_aarch64_sme_umlal_vg2x2>;498defm UMLAL_VG4_M4Z4Z : sme2_int_mla_long_array_vg4_multi<"umlal", 0b10, int_aarch64_sme_umlal_vg2x4>;499 500defm UMLSL_MZZI : sme2_mla_long_array_index<"umlsl", 0b11, 0b11, nxv8i16, int_aarch64_sme_umlsl_lane_vg2x1>;501defm UMLSL_VG2_M2ZZI : sme2_int_mla_long_array_vg2_index<"umlsl", 0b11, int_aarch64_sme_umlsl_lane_vg2x2>;502defm UMLSL_VG4_M4ZZI : sme2_int_mla_long_array_vg4_index<"umlsl", 0b11, int_aarch64_sme_umlsl_lane_vg2x4>;503defm UMLSL_MZZ : sme2_mla_long_array_single<"umlsl",0b01, 0b11, nxv8i16, int_aarch64_sme_umlsl_single_vg2x1>;504defm UMLSL_VG2_M2ZZ : sme2_int_mla_long_array_vg2_single<"umlsl", 0b11, int_aarch64_sme_umlsl_single_vg2x2>;505defm UMLSL_VG4_M4ZZ : sme2_int_mla_long_array_vg4_single<"umlsl", 0b11, int_aarch64_sme_umlsl_single_vg2x4>;506defm UMLSL_VG2_M2Z2Z : sme2_int_mla_long_array_vg2_multi<"umlsl", 0b11, int_aarch64_sme_umlsl_vg2x2>;507defm UMLSL_VG4_M4Z4Z : sme2_int_mla_long_array_vg4_multi<"umlsl", 0b11, int_aarch64_sme_umlsl_vg2x4>;508 509defm FCVT_Z2Z_StoH : sme2_cvt_vg2_single<"fcvt", 0b00000, nxv8f16, nxv4f32, int_aarch64_sve_fcvt_x2>;510defm FCVTN_Z2Z_StoH : sme2_cvt_vg2_single<"fcvtn", 0b00001, nxv8f16, nxv4f32, int_aarch64_sve_fcvtn_x2>;511defm BFCVT_Z2Z_StoH : sme2_cvt_vg2_single<"bfcvt", 0b10000, nxv8bf16, nxv4f32, int_aarch64_sve_bfcvt_x2>;512defm BFCVTN_Z2Z_StoH : sme2_cvt_vg2_single<"bfcvtn", 0b10001, nxv8bf16, nxv4f32, int_aarch64_sve_bfcvtn_x2>;513 514defm SQCVT_Z2Z_StoH : sme2_cvt_vg2_single<"sqcvt", 0b00110, nxv8i16, nxv4i32, int_aarch64_sve_sqcvt_x2>;515defm UQCVT_Z2Z_StoH : sme2_cvt_vg2_single<"uqcvt", 0b00111, nxv8i16, nxv4i32, int_aarch64_sve_uqcvt_x2>;516defm SQCVTU_Z2Z_StoH : sme2_cvt_vg2_single<"sqcvtu", 0b10110, nxv8i16, nxv4i32, int_aarch64_sve_sqcvtu_x2>;517defm SQCVT_Z4Z : sme2_int_cvt_vg4_single<"sqcvt", 0b000, int_aarch64_sve_sqcvt_x4>;518defm UQCVT_Z4Z : sme2_int_cvt_vg4_single<"uqcvt", 0b001, int_aarch64_sve_uqcvt_x4>;519defm SQCVTU_Z4Z : sme2_int_cvt_vg4_single<"sqcvtu", 0b100, int_aarch64_sve_sqcvtu_x4>;520defm SQCVTN_Z4Z : sme2_int_cvt_vg4_single<"sqcvtn", 0b010, int_aarch64_sve_sqcvtn_x4>;521defm SQCVTUN_Z4Z : sme2_int_cvt_vg4_single<"sqcvtun", 0b110, int_aarch64_sve_sqcvtun_x4>;522defm UQCVTN_Z4Z : sme2_int_cvt_vg4_single<"uqcvtn", 0b011, int_aarch64_sve_uqcvtn_x4>;523 524defm FCVTZS_2Z2Z_StoS : sme2_fp_cvt_vg2_multi<"fcvtzs", 0b00010>;525defm FCVTZS_4Z4Z_StoS : sme2_fp_cvt_vg4_multi<"fcvtzs", 0b0001000>;526defm FCVTZU_2Z2Z_StoS : sme2_fp_cvt_vg2_multi<"fcvtzu", 0b00011>;527defm FCVTZU_4Z4Z_StoS : sme2_fp_cvt_vg4_multi<"fcvtzu", 0b0001010>;528defm SCVTF_2Z2Z_StoS : sme2_fp_cvt_vg2_multi<"scvtf", 0b00100>;529defm SCVTF_4Z4Z_StoS : sme2_fp_cvt_vg4_multi<"scvtf", 0b0010000>;530defm UCVTF_2Z2Z_StoS : sme2_fp_cvt_vg2_multi<"ucvtf", 0b00101>;531defm UCVTF_4Z4Z_StoS : sme2_fp_cvt_vg4_multi<"ucvtf", 0b0010010>;532 533defm SMAX_VG2_2ZZ : sme2_int_sve_destructive_vector_vg2_single<"smax", 0b0000000>;534defm SMAX_VG4_4ZZ : sme2_int_sve_destructive_vector_vg4_single<"smax", 0b0000000>;535defm SMAX_VG2_2Z2Z : sme2_int_sve_destructive_vector_vg2_multi<"smax", 0b0000000>;536defm SMAX_VG4_4Z4Z : sme2_int_sve_destructive_vector_vg4_multi<"smax", 0b0000000>;537 538defm UMAX_VG2_2ZZ : sme2_int_sve_destructive_vector_vg2_single<"umax", 0b0000001>;539defm UMAX_VG4_4ZZ : sme2_int_sve_destructive_vector_vg4_single<"umax", 0b0000001>;540defm UMAX_VG2_2Z2Z : sme2_int_sve_destructive_vector_vg2_multi<"umax", 0b0000001>;541defm UMAX_VG4_4Z4Z : sme2_int_sve_destructive_vector_vg4_multi<"umax", 0b0000001>;542 543defm SMIN_VG2_2ZZ : sme2_int_sve_destructive_vector_vg2_single<"smin", 0b0000010>;544defm SMIN_VG4_4ZZ : sme2_int_sve_destructive_vector_vg4_single<"smin", 0b0000010>;545defm SMIN_VG2_2Z2Z : sme2_int_sve_destructive_vector_vg2_multi<"smin", 0b0000010>;546defm SMIN_VG4_4Z4Z : sme2_int_sve_destructive_vector_vg4_multi<"smin", 0b0000010>;547 548defm UMIN_VG2_2ZZ : sme2_int_sve_destructive_vector_vg2_single<"umin", 0b0000011>;549defm UMIN_VG4_4ZZ : sme2_int_sve_destructive_vector_vg4_single<"umin", 0b0000011>;550defm UMIN_VG2_2Z2Z : sme2_int_sve_destructive_vector_vg2_multi<"umin", 0b0000011>;551defm UMIN_VG4_4Z4Z : sme2_int_sve_destructive_vector_vg4_multi<"umin", 0b0000011>;552 553defm FMAX_VG2_2ZZ : sme2_fp_sve_destructive_vector_vg2_single<"fmax", 0b0010000>;554defm FMAX_VG4_4ZZ : sme2_fp_sve_destructive_vector_vg4_single<"fmax", 0b0010000>;555defm FMAX_VG2_2Z2Z : sme2_fp_sve_destructive_vector_vg2_multi<"fmax", 0b0010000>;556defm FMAX_VG4_4Z4Z : sme2_fp_sve_destructive_vector_vg4_multi<"fmax", 0b0010000>;557 558defm FMIN_VG2_2ZZ : sme2_fp_sve_destructive_vector_vg2_single<"fmin", 0b0010001>;559defm FMIN_VG4_4ZZ : sme2_fp_sve_destructive_vector_vg4_single<"fmin", 0b0010001>;560defm FMIN_VG2_2Z2Z : sme2_fp_sve_destructive_vector_vg2_multi<"fmin", 0b0010001>;561defm FMIN_VG4_4Z4Z : sme2_fp_sve_destructive_vector_vg4_multi<"fmin", 0b0010001>;562 563defm FMAXNM_VG2_2ZZ : sme2_fp_sve_destructive_vector_vg2_single<"fmaxnm", 0b0010010>;564defm FMAXNM_VG4_4ZZ : sme2_fp_sve_destructive_vector_vg4_single<"fmaxnm", 0b0010010>;565defm FMAXNM_VG2_2Z2Z : sme2_fp_sve_destructive_vector_vg2_multi<"fmaxnm", 0b0010010>;566defm FMAXNM_VG4_4Z4Z : sme2_fp_sve_destructive_vector_vg4_multi<"fmaxnm", 0b0010010>;567 568defm FMINNM_VG2_2ZZ : sme2_fp_sve_destructive_vector_vg2_single<"fminnm", 0b0010011>;569defm FMINNM_VG4_4ZZ : sme2_fp_sve_destructive_vector_vg4_single<"fminnm", 0b0010011>;570defm FMINNM_VG2_2Z2Z : sme2_fp_sve_destructive_vector_vg2_multi<"fminnm", 0b0010011>;571defm FMINNM_VG4_4Z4Z : sme2_fp_sve_destructive_vector_vg4_multi<"fminnm", 0b0010011>;572 573defm SRSHL_VG2_2ZZ : sme2_int_sve_destructive_vector_vg2_single<"srshl", 0b0100010>;574defm SRSHL_VG4_4ZZ : sme2_int_sve_destructive_vector_vg4_single<"srshl", 0b0100010>;575defm SRSHL_VG2_2Z2Z : sme2_int_sve_destructive_vector_vg2_multi<"srshl", 0b0100010>;576defm SRSHL_VG4_4Z4Z : sme2_int_sve_destructive_vector_vg4_multi<"srshl", 0b0100010>;577 578defm URSHL_VG2_2ZZ : sme2_int_sve_destructive_vector_vg2_single<"urshl", 0b0100011>;579defm URSHL_VG4_4ZZ : sme2_int_sve_destructive_vector_vg4_single<"urshl", 0b0100011>;580defm URSHL_VG2_2Z2Z : sme2_int_sve_destructive_vector_vg2_multi<"urshl", 0b0100011>;581defm URSHL_VG4_4Z4Z : sme2_int_sve_destructive_vector_vg4_multi<"urshl", 0b0100011>;582 583defm FCLAMP_VG2_2Z2Z : sme2_fp_clamp_vector_vg2_multi<"fclamp">;584defm FCLAMP_VG4_4Z4Z : sme2_fp_clamp_vector_vg4_multi<"fclamp">;585 586defm SCLAMP_VG2_2Z2Z : sme2_int_clamp_vector_vg2_multi<"sclamp", 0b0>;587defm SCLAMP_VG4_4Z4Z : sme2_int_clamp_vector_vg4_multi<"sclamp", 0b0>;588 589defm UCLAMP_VG2_2Z2Z : sme2_int_clamp_vector_vg2_multi<"uclamp", 0b1>;590defm UCLAMP_VG4_4Z4Z : sme2_int_clamp_vector_vg4_multi<"uclamp", 0b1>;591 592defm FDOT_VG2_M2ZZI_HtoS : sme2_multi_vec_array_vg2_index_32b<"fdot", 0b01, 0b1001, ZZ_h_mul_r, ZPR4b16, nxv8f16, int_aarch64_sme_fdot_lane_za32_vg1x2>;593defm FDOT_VG4_M4ZZI_HtoS : sme2_multi_vec_array_vg4_index_32b<"fdot", 0b1001, ZZZZ_h_mul_r, ZPR4b16, nxv8f16, int_aarch64_sme_fdot_lane_za32_vg1x4>;594defm FDOT_VG2_M2ZZ_HtoS : sme2_dot_mla_add_sub_array_vg2_single<"fdot", 0b0010000, MatrixOp32, ZZ_h, ZPR4b16, nxv8f16, int_aarch64_sme_fdot_single_za32_vg1x2>;595defm FDOT_VG4_M4ZZ_HtoS : sme2_dot_mla_add_sub_array_vg4_single<"fdot", 0b0110000, MatrixOp32, ZZZZ_h, ZPR4b16, nxv8f16, int_aarch64_sme_fdot_single_za32_vg1x4>;596defm FDOT_VG2_M2Z2Z_HtoS : sme2_dot_mla_add_sub_array_vg2_multi<"fdot", 0b0100000, MatrixOp32, ZZ_h_mul_r, nxv8f16, int_aarch64_sme_fdot_za32_vg1x2>;597defm FDOT_VG4_M4Z4Z_HtoS : sme2_dot_mla_add_sub_array_vg4_multi<"fdot", 0b0100000, MatrixOp32, ZZZZ_h_mul_r, nxv8f16, int_aarch64_sme_fdot_za32_vg1x4>;598 599defm BFDOT_VG2_M2ZZI_HtoS : sme2_multi_vec_array_vg2_index_32b<"bfdot", 0b01, 0b1011, ZZ_h_mul_r, ZPR4b16, nxv8bf16, int_aarch64_sme_fdot_lane_za32_vg1x2>;600defm BFDOT_VG4_M4ZZI_HtoS : sme2_multi_vec_array_vg4_index_32b<"bfdot", 0b1011, ZZZZ_h_mul_r, ZPR4b16, nxv8bf16, int_aarch64_sme_fdot_lane_za32_vg1x4>;601defm BFDOT_VG2_M2ZZ_HtoS : sme2_dot_mla_add_sub_array_vg2_single<"bfdot", 0b0010010, MatrixOp32, ZZ_h, ZPR4b16, nxv8bf16, int_aarch64_sme_fdot_single_za32_vg1x2>;602defm BFDOT_VG4_M4ZZ_HtoS : sme2_dot_mla_add_sub_array_vg4_single<"bfdot", 0b0110010, MatrixOp32, ZZZZ_h, ZPR4b16, nxv8bf16, int_aarch64_sme_fdot_single_za32_vg1x4>;603defm BFDOT_VG2_M2Z2Z_HtoS : sme2_dot_mla_add_sub_array_vg2_multi<"bfdot", 0b0100010, MatrixOp32, ZZ_h_mul_r, nxv8bf16, int_aarch64_sme_fdot_za32_vg1x2>;604defm BFDOT_VG4_M4Z4Z_HtoS : sme2_dot_mla_add_sub_array_vg4_multi<"bfdot", 0b0100010, MatrixOp32, ZZZZ_h_mul_r, nxv8bf16, int_aarch64_sme_fdot_za32_vg1x4>;605 606defm BFVDOT_VG2_M2ZZI_HtoS : sme2_multi_vec_array_vg2_index_32b<"bfvdot", 0b01, 0b0011, ZZ_h_mul_r, ZPR4b16, nxv8bf16, int_aarch64_sme_fvdot_lane_za32_vg1x2>;607 608defm FVDOT_VG2_M2ZZI_HtoS : sme2_multi_vec_array_vg2_index_32b<"fvdot", 0b01, 0b0001, ZZ_h_mul_r, ZPR4b16, nxv8f16, int_aarch64_sme_fvdot_lane_za32_vg1x2>;609 610defm SDOT_VG2_M2ZZI_HToS : sme2_multi_vec_array_vg2_index_32b<"sdot", 0b01, 0b1000, ZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_sdot_lane_za32_vg1x2>;611defm SDOT_VG2_M2ZZI_BToS : sme2_multi_vec_array_vg2_index_32b<"sdot", 0b01, 0b1100, ZZ_b_mul_r, ZPR4b8, nxv16i8, int_aarch64_sme_sdot_lane_za32_vg1x2>;612defm SDOT_VG4_M4ZZI_HToS : sme2_multi_vec_array_vg4_index_32b<"sdot", 0b1000, ZZZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_sdot_lane_za32_vg1x4>;613defm SDOT_VG4_M4ZZI_BToS : sme2_multi_vec_array_vg4_index_32b<"sdot", 0b1100, ZZZZ_b_mul_r, ZPR4b8, nxv16i8, int_aarch64_sme_sdot_lane_za32_vg1x4>;614defm SDOT_VG2_M2ZZ_HtoS : sme2_dot_mla_add_sub_array_vg2_single<"sdot", 0b1010101, MatrixOp32, ZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_sdot_single_za32_vg1x2>;615defm SDOT_VG4_M4ZZ_HtoS : sme2_dot_mla_add_sub_array_vg4_single<"sdot", 0b1110101, MatrixOp32, ZZZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_sdot_single_za32_vg1x4>;616defm SDOT_VG2_M2Z2Z_HtoS : sme2_dot_mla_add_sub_array_vg2_multi<"sdot", 0b1101001, MatrixOp32, ZZ_h_mul_r, nxv8i16, int_aarch64_sme_sdot_za32_vg1x2>;617defm SDOT_VG4_M4Z4Z_HtoS : sme2_dot_mla_add_sub_array_vg4_multi<"sdot", 0b1101001, MatrixOp32, ZZZZ_h_mul_r, nxv8i16, int_aarch64_sme_sdot_za32_vg1x4>;618defm SDOT_VG2_M2ZZ_BtoS : sme2_dot_mla_add_sub_array_vg2_single<"sdot", 0b0010100, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_sdot_single_za32_vg1x2>;619defm SDOT_VG4_M4ZZ_BtoS : sme2_dot_mla_add_sub_array_vg4_single<"sdot", 0b0110100, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_sdot_single_za32_vg1x4>;620defm SDOT_VG2_M2Z2Z_BtoS : sme2_dot_mla_add_sub_array_vg2_multi<"sdot", 0b0101000, MatrixOp32, ZZ_b_mul_r, nxv16i8, int_aarch64_sme_sdot_za32_vg1x2>;621defm SDOT_VG4_M4Z4Z_BtoS : sme2_dot_mla_add_sub_array_vg4_multi<"sdot", 0b0101000, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, int_aarch64_sme_sdot_za32_vg1x4>;622 623defm SUDOT_VG2_M2ZZI_BToS : sme2_multi_vec_array_vg2_index_32b<"sudot", 0b01, 0b1111, ZZ_b_mul_r, ZPR4b8, nxv16i8, int_aarch64_sme_sudot_lane_za32_vg1x2>;624defm SUDOT_VG4_M4ZZI_BToS : sme2_multi_vec_array_vg4_index_32b<"sudot", 0b1111, ZZZZ_b_mul_r, ZPR4b8, nxv16i8, int_aarch64_sme_sudot_lane_za32_vg1x4>;625defm SUDOT_VG2_M2ZZ_BToS : sme2_dot_mla_add_sub_array_vg2_single<"sudot", 0b0010111, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_sudot_single_za32_vg1x2>;626defm SUDOT_VG4_M4ZZ_BToS : sme2_dot_mla_add_sub_array_vg4_single<"sudot", 0b0110111, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_sudot_single_za32_vg1x4>;627 628defm SVDOT_VG2_M2ZZI_HtoS : sme2_multi_vec_array_vg2_index_32b<"svdot", 0b01, 0b0100, ZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_svdot_lane_za32_vg1x2>;629defm SVDOT_VG4_M4ZZI_BtoS : sme2_multi_vec_array_vg4_index_32b<"svdot", 0b0100, ZZZZ_b_mul_r, ZPR4b8, nxv16i8, int_aarch64_sme_svdot_lane_za32_vg1x4>;630 631defm SUVDOT_VG4_M4ZZI_BToS : sme2_multi_vec_array_vg4_index_32b<"suvdot", 0b0111, ZZZZ_b_mul_r, ZPR4b8, nxv16i8, int_aarch64_sme_suvdot_lane_za32_vg1x4>;632 633defm UDOT_VG2_M2ZZI_HToS : sme2_multi_vec_array_vg2_index_32b<"udot", 0b01, 0b1010, ZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_udot_lane_za32_vg1x2>;634defm UDOT_VG2_M2ZZI_BToS : sme2_multi_vec_array_vg2_index_32b<"udot", 0b01, 0b1110, ZZ_b_mul_r, ZPR4b8, nxv16i8, int_aarch64_sme_udot_lane_za32_vg1x2>;635defm UDOT_VG4_M4ZZI_BtoS : sme2_multi_vec_array_vg4_index_32b<"udot", 0b1110, ZZZZ_b_mul_r, ZPR4b8, nxv16i8, int_aarch64_sme_udot_lane_za32_vg1x4>;636defm UDOT_VG4_M4ZZI_HToS : sme2_multi_vec_array_vg4_index_32b<"udot", 0b1010, ZZZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_udot_lane_za32_vg1x4>;637defm UDOT_VG2_M2ZZ_HtoS : sme2_dot_mla_add_sub_array_vg2_single<"udot", 0b1010111, MatrixOp32, ZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_udot_single_za32_vg1x2>;638defm UDOT_VG4_M4ZZ_HtoS : sme2_dot_mla_add_sub_array_vg4_single<"udot", 0b1110111, MatrixOp32, ZZZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_udot_single_za32_vg1x4>;639defm UDOT_VG2_M2Z2Z_HtoS : sme2_dot_mla_add_sub_array_vg2_multi<"udot", 0b1101011, MatrixOp32, ZZ_h_mul_r, nxv8i16, int_aarch64_sme_udot_za32_vg1x2>;640defm UDOT_VG4_M4Z4Z_HtoS : sme2_dot_mla_add_sub_array_vg4_multi<"udot", 0b1101011, MatrixOp32, ZZZZ_h_mul_r, nxv8i16, int_aarch64_sme_udot_za32_vg1x4>;641defm UDOT_VG2_M2ZZ_BtoS : sme2_dot_mla_add_sub_array_vg2_single<"udot", 0b0010110, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_udot_single_za32_vg1x2>;642defm UDOT_VG4_M4ZZ_BtoS : sme2_dot_mla_add_sub_array_vg4_single<"udot", 0b0110110, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_udot_single_za32_vg1x4>;643defm UDOT_VG2_M2Z2Z_BtoS : sme2_dot_mla_add_sub_array_vg2_multi<"udot", 0b0101010, MatrixOp32, ZZ_b_mul_r, nxv16i8, int_aarch64_sme_udot_za32_vg1x2>;644defm UDOT_VG4_M4Z4Z_BtoS : sme2_dot_mla_add_sub_array_vg4_multi<"udot", 0b0101010, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, int_aarch64_sme_udot_za32_vg1x4>;645 646defm USDOT_VG2_M2ZZI_BToS: sme2_multi_vec_array_vg2_index_32b<"usdot", 0b01, 0b1101, ZZ_b_mul_r, ZPR4b8, nxv16i8, int_aarch64_sme_usdot_lane_za32_vg1x2>;647defm USDOT_VG4_M4ZZI_BToS: sme2_multi_vec_array_vg4_index_32b<"usdot", 0b1101, ZZZZ_b_mul_r, ZPR4b8, nxv16i8, int_aarch64_sme_usdot_lane_za32_vg1x4>;648defm USDOT_VG2_M2ZZ_BToS : sme2_dot_mla_add_sub_array_vg2_single<"usdot", 0b0010101, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_usdot_single_za32_vg1x2>;649defm USDOT_VG4_M4ZZ_BToS : sme2_dot_mla_add_sub_array_vg4_single<"usdot", 0b0110101, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_usdot_single_za32_vg1x4>;650defm USDOT_VG2_M2Z2Z_BToS : sme2_dot_mla_add_sub_array_vg2_multi<"usdot", 0b0101001, MatrixOp32, ZZ_b_mul_r, nxv16i8, int_aarch64_sme_usdot_za32_vg1x2>;651defm USDOT_VG4_M4Z4Z_BToS : sme2_dot_mla_add_sub_array_vg4_multi<"usdot", 0b0101001, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, int_aarch64_sme_usdot_za32_vg1x4>;652 653defm USVDOT_VG4_M4ZZI_BToS : sme2_multi_vec_array_vg4_index_32b<"usvdot", 0b0101, ZZZZ_b_mul_r, ZPR4b8, nxv16i8, int_aarch64_sme_usvdot_lane_za32_vg1x4>;654 655defm UVDOT_VG2_M2ZZI_HtoS : sme2_multi_vec_array_vg2_index_32b<"uvdot", 0b01, 0b0110, ZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_uvdot_lane_za32_vg1x2>;656defm UVDOT_VG4_M4ZZI_BtoS : sme2_multi_vec_array_vg4_index_32b<"uvdot", 0b0110, ZZZZ_b_mul_r, ZPR4b8, nxv16i8, int_aarch64_sme_uvdot_lane_za32_vg1x4>;657 658defm SMLALL_MZZI_BtoS : sme2_mla_ll_array_index_32b<"smlall", 0b00, 0b000, int_aarch64_sme_smla_za32_lane_vg4x1>;659defm SMLALL_VG2_M2ZZI_BtoS : sme2_mla_ll_array_vg2_index_32b<"smlall", 0b00, 0b000, int_aarch64_sme_smla_za32_lane_vg4x2>;660defm SMLALL_VG4_M4ZZI_BtoS : sme2_mla_ll_array_vg4_index_32b<"smlall", 0b00, 0b0000, int_aarch64_sme_smla_za32_lane_vg4x4>;661defm SMLALL_MZZ_BtoS : sme2_mla_ll_array_single<"smlall", 0b00000, MatrixOp32, ZPR8, ZPR4b8, nxv16i8, int_aarch64_sme_smla_za32_single_vg4x1>;662defm SMLALL_VG2_M2ZZ_BtoS : sme2_mla_ll_array_vg2_single<"smlall", 0b000000, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_smla_za32_single_vg4x2>;663defm SMLALL_VG4_M4ZZ_BtoS : sme2_mla_ll_array_vg4_single<"smlall", 0b010000, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_smla_za32_single_vg4x4>;664defm SMLALL_VG2_M2Z2Z_BtoS : sme2_mla_ll_array_vg2_multi<"smlall", 0b00000, MatrixOp32, ZZ_b_mul_r, nxv16i8, int_aarch64_sme_smla_za32_vg4x2>;665defm SMLALL_VG4_M4Z4Z_BtoS : sme2_mla_ll_array_vg4_multi<"smlall", 0b00000, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, int_aarch64_sme_smla_za32_vg4x4>;666 667defm USMLALL_MZZI_BtoS : sme2_mla_ll_array_index_32b<"usmlall", 0b00, 0b001, int_aarch64_sme_usmla_za32_lane_vg4x1>;668defm USMLALL_VG2_M2ZZI_BtoS : sme2_mla_ll_array_vg2_index_32b<"usmlall", 0b00, 0b100, int_aarch64_sme_usmla_za32_lane_vg4x2>;669defm USMLALL_VG4_M4ZZI_BtoS : sme2_mla_ll_array_vg4_index_32b<"usmlall", 0b00, 0b0100, int_aarch64_sme_usmla_za32_lane_vg4x4>;670defm USMLALL_MZZ_BtoS : sme2_mla_ll_array_single<"usmlall", 0b00001, MatrixOp32, ZPR8, ZPR4b8, nxv16i8, int_aarch64_sme_usmla_za32_single_vg4x1>;671defm USMLALL_VG2_M2ZZ_BtoS : sme2_mla_ll_array_vg2_single<"usmlall", 0b000010, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_usmla_za32_single_vg4x2>;672defm USMLALL_VG4_M4ZZ_BtoS : sme2_mla_ll_array_vg4_single<"usmlall", 0b010010, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_usmla_za32_single_vg4x4>;673defm USMLALL_VG2_M2Z2Z_BtoS : sme2_mla_ll_array_vg2_multi<"usmlall", 0b00001, MatrixOp32, ZZ_b_mul_r, nxv16i8, int_aarch64_sme_usmla_za32_vg4x2>;674defm USMLALL_VG4_M4Z4Z_BtoS : sme2_mla_ll_array_vg4_multi<"usmlall", 0b00001, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, int_aarch64_sme_usmla_za32_vg4x4>;675 676defm SMLSLL_MZZI_BtoS : sme2_mla_ll_array_index_32b<"smlsll", 0b00, 0b010, int_aarch64_sme_smls_za32_lane_vg4x1>;677defm SMLSLL_VG2_M2ZZI_BtoS : sme2_mla_ll_array_vg2_index_32b<"smlsll", 0b00, 0b001, int_aarch64_sme_smls_za32_lane_vg4x2>;678defm SMLSLL_VG4_M4ZZI_BtoS : sme2_mla_ll_array_vg4_index_32b<"smlsll", 0b00, 0b0001, int_aarch64_sme_smls_za32_lane_vg4x4>;679defm SMLSLL_MZZ_BtoS : sme2_mla_ll_array_single<"smlsll", 0b00010, MatrixOp32, ZPR8, ZPR4b8, nxv16i8, int_aarch64_sme_smls_za32_single_vg4x1>;680defm SMLSLL_VG2_M2ZZ_BtoS : sme2_mla_ll_array_vg2_single<"smlsll", 0b000100, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_smls_za32_single_vg4x2>;681defm SMLSLL_VG4_M4ZZ_BtoS : sme2_mla_ll_array_vg4_single<"smlsll", 0b010100, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_smls_za32_single_vg4x4>;682defm SMLSLL_VG2_M2Z2Z_BtoS : sme2_mla_ll_array_vg2_multi<"smlsll", 0b00010, MatrixOp32, ZZ_b_mul_r, nxv16i8, int_aarch64_sme_smls_za32_vg4x2>;683defm SMLSLL_VG4_M4Z4Z_BtoS : sme2_mla_ll_array_vg4_multi<"smlsll", 0b00010, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, int_aarch64_sme_smls_za32_vg4x4>;684 685defm UMLALL_MZZI_BtoS : sme2_mla_ll_array_index_32b<"umlall", 0b00, 0b100, int_aarch64_sme_umla_za32_lane_vg4x1>;686defm UMLALL_VG2_M2ZZI_BtoS : sme2_mla_ll_array_vg2_index_32b<"umlall", 0b00, 0b010, int_aarch64_sme_umla_za32_lane_vg4x2>;687defm UMLALL_VG4_M4ZZI_BtoS : sme2_mla_ll_array_vg4_index_32b<"umlall", 0b00, 0b0010, int_aarch64_sme_umla_za32_lane_vg4x4>;688defm UMLALL_MZZ_BtoS : sme2_mla_ll_array_single<"umlall", 0b00100, MatrixOp32, ZPR8, ZPR4b8, nxv16i8, int_aarch64_sme_umla_za32_single_vg4x1>;689defm UMLALL_VG2_M2ZZ_BtoS : sme2_mla_ll_array_vg2_single<"umlall", 0b001000, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_umla_za32_single_vg4x2>;690defm UMLALL_VG4_M4ZZ_BtoS : sme2_mla_ll_array_vg4_single<"umlall", 0b011000, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_umla_za32_single_vg4x4>;691defm UMLALL_VG2_M2Z2Z_BtoS : sme2_mla_ll_array_vg2_multi<"umlall", 0b00100, MatrixOp32, ZZ_b_mul_r, nxv16i8, int_aarch64_sme_umla_za32_vg4x2>;692defm UMLALL_VG4_M4Z4Z_BtoS : sme2_mla_ll_array_vg4_multi<"umlall", 0b00100, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, int_aarch64_sme_umla_za32_vg4x4>;693 694defm SUMLALL_MZZI_BtoS : sme2_mla_ll_array_index_32b<"sumlall", 0b00, 0b101, int_aarch64_sme_sumla_za32_lane_vg4x1>;695defm SUMLALL_VG2_M2ZZI_BtoS : sme2_mla_ll_array_vg2_index_32b<"sumlall", 0b00, 0b110, int_aarch64_sme_sumla_za32_lane_vg4x2>;696defm SUMLALL_VG4_M4ZZI_BtoS : sme2_mla_ll_array_vg4_index_32b<"sumlall", 0b00, 0b0110, int_aarch64_sme_sumla_za32_lane_vg4x4>;697defm SUMLALL_VG2_M2ZZ_BtoS : sme2_mla_ll_array_vg2_single<"sumlall", 0b001010, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_sumla_za32_single_vg4x2>;698defm SUMLALL_VG4_M4ZZ_BtoS : sme2_mla_ll_array_vg4_single<"sumlall", 0b011010, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_sumla_za32_single_vg4x4>;699 700defm UMLSLL_MZZI_BtoS : sme2_mla_ll_array_index_32b<"umlsll", 0b00, 0b110, int_aarch64_sme_umls_za32_lane_vg4x1>;701defm UMLSLL_VG2_M2ZZI_BtoS : sme2_mla_ll_array_vg2_index_32b<"umlsll", 0b00, 0b011, int_aarch64_sme_umls_za32_lane_vg4x2>;702defm UMLSLL_VG4_M4ZZI_BtoS : sme2_mla_ll_array_vg4_index_32b<"umlsll", 0b00, 0b0011, int_aarch64_sme_umls_za32_lane_vg4x4>;703defm UMLSLL_MZZ_BtoS : sme2_mla_ll_array_single<"umlsll", 0b00110, MatrixOp32, ZPR8, ZPR4b8, nxv16i8, int_aarch64_sme_umls_za32_single_vg4x1>;704defm UMLSLL_VG2_M2ZZ_BtoS : sme2_mla_ll_array_vg2_single<"umlsll", 0b001100, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_umls_za32_single_vg4x2>;705defm UMLSLL_VG4_M4ZZ_BtoS : sme2_mla_ll_array_vg4_single<"umlsll", 0b011100, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_umls_za32_single_vg4x4>;706defm UMLSLL_VG2_M2Z2Z_BtoS : sme2_mla_ll_array_vg2_multi<"umlsll", 0b00110, MatrixOp32, ZZ_b_mul_r, nxv16i8, int_aarch64_sme_umls_za32_vg4x2>;707defm UMLSLL_VG4_M4Z4Z_BtoS : sme2_mla_ll_array_vg4_multi<"umlsll", 0b00110, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, int_aarch64_sme_umls_za32_vg4x4>;708 709defm BMOPA_MPPZZ_S : sme2_int_bmopx_tile<"bmopa", 0b100, int_aarch64_sme_bmopa_za32>;710defm BMOPS_MPPZZ_S : sme2_int_bmopx_tile<"bmops", 0b101, int_aarch64_sme_bmops_za32>;711 712defm SMOPA_MPPZZ_HtoS : sme2_int_mopx_tile<"smopa", 0b000, int_aarch64_sme_smopa_za32>;713defm SMOPS_MPPZZ_HtoS : sme2_int_mopx_tile<"smops", 0b001, int_aarch64_sme_smops_za32>;714 715defm UMOPA_MPPZZ_HtoS : sme2_int_mopx_tile<"umopa", 0b100, int_aarch64_sme_umopa_za32>;716defm UMOPS_MPPZZ_HtoS : sme2_int_mopx_tile<"umops", 0b101, int_aarch64_sme_umops_za32>;717 718def MOVT_XTI : sme2_movt_zt_to_scalar<"movt", 0b0011111>;719def MOVT_TIX : sme2_movt_scalar_to_zt<"movt", 0b0011111>;720 721defm LUTI2_ZTZI : sme2_luti2_vector_index<"luti2", int_aarch64_sme_luti2_lane_zt>;722defm LUTI2_2ZTZI : sme2_luti2_vector_vg2_index<"luti2">;723defm LUTI2_4ZTZI : sme2_luti2_vector_vg4_index<"luti2">;724 725defm LUTI4_ZTZI : sme2_luti4_vector_index<"luti4", int_aarch64_sme_luti4_lane_zt>;726defm LUTI4_2ZTZI : sme2_luti4_vector_vg2_index<"luti4">;727defm LUTI4_4ZTZI : sme2_luti4_vector_vg4_index<"luti4">;728 729defm SUNPK_VG2_2ZZ : sme2_unpk_vector_vg2<"sunpk", 0b0>;730defm SUNPK_VG4_4Z2Z : sme2_unpk_vector_vg4<"sunpk", 0b0>;731defm UUNPK_VG2_2ZZ : sme2_unpk_vector_vg2<"uunpk", 0b1>;732defm UUNPK_VG4_4Z2Z : sme2_unpk_vector_vg4<"uunpk", 0b1>;733 734defm ZIP_VG2_2ZZZ : sme2_zip_vector_vg2<"zip", 0b0>;735defm UZP_VG2_2ZZZ : sme2_zip_vector_vg2<"uzp", 0b1>;736defm ZIP_VG4_4Z4Z : sme2_zip_vector_vg4<"zip", 0b0110000>;737defm UZP_VG4_4Z4Z : sme2_zip_vector_vg4<"uzp", 0b0110001>;738defm ZIP_VG4_4Z4Z_Q : sme2_zip_vector_vg4_Q<"zip", 0b0111000>;739defm UZP_VG4_4Z4Z_Q : sme2_zip_vector_vg4_Q<"uzp", 0b0111001>;740 741defm FRINTA_2Z2Z: sme2_frint_vector_vg2_multi<"frinta", 0b11000>;742defm FRINTA_4Z4Z: sme2_frint_vector_vg4_multi<"frinta", 0b1100000>;743defm FRINTM_2Z2Z: sme2_frint_vector_vg2_multi<"frintm", 0b10100>;744defm FRINTM_4Z4Z: sme2_frint_vector_vg4_multi<"frintm", 0b1010000>;745defm FRINTN_2Z2Z: sme2_frint_vector_vg2_multi<"frintn", 0b10000>;746defm FRINTN_4Z4Z: sme2_frint_vector_vg4_multi<"frintn", 0b1000000>;747defm FRINTP_2Z2Z: sme2_frint_vector_vg2_multi<"frintp", 0b10010>;748defm FRINTP_4Z4Z: sme2_frint_vector_vg4_multi<"frintp", 0b1001000>;749 750defm MOVA_MXI2Z : sme2_mova_vec_to_tile_vg2_multi<"mova", int_aarch64_sme_write_hor_vg2, int_aarch64_sme_write_ver_vg2>;751defm MOVA_MXI4Z : sme2_mova_vec_to_tile_vg4_multi<"mova", int_aarch64_sme_write_hor_vg4, int_aarch64_sme_write_ver_vg4>;752defm MOVA_2ZMXI : sme2_mova_tile_to_vec_vg2_multi<"mova">;753defm MOVA_4ZMXI : sme2_mova_tile_to_vec_vg4_multi<"mova">;754 755defm MOVA_VG2_MXI2Z : sme2_mova_vec_to_array_vg2_multi<"mova", int_aarch64_sme_write_vg1x2>;756defm MOVA_VG4_MXI4Z : sme2_mova_vec_to_array_vg4_multi<"mova", int_aarch64_sme_write_vg1x4>;757defm MOVA_VG2_2ZMXI : sme2_mova_array_to_vec_vg2_multi<0b000, "mova">;758defm MOVA_VG4_4ZMXI : sme2_mova_array_to_vec_vg4_multi<0b1000, "mova">;759 760defm SQRSHR_VG2_Z2ZI : sme2_sat_shift_vector_vg2<"sqrshr", 0b0, 0b0, int_aarch64_sve_sqrshr_x2>;761defm SQRSHR_VG4_Z4ZI : sme2_sat_shift_vector_vg4<"sqrshr", 0b000, int_aarch64_sve_sqrshr_x4>;762 763defm UQRSHR_VG2_Z2ZI : sme2_sat_shift_vector_vg2<"uqrshr", 0b0, 0b1, int_aarch64_sve_uqrshr_x2>;764defm UQRSHR_VG4_Z4ZI : sme2_sat_shift_vector_vg4<"uqrshr", 0b001, int_aarch64_sve_uqrshr_x4>;765 766defm SQRSHRU_VG2_Z2ZI : sme2_sat_shift_vector_vg2<"sqrshru", 0b1, 0b0, int_aarch64_sve_sqrshru_x2>;767defm SQRSHRU_VG4_Z4ZI : sme2_sat_shift_vector_vg4<"sqrshru", 0b010, int_aarch64_sve_sqrshru_x4>;768 769defm SQRSHRN_VG4_Z4ZI : sme2_sat_shift_vector_vg4<"sqrshrn", 0b100, int_aarch64_sve_sqrshrn_x4>;770defm UQRSHRN_VG4_Z4ZI : sme2_sat_shift_vector_vg4<"uqrshrn", 0b101, int_aarch64_sve_uqrshrn_x4>;771defm SQRSHRUN_VG4_Z4ZI : sme2_sat_shift_vector_vg4<"sqrshrun", 0b110, int_aarch64_sve_sqrshrun_x4>;772 773defm SEL_VG2_2ZC2Z2Z: sme2_sel_vector_vg2<"sel">;774defm SEL_VG4_4ZC4Z4Z: sme2_sel_vector_vg4<"sel">;775 776def LD1B_2Z_STRIDED : sme2_ld_vector_vg2_multi_scalar_scalar<0b00, 0b0, ZZ_b_strided, GPR64shifted8, "ld1b">;777def LD1B_4Z_STRIDED : sme2_ld_vector_vg4_multi_scalar_scalar<0b00, 0b0, ZZZZ_b_strided, GPR64shifted8, "ld1b">;778defm LD1B_2Z_STRIDED_IMM : sme2_ld_vector_vg2_multi_scalar_immediate<0b00, 0b0, ZZ_b_strided, simm4s2, "ld1b">;779defm LD1B_4Z_STRIDED_IMM : sme2_ld_vector_vg4_multi_scalar_immediate<0b00, 0b0, ZZZZ_b_strided, simm4s4, "ld1b">;780def LD1H_2Z_STRIDED : sme2_ld_vector_vg2_multi_scalar_scalar<0b01, 0b0, ZZ_h_strided, GPR64shifted16, "ld1h">;781def LD1H_4Z_STRIDED : sme2_ld_vector_vg4_multi_scalar_scalar<0b01, 0b0, ZZZZ_h_strided, GPR64shifted16, "ld1h">;782defm LD1H_2Z_STRIDED_IMM : sme2_ld_vector_vg2_multi_scalar_immediate<0b01, 0b0, ZZ_h_strided, simm4s2, "ld1h">;783defm LD1H_4Z_STRIDED_IMM : sme2_ld_vector_vg4_multi_scalar_immediate<0b01, 0b0, ZZZZ_h_strided, simm4s4, "ld1h">;784def LD1W_2Z_STRIDED : sme2_ld_vector_vg2_multi_scalar_scalar<0b10, 0b0, ZZ_s_strided, GPR64shifted32, "ld1w">;785def LD1W_4Z_STRIDED : sme2_ld_vector_vg4_multi_scalar_scalar<0b10, 0b0, ZZZZ_s_strided, GPR64shifted32, "ld1w">;786defm LD1W_2Z_STRIDED_IMM : sme2_ld_vector_vg2_multi_scalar_immediate<0b10, 0b0, ZZ_s_strided, simm4s2, "ld1w">;787defm LD1W_4Z_STRIDED_IMM : sme2_ld_vector_vg4_multi_scalar_immediate<0b10, 0b0, ZZZZ_s_strided, simm4s4, "ld1w">;788def LD1D_2Z_STRIDED : sme2_ld_vector_vg2_multi_scalar_scalar<0b11, 0b0, ZZ_d_strided, GPR64shifted64, "ld1d">;789def LD1D_4Z_STRIDED : sme2_ld_vector_vg4_multi_scalar_scalar<0b11, 0b0, ZZZZ_d_strided, GPR64shifted64, "ld1d">;790defm LD1D_2Z_STRIDED_IMM : sme2_ld_vector_vg2_multi_scalar_immediate<0b11, 0b0, ZZ_d_strided, simm4s2, "ld1d">;791defm LD1D_4Z_STRIDED_IMM : sme2_ld_vector_vg4_multi_scalar_immediate<0b11, 0b0, ZZZZ_d_strided, simm4s4, "ld1d">;792 793def LDNT1B_2Z_STRIDED : sme2_ld_vector_vg2_multi_scalar_scalar<0b00, 0b1, ZZ_b_strided, GPR64shifted8, "ldnt1b">;794def LDNT1B_4Z_STRIDED : sme2_ld_vector_vg4_multi_scalar_scalar<0b00, 0b1, ZZZZ_b_strided, GPR64shifted8, "ldnt1b">;795defm LDNT1B_2Z_STRIDED_IMM : sme2_ld_vector_vg2_multi_scalar_immediate<0b00, 0b1, ZZ_b_strided, simm4s2, "ldnt1b">;796defm LDNT1B_4Z_STRIDED_IMM : sme2_ld_vector_vg4_multi_scalar_immediate<0b00, 0b1, ZZZZ_b_strided, simm4s4, "ldnt1b">;797def LDNT1H_2Z_STRIDED : sme2_ld_vector_vg2_multi_scalar_scalar<0b01, 0b1, ZZ_h_strided, GPR64shifted16, "ldnt1h">;798def LDNT1H_4Z_STRIDED : sme2_ld_vector_vg4_multi_scalar_scalar<0b01, 0b1, ZZZZ_h_strided, GPR64shifted16, "ldnt1h">;799defm LDNT1H_2Z_STRIDED_IMM : sme2_ld_vector_vg2_multi_scalar_immediate<0b01, 0b1, ZZ_h_strided, simm4s2, "ldnt1h">;800defm LDNT1H_4Z_STRIDED_IMM : sme2_ld_vector_vg4_multi_scalar_immediate<0b01, 0b1, ZZZZ_h_strided, simm4s4, "ldnt1h">;801def LDNT1W_2Z_STRIDED : sme2_ld_vector_vg2_multi_scalar_scalar<0b10, 0b1, ZZ_s_strided, GPR64shifted32, "ldnt1w">;802def LDNT1W_4Z_STRIDED : sme2_ld_vector_vg4_multi_scalar_scalar<0b10, 0b1, ZZZZ_s_strided, GPR64shifted32, "ldnt1w">;803defm LDNT1W_2Z_STRIDED_IMM : sme2_ld_vector_vg2_multi_scalar_immediate<0b10, 0b1, ZZ_s_strided, simm4s2, "ldnt1w">;804defm LDNT1W_4Z_STRIDED_IMM : sme2_ld_vector_vg4_multi_scalar_immediate<0b10, 0b1, ZZZZ_s_strided, simm4s4, "ldnt1w">;805def LDNT1D_2Z_STRIDED : sme2_ld_vector_vg2_multi_scalar_scalar<0b11, 0b1, ZZ_d_strided, GPR64shifted64, "ldnt1d">;806def LDNT1D_4Z_STRIDED : sme2_ld_vector_vg4_multi_scalar_scalar<0b11, 0b1, ZZZZ_d_strided, GPR64shifted64, "ldnt1d">;807defm LDNT1D_2Z_STRIDED_IMM : sme2_ld_vector_vg2_multi_scalar_immediate<0b11, 0b1, ZZ_d_strided, simm4s2, "ldnt1d">;808defm LDNT1D_4Z_STRIDED_IMM : sme2_ld_vector_vg4_multi_scalar_immediate<0b11, 0b1, ZZZZ_d_strided, simm4s4, "ldnt1d">;809 810def ST1B_2Z_STRIDED : sme2_st_vector_vg2_multi_scalar_scalar<0b00, 0b0, ZZ_b_strided, GPR64shifted8, "st1b">;811def ST1B_4Z_STRIDED : sme2_st_vector_vg4_multi_scalar_scalar<0b00, 0b0, ZZZZ_b_strided, GPR64shifted8, "st1b">;812defm ST1B_2Z_STRIDED_IMM : sme2_st_vector_vg2_multi_scalar_immediate<0b00, 0b0, ZZ_b_strided, simm4s2, "st1b">;813defm ST1B_4Z_STRIDED_IMM : sme2_st_vector_vg4_multi_scalar_immediate<0b00, 0b0, ZZZZ_b_strided, simm4s4, "st1b">;814def ST1H_2Z_STRIDED : sme2_st_vector_vg2_multi_scalar_scalar<0b01, 0b0, ZZ_h_strided, GPR64shifted16, "st1h">;815def ST1H_4Z_STRIDED : sme2_st_vector_vg4_multi_scalar_scalar<0b01, 0b0, ZZZZ_h_strided, GPR64shifted16, "st1h">;816defm ST1H_2Z_STRIDED_IMM : sme2_st_vector_vg2_multi_scalar_immediate<0b01, 0b0, ZZ_h_strided, simm4s2, "st1h">;817defm ST1H_4Z_STRIDED_IMM : sme2_st_vector_vg4_multi_scalar_immediate<0b01, 0b0, ZZZZ_h_strided, simm4s4, "st1h">;818def ST1W_2Z_STRIDED : sme2_st_vector_vg2_multi_scalar_scalar<0b10, 0b0, ZZ_s_strided, GPR64shifted32, "st1w">;819def ST1W_4Z_STRIDED : sme2_st_vector_vg4_multi_scalar_scalar<0b10, 0b0, ZZZZ_s_strided, GPR64shifted32, "st1w">;820defm ST1W_2Z_STRIDED_IMM : sme2_st_vector_vg2_multi_scalar_immediate<0b10, 0b0, ZZ_s_strided, simm4s2, "st1w">;821defm ST1W_4Z_STRIDED_IMM : sme2_st_vector_vg4_multi_scalar_immediate<0b10, 0b0, ZZZZ_s_strided, simm4s4, "st1w">;822def ST1D_2Z_STRIDED : sme2_st_vector_vg2_multi_scalar_scalar<0b11, 0b0, ZZ_d_strided, GPR64shifted64, "st1d">;823def ST1D_4Z_STRIDED : sme2_st_vector_vg4_multi_scalar_scalar<0b11, 0b0, ZZZZ_d_strided, GPR64shifted64, "st1d">;824defm ST1D_2Z_STRIDED_IMM : sme2_st_vector_vg2_multi_scalar_immediate<0b11, 0b0, ZZ_d_strided, simm4s2, "st1d">;825defm ST1D_4Z_STRIDED_IMM : sme2_st_vector_vg4_multi_scalar_immediate<0b11, 0b0, ZZZZ_d_strided, simm4s4, "st1d">;826 827def STNT1B_2Z_STRIDED : sme2_st_vector_vg2_multi_scalar_scalar<0b00, 0b1, ZZ_b_strided, GPR64shifted8, "stnt1b">;828def STNT1B_4Z_STRIDED : sme2_st_vector_vg4_multi_scalar_scalar<0b00, 0b1, ZZZZ_b_strided, GPR64shifted8, "stnt1b">;829defm STNT1B_2Z_STRIDED_IMM : sme2_st_vector_vg2_multi_scalar_immediate<0b00, 0b1, ZZ_b_strided, simm4s2, "stnt1b">;830defm STNT1B_4Z_STRIDED_IMM : sme2_st_vector_vg4_multi_scalar_immediate<0b00, 0b1, ZZZZ_b_strided, simm4s4, "stnt1b">;831def STNT1H_2Z_STRIDED : sme2_st_vector_vg2_multi_scalar_scalar<0b01, 0b1, ZZ_h_strided, GPR64shifted16, "stnt1h">;832def STNT1H_4Z_STRIDED : sme2_st_vector_vg4_multi_scalar_scalar<0b01, 0b1, ZZZZ_h_strided, GPR64shifted16, "stnt1h">;833defm STNT1H_2Z_STRIDED_IMM : sme2_st_vector_vg2_multi_scalar_immediate<0b01, 0b1, ZZ_h_strided, simm4s2, "stnt1h">;834defm STNT1H_4Z_STRIDED_IMM : sme2_st_vector_vg4_multi_scalar_immediate<0b01, 0b1, ZZZZ_h_strided, simm4s4, "stnt1h">;835def STNT1W_2Z_STRIDED : sme2_st_vector_vg2_multi_scalar_scalar<0b10, 0b1, ZZ_s_strided, GPR64shifted32, "stnt1w">;836def STNT1W_4Z_STRIDED : sme2_st_vector_vg4_multi_scalar_scalar<0b10, 0b1, ZZZZ_s_strided, GPR64shifted32, "stnt1w">;837defm STNT1W_2Z_STRIDED_IMM : sme2_st_vector_vg2_multi_scalar_immediate<0b10, 0b1, ZZ_s_strided, simm4s2, "stnt1w">;838defm STNT1W_4Z_STRIDED_IMM : sme2_st_vector_vg4_multi_scalar_immediate<0b10, 0b1, ZZZZ_s_strided, simm4s4, "stnt1w">;839def STNT1D_2Z_STRIDED : sme2_st_vector_vg2_multi_scalar_scalar<0b11, 0b1, ZZ_d_strided, GPR64shifted64, "stnt1d">;840def STNT1D_4Z_STRIDED : sme2_st_vector_vg4_multi_scalar_scalar<0b11, 0b1, ZZZZ_d_strided, GPR64shifted64, "stnt1d">;841defm STNT1D_2Z_STRIDED_IMM : sme2_st_vector_vg2_multi_scalar_immediate<0b11, 0b1, ZZ_d_strided, simm4s2, "stnt1d">;842defm STNT1D_4Z_STRIDED_IMM : sme2_st_vector_vg4_multi_scalar_immediate<0b11, 0b1, ZZZZ_d_strided, simm4s4, "stnt1d">;843} // End let Predicates = [HasSME2]844 845 846let Predicates = [HasSME2andIsNonStreamingSafe] in {847defm ZERO_T : sme2_zero_zt<"zero", 0b0001>;848 849defm LDR_TX : sme2_spill_fill_vector<"ldr", 0b01111100, AArch64_restore_zt>;850defm STR_TX : sme2_spill_fill_vector<"str", 0b11111100, AArch64_save_zt>;851} // End let Predicates = [HasSME2andIsNonStreamingSafe]852 853let Predicates = [HasSME2, HasSMEI16I64] in {854defm ADD_VG2_M2ZZ_D : sme2_dot_mla_add_sub_array_vg2_single<"add", 0b1011010, MatrixOp64, ZZ_d, ZPR4b64, nxv2i64, int_aarch64_sme_add_write_single_za_vg1x2>;855defm ADD_VG4_M4ZZ_D : sme2_dot_mla_add_sub_array_vg4_single<"add", 0b1111010, MatrixOp64, ZZZZ_d, ZPR4b64, nxv2i64, int_aarch64_sme_add_write_single_za_vg1x4>;856defm ADD_VG2_M2Z2Z_D : sme2_dot_mla_add_sub_array_vg2_multi<"add", 0b1110010, MatrixOp64, ZZ_d_mul_r, nxv2i64, int_aarch64_sme_add_write_za_vg1x2>;857defm ADD_VG4_M4Z4Z_D : sme2_dot_mla_add_sub_array_vg4_multi<"add", 0b1110010, MatrixOp64, ZZZZ_d_mul_r, nxv2i64, int_aarch64_sme_add_write_za_vg1x4>;858 859defm SUB_VG2_M2ZZ_D : sme2_dot_mla_add_sub_array_vg2_single<"sub", 0b1011011, MatrixOp64, ZZ_d, ZPR4b64, nxv2i64, int_aarch64_sme_sub_write_single_za_vg1x2>;860defm SUB_VG4_M4ZZ_D : sme2_dot_mla_add_sub_array_vg4_single<"sub", 0b1111011, MatrixOp64, ZZZZ_d, ZPR4b64, nxv2i64, int_aarch64_sme_sub_write_single_za_vg1x4>;861defm SUB_VG2_M2Z2Z_D : sme2_dot_mla_add_sub_array_vg2_multi<"sub", 0b1110011, MatrixOp64, ZZ_d_mul_r, nxv2i64, int_aarch64_sme_sub_write_za_vg1x2>;862defm SUB_VG4_M4Z4Z_D : sme2_dot_mla_add_sub_array_vg4_multi<"sub", 0b1110011, MatrixOp64, ZZZZ_d_mul_r, nxv2i64, int_aarch64_sme_sub_write_za_vg1x4>;863 864defm ADD_VG2_M2Z_D : sme2_multivec_accum_add_sub_vg2<"add", 0b1010, MatrixOp64, ZZ_d_mul_r, nxv2i64, int_aarch64_sme_add_za64_vg1x2>;865defm ADD_VG4_M4Z_D : sme2_multivec_accum_add_sub_vg4<"add", 0b1010, MatrixOp64, ZZZZ_d_mul_r, nxv2i64, int_aarch64_sme_add_za64_vg1x4>;866 867defm SUB_VG2_M2Z_D : sme2_multivec_accum_add_sub_vg2<"sub", 0b1011, MatrixOp64, ZZ_d_mul_r, nxv2i64, int_aarch64_sme_sub_za64_vg1x2>;868defm SUB_VG4_M4Z_D : sme2_multivec_accum_add_sub_vg4<"sub", 0b1011, MatrixOp64, ZZZZ_d_mul_r, nxv2i64, int_aarch64_sme_sub_za64_vg1x4>;869 870defm SDOT_VG2_M2ZZI_HtoD : sme2_multi_vec_array_vg2_index_64b<"sdot", 0b01, ZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_sdot_lane_za64_vg1x2>;871defm SDOT_VG4_M4ZZI_HtoD : sme2_multi_vec_array_vg4_index_64b<"sdot", 0b001, ZZZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_sdot_lane_za64_vg1x4>;872defm SDOT_VG2_M2ZZ_HtoD : sme2_dot_mla_add_sub_array_vg2_single<"sdot", 0b1010100, MatrixOp64, ZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_sdot_single_za64_vg1x2>;873defm SDOT_VG4_M4ZZ_HtoD : sme2_dot_mla_add_sub_array_vg4_single<"sdot", 0b1110100, MatrixOp64, ZZZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_sdot_single_za64_vg1x4>;874defm SDOT_VG2_M2Z2Z_HtoD : sme2_dot_mla_add_sub_array_vg2_multi<"sdot", 0b1101000, MatrixOp64, ZZ_h_mul_r, nxv8i16, int_aarch64_sme_sdot_za64_vg1x2>;875defm SDOT_VG4_M4Z4Z_HtoD : sme2_dot_mla_add_sub_array_vg4_multi<"sdot", 0b1101000, MatrixOp64, ZZZZ_h_mul_r, nxv8i16, int_aarch64_sme_sdot_za64_vg1x4>;876 877defm SVDOT_VG4_M4ZZI_HtoD : sme2_multi_vec_array_vg4_index_64b<"svdot", 0b101, ZZZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_svdot_lane_za64_vg1x4>;878 879defm UDOT_VG2_M2ZZI_HtoD : sme2_multi_vec_array_vg2_index_64b<"udot", 0b11, ZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_udot_lane_za64_vg1x2>;880defm UDOT_VG4_M4ZZI_HtoD : sme2_multi_vec_array_vg4_index_64b<"udot", 0b011, ZZZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_udot_lane_za64_vg1x4>;881defm UDOT_VG2_M2ZZ_HtoD : sme2_dot_mla_add_sub_array_vg2_single<"udot", 0b1010110, MatrixOp64, ZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_udot_single_za64_vg1x2>;882defm UDOT_VG4_M4ZZ_HtoD : sme2_dot_mla_add_sub_array_vg4_single<"udot", 0b1110110, MatrixOp64, ZZZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_udot_single_za64_vg1x4>;883defm UDOT_VG2_M2Z2Z_HtoD : sme2_dot_mla_add_sub_array_vg2_multi<"udot", 0b1101010, MatrixOp64, ZZ_h_mul_r, nxv8i16, int_aarch64_sme_udot_za64_vg1x2>;884defm UDOT_VG4_M4Z4Z_HtoD : sme2_dot_mla_add_sub_array_vg4_multi<"udot", 0b1101010, MatrixOp64, ZZZZ_h_mul_r, nxv8i16, int_aarch64_sme_udot_za64_vg1x4>;885 886defm UVDOT_VG4_M4ZZI_HtoD : sme2_multi_vec_array_vg4_index_64b<"uvdot", 0b111, ZZZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_uvdot_lane_za64_vg1x4>;887 888defm SMLALL_MZZI_HtoD : sme2_mla_ll_array_index_64b<"smlall", 0b00, int_aarch64_sme_smla_za64_lane_vg4x1>;889defm SMLALL_VG2_M2ZZI_HtoD : sme2_mla_ll_array_vg2_index_64b<"smlall", 0b00, int_aarch64_sme_smla_za64_lane_vg4x2>;890defm SMLALL_VG4_M4ZZI_HtoD : sme2_mla_ll_array_vg4_index_64b<"smlall", 0b00, int_aarch64_sme_smla_za64_lane_vg4x4>;891defm SMLALL_MZZ_HtoD : sme2_mla_ll_array_single<"smlall", 0b10000, MatrixOp64, ZPR16, ZPR4b16, nxv8i16, int_aarch64_sme_smla_za64_single_vg4x1>;892defm SMLALL_VG2_M2ZZ_HtoD : sme2_mla_ll_array_vg2_single<"smlall", 0b100000, MatrixOp64, ZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_smla_za64_single_vg4x2>;893defm SMLALL_VG4_M4ZZ_HtoD : sme2_mla_ll_array_vg4_single<"smlall", 0b110000, MatrixOp64, ZZZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_smla_za64_single_vg4x4>;894defm SMLALL_VG2_M2Z2Z_HtoD : sme2_mla_ll_array_vg2_multi<"smlall", 0b10000, MatrixOp64, ZZ_h_mul_r, nxv8i16, int_aarch64_sme_smla_za64_vg4x2>;895defm SMLALL_VG4_M4Z4Z_HtoD : sme2_mla_ll_array_vg4_multi<"smlall", 0b10000, MatrixOp64, ZZZZ_h_mul_r, nxv8i16, int_aarch64_sme_smla_za64_vg4x4>;896 897defm SMLSLL_MZZI_HtoD : sme2_mla_ll_array_index_64b<"smlsll", 0b01, int_aarch64_sme_smls_za64_lane_vg4x1>;898defm SMLSLL_VG2_M2ZZI_HtoD : sme2_mla_ll_array_vg2_index_64b<"smlsll", 0b01, int_aarch64_sme_smls_za64_lane_vg4x2>;899defm SMLSLL_VG4_M4ZZI_HtoD : sme2_mla_ll_array_vg4_index_64b<"smlsll", 0b01, int_aarch64_sme_smls_za64_lane_vg4x4>;900defm SMLSLL_MZZ_HtoD : sme2_mla_ll_array_single<"smlsll", 0b10010, MatrixOp64, ZPR16, ZPR4b16, nxv8i16, int_aarch64_sme_smls_za64_single_vg4x1>;901defm SMLSLL_VG2_M2ZZ_HtoD : sme2_mla_ll_array_vg2_single<"smlsll", 0b100100, MatrixOp64, ZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_smls_za64_single_vg4x2>;902defm SMLSLL_VG4_M4ZZ_HtoD : sme2_mla_ll_array_vg4_single<"smlsll", 0b110100, MatrixOp64, ZZZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_smls_za64_single_vg4x4>;903defm SMLSLL_VG2_M2Z2Z_HtoD : sme2_mla_ll_array_vg2_multi<"smlsll", 0b10010, MatrixOp64, ZZ_h_mul_r, nxv8i16, int_aarch64_sme_smls_za64_vg4x2>;904defm SMLSLL_VG4_M4Z4Z_HtoD : sme2_mla_ll_array_vg4_multi<"smlsll", 0b10010, MatrixOp64, ZZZZ_h_mul_r, nxv8i16, int_aarch64_sme_smls_za64_vg4x4>;905 906defm UMLALL_MZZI_HtoD : sme2_mla_ll_array_index_64b<"umlall", 0b10, int_aarch64_sme_umla_za64_lane_vg4x1>;907defm UMLALL_VG2_M2ZZI_HtoD : sme2_mla_ll_array_vg2_index_64b<"umlall", 0b10, int_aarch64_sme_umla_za64_lane_vg4x2>;908defm UMLALL_VG4_M4ZZI_HtoD : sme2_mla_ll_array_vg4_index_64b<"umlall", 0b10, int_aarch64_sme_umla_za64_lane_vg4x4>;909defm UMLALL_MZZ_HtoD : sme2_mla_ll_array_single<"umlall", 0b10100, MatrixOp64, ZPR16, ZPR4b16, nxv8i16, int_aarch64_sme_umla_za64_single_vg4x1>;910defm UMLALL_VG2_M2ZZ_HtoD : sme2_mla_ll_array_vg2_single<"umlall", 0b101000, MatrixOp64, ZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_umla_za64_single_vg4x2>;911defm UMLALL_VG4_M4ZZ_HtoD : sme2_mla_ll_array_vg4_single<"umlall", 0b111000, MatrixOp64, ZZZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_umla_za64_single_vg4x4>;912defm UMLALL_VG2_M2Z2Z_HtoD : sme2_mla_ll_array_vg2_multi<"umlall", 0b10100, MatrixOp64, ZZ_h_mul_r, nxv8i16, int_aarch64_sme_umla_za64_vg4x2>;913defm UMLALL_VG4_M4Z4Z_HtoD : sme2_mla_ll_array_vg4_multi<"umlall", 0b10100, MatrixOp64, ZZZZ_h_mul_r, nxv8i16, int_aarch64_sme_umla_za64_vg4x4>;914 915defm UMLSLL_MZZI_HtoD : sme2_mla_ll_array_index_64b<"umlsll", 0b11, int_aarch64_sme_umls_za64_lane_vg4x1>;916defm UMLSLL_VG2_M2ZZI_HtoD : sme2_mla_ll_array_vg2_index_64b<"umlsll", 0b11, int_aarch64_sme_umls_za64_lane_vg4x2>;917defm UMLSLL_VG4_M4ZZI_HtoD : sme2_mla_ll_array_vg4_index_64b<"umlsll", 0b11, int_aarch64_sme_umls_za64_lane_vg4x4>;918defm UMLSLL_MZZ_HtoD : sme2_mla_ll_array_single<"umlsll", 0b10110, MatrixOp64, ZPR16, ZPR4b16, nxv8i16, int_aarch64_sme_umls_za64_single_vg4x1>;919defm UMLSLL_VG2_M2ZZ_HtoD : sme2_mla_ll_array_vg2_single<"umlsll", 0b101100, MatrixOp64, ZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_umls_za64_single_vg4x2>;920defm UMLSLL_VG4_M4ZZ_HtoD : sme2_mla_ll_array_vg4_single<"umlsll", 0b111100, MatrixOp64, ZZZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_umls_za64_single_vg4x4>;921defm UMLSLL_VG2_M2Z2Z_HtoD : sme2_mla_ll_array_vg2_multi<"umlsll", 0b10110, MatrixOp64, ZZ_h_mul_r, nxv8i16, int_aarch64_sme_umls_za64_vg4x2>;922defm UMLSLL_VG4_M4Z4Z_HtoD : sme2_mla_ll_array_vg4_multi<"umlsll", 0b10110, MatrixOp64, ZZZZ_h_mul_r, nxv8i16, int_aarch64_sme_umls_za64_vg4x4>;923}924 925let Predicates = [HasSME2, HasSMEF64F64] in {926defm FMLA_VG2_M2ZZI_D : sme2_multi_vec_array_vg2_index_64b<"fmla", 0b00, ZZ_d_mul_r, ZPR4b64, nxv2f64, int_aarch64_sme_fmla_lane_vg1x2>;927defm FMLA_VG4_M4ZZI_D : sme2_multi_vec_array_vg4_index_64b<"fmla", 0b000, ZZZZ_d_mul_r, ZPR4b64, nxv2f64, int_aarch64_sme_fmla_lane_vg1x4>;928defm FMLA_VG2_M2ZZ_D : sme2_dot_mla_add_sub_array_vg2_single<"fmla", 0b1011000, MatrixOp64, ZZ_d, ZPR4b64, nxv2f64, int_aarch64_sme_fmla_single_vg1x2>;929defm FMLA_VG4_M4ZZ_D : sme2_dot_mla_add_sub_array_vg4_single<"fmla", 0b1111000, MatrixOp64, ZZZZ_d, ZPR4b64, nxv2f64, int_aarch64_sme_fmla_single_vg1x4>;930defm FMLA_VG2_M2Z2Z_D : sme2_dot_mla_add_sub_array_vg2_multi<"fmla", 0b1110000, MatrixOp64, ZZ_d_mul_r, nxv2f64, int_aarch64_sme_fmla_vg1x2>;931defm FMLA_VG4_M4Z4Z_D : sme2_dot_mla_add_sub_array_vg4_multi<"fmla", 0b1110000, MatrixOp64, ZZZZ_d_mul_r, nxv2f64, int_aarch64_sme_fmla_vg1x4>;932 933defm FMLS_VG2_M2ZZI_D : sme2_multi_vec_array_vg2_index_64b<"fmls", 0b10, ZZ_d_mul_r, ZPR4b64, nxv2f64, int_aarch64_sme_fmls_lane_vg1x2>;934defm FMLS_VG4_M4ZZI_D : sme2_multi_vec_array_vg4_index_64b<"fmls", 0b010, ZZZZ_d_mul_r, ZPR4b64, nxv2f64, int_aarch64_sme_fmls_lane_vg1x4>;935defm FMLS_VG2_M2ZZ_D : sme2_dot_mla_add_sub_array_vg2_single<"fmls", 0b1011001, MatrixOp64, ZZ_d, ZPR4b64, nxv2f64, int_aarch64_sme_fmls_single_vg1x2>;936defm FMLS_VG4_M4ZZ_D : sme2_dot_mla_add_sub_array_vg4_single<"fmls", 0b1111001, MatrixOp64, ZZZZ_d, ZPR4b64, nxv2f64, int_aarch64_sme_fmls_single_vg1x4>;937defm FMLS_VG2_M2Z2Z_D : sme2_dot_mla_add_sub_array_vg2_multi<"fmls", 0b1110001, MatrixOp64, ZZ_d_mul_r, nxv2f64, int_aarch64_sme_fmls_vg1x2>;938defm FMLS_VG4_M4Z4Z_D : sme2_dot_mla_add_sub_array_vg4_multi<"fmls", 0b1110001, MatrixOp64, ZZZZ_d_mul_r, nxv2f64, int_aarch64_sme_fmls_vg1x4>;939 940defm FADD_VG2_M2Z_D : sme2_multivec_accum_add_sub_vg2<"fadd", 0b1000, MatrixOp64, ZZ_d_mul_r, nxv2f64, int_aarch64_sme_add_za64_vg1x2>;941defm FADD_VG4_M4Z_D : sme2_multivec_accum_add_sub_vg4<"fadd", 0b1000, MatrixOp64, ZZZZ_d_mul_r, nxv2f64, int_aarch64_sme_add_za64_vg1x4>;942 943defm FSUB_VG2_M2Z_D : sme2_multivec_accum_add_sub_vg2<"fsub", 0b1001, MatrixOp64, ZZ_d_mul_r, nxv2f64, int_aarch64_sme_sub_za64_vg1x2>;944defm FSUB_VG4_M4Z_D : sme2_multivec_accum_add_sub_vg4<"fsub", 0b1001, MatrixOp64, ZZZZ_d_mul_r, nxv2f64, int_aarch64_sme_sub_za64_vg1x4>;945}946 947let Predicates = [HasSME2p1] in {948defm MOVAZ_ZMI : sme2p1_movaz_tile_to_vec<"movaz", int_aarch64_sme_readz_horiz, int_aarch64_sme_readz_vert,949 int_aarch64_sme_readz_q_horiz, int_aarch64_sme_readz_q_vert>;950defm MOVAZ_2ZMI : sme2p1_movaz_tile_to_vec_vg2<"movaz">;951defm MOVAZ_4ZMI : sme2p1_movaz_tile_to_vec_vg4<"movaz">;952defm MOVAZ_VG2_2ZMXI : sme2_movaz_array_to_vec_vg2_multi<"movaz">;953defm MOVAZ_VG4_4ZMXI : sme2_movaz_array_to_vec_vg4_multi<"movaz">;954 955defm ZERO_MXI : sme2p1_zero_matrix<"zero">;956 957defm LUTI2_S_2ZTZI : sme2p1_luti2_vector_vg2_index<"luti2">;958defm LUTI2_S_4ZTZI : sme2p1_luti2_vector_vg4_index<"luti2">;959 960defm LUTI4_S_2ZTZI : sme2p1_luti4_vector_vg2_index<"luti4">;961defm LUTI4_S_4ZTZI : sme2p1_luti4_vector_vg4_index<"luti4">;962}963 964let Predicates = [HasSMEF16F16_or_SMEF8F16] in {965defm FADD_VG2_M2Z_H : sme2_multivec_accum_add_sub_vg2<"fadd", 0b0100, MatrixOp16, ZZ_h_mul_r, nxv8f16, int_aarch64_sme_add_za16_vg1x2>;966defm FADD_VG4_M4Z_H : sme2_multivec_accum_add_sub_vg4<"fadd", 0b0100, MatrixOp16, ZZZZ_h_mul_r, nxv8f16, int_aarch64_sme_add_za16_vg1x4>;967defm FSUB_VG2_M2Z_H : sme2_multivec_accum_add_sub_vg2<"fsub", 0b0101, MatrixOp16, ZZ_h_mul_r, nxv8f16, int_aarch64_sme_sub_za16_vg1x2>;968defm FSUB_VG4_M4Z_H : sme2_multivec_accum_add_sub_vg4<"fsub", 0b0101, MatrixOp16, ZZZZ_h_mul_r, nxv8f16, int_aarch64_sme_sub_za16_vg1x4>;969}970 971let Predicates = [HasSMEF16F16] in {972defm FMLA_VG2_M2ZZI_H : sme2p1_multi_vec_array_vg2_index_16b<"fmla", 0b00, 0b100, ZZ_h_mul_r, ZPR4b16, nxv8f16, int_aarch64_sme_fmla_lane_vg1x2>;973defm FMLA_VG4_M4ZZI_H : sme2p1_multi_vec_array_vg4_index_16b<"fmla", 0b000, ZZZZ_h_mul_r, ZPR4b16, nxv8f16, int_aarch64_sme_fmla_lane_vg1x4>;974defm FMLA_VG2_M2ZZ_H : sme2_dot_mla_add_sub_array_vg2_single<"fmla", 0b0011100, MatrixOp16, ZZ_h, ZPR4b16, nxv8f16, int_aarch64_sme_fmla_single_vg1x2>;975defm FMLA_VG4_M4ZZ_H : sme2_dot_mla_add_sub_array_vg4_single<"fmla", 0b0111100, MatrixOp16, ZZZZ_h, ZPR4b16, nxv8f16, int_aarch64_sme_fmla_single_vg1x4>;976defm FMLA_VG2_M2Z2Z_H : sme2_dot_mla_add_sub_array_vg2_multi<"fmla", 0b0100001, MatrixOp16, ZZ_h_mul_r, nxv8f16, int_aarch64_sme_fmla_vg1x2>;977defm FMLA_VG4_M4Z4Z_H : sme2_dot_mla_add_sub_array_vg4_multi<"fmla", 0b0100001, MatrixOp16, ZZZZ_h_mul_r, nxv8f16, int_aarch64_sme_fmla_vg1x4>;978 979defm FMLS_VG2_M2ZZI_H : sme2p1_multi_vec_array_vg2_index_16b<"fmls", 0b00, 0b101, ZZ_h_mul_r, ZPR4b16, nxv8f16, int_aarch64_sme_fmls_lane_vg1x2>;980defm FMLS_VG4_M4ZZI_H : sme2p1_multi_vec_array_vg4_index_16b<"fmls", 0b001, ZZZZ_h_mul_r, ZPR4b16, nxv8f16, int_aarch64_sme_fmls_lane_vg1x4>;981defm FMLS_VG2_M2ZZ_H : sme2_dot_mla_add_sub_array_vg2_single<"fmls", 0b0011101, MatrixOp16, ZZ_h, ZPR4b16, nxv8f16, int_aarch64_sme_fmls_single_vg1x2>;982defm FMLS_VG4_M4ZZ_H : sme2_dot_mla_add_sub_array_vg4_single<"fmls", 0b0111101, MatrixOp16, ZZZZ_h, ZPR4b16, nxv8f16, int_aarch64_sme_fmls_single_vg1x4>;983defm FMLS_VG2_M2Z2Z_H : sme2_dot_mla_add_sub_array_vg2_multi<"fmls", 0b0100011, MatrixOp16, ZZ_h_mul_r, nxv8f16, int_aarch64_sme_fmls_vg1x2>;984defm FMLS_VG4_M4Z4Z_H : sme2_dot_mla_add_sub_array_vg4_multi<"fmls", 0b0100011, MatrixOp16, ZZZZ_h_mul_r, nxv8f16, int_aarch64_sme_fmls_vg1x4>;985 986defm FCVT_2ZZ_H : sme2p1_fp_cvt_vector_vg2_single<"fcvt", 0b0>;987defm FCVTL_2ZZ_H : sme2p1_fp_cvt_vector_vg2_single<"fcvtl", 0b1>;988 989defm FMOPA_MPPZZ_H : sme2p1_fmop_tile_fp16<"fmopa", 0b0, 0b0, nxv8f16, int_aarch64_sme_mopa>;990defm FMOPS_MPPZZ_H : sme2p1_fmop_tile_fp16<"fmops", 0b0, 0b1, nxv8f16, int_aarch64_sme_mops>;991}992 993// SME2 ZA-targeting non-widening BFloat16 instructions994let Predicates = [HasSMEB16B16] in {995defm BFADD_VG2_M2Z_H : sme2_multivec_accum_add_sub_vg2<"bfadd", 0b1100, MatrixOp16, ZZ_h_mul_r, nxv8bf16, int_aarch64_sme_add_za16_vg1x2>;996defm BFADD_VG4_M4Z_H : sme2_multivec_accum_add_sub_vg4<"bfadd", 0b1100, MatrixOp16, ZZZZ_h_mul_r, nxv8bf16, int_aarch64_sme_add_za16_vg1x4>;997defm BFSUB_VG2_M2Z_H : sme2_multivec_accum_add_sub_vg2<"bfsub", 0b1101, MatrixOp16, ZZ_h_mul_r, nxv8bf16, int_aarch64_sme_sub_za16_vg1x2>;998defm BFSUB_VG4_M4Z_H : sme2_multivec_accum_add_sub_vg4<"bfsub", 0b1101, MatrixOp16, ZZZZ_h_mul_r, nxv8bf16, int_aarch64_sme_sub_za16_vg1x4>;999 1000defm BFMLA_VG2_M2ZZI : sme2p1_multi_vec_array_vg2_index_16b<"bfmla", 0b00, 0b110, ZZ_h_mul_r, ZPR4b16, nxv8bf16, int_aarch64_sme_fmla_lane_vg1x2>;1001defm BFMLA_VG4_M4ZZI : sme2p1_multi_vec_array_vg4_index_16b<"bfmla", 0b010, ZZZZ_h_mul_r, ZPR4b16, nxv8bf16, int_aarch64_sme_fmla_lane_vg1x4>;1002defm BFMLA_VG2_M2ZZ : sme2_dot_mla_add_sub_array_vg2_single<"bfmla", 0b1011100, MatrixOp16, ZZ_h, ZPR4b16, nxv8bf16, int_aarch64_sme_fmla_single_vg1x2>;1003defm BFMLA_VG4_M4ZZ : sme2_dot_mla_add_sub_array_vg4_single<"bfmla", 0b1111100, MatrixOp16, ZZZZ_h, ZPR4b16, nxv8bf16, int_aarch64_sme_fmla_single_vg1x4>;1004defm BFMLA_VG2_M2Z2Z : sme2_dot_mla_add_sub_array_vg2_multi<"bfmla", 0b1100001, MatrixOp16, ZZ_h_mul_r, nxv8bf16, int_aarch64_sme_fmla_vg1x2>;1005defm BFMLA_VG4_M4Z4Z : sme2_dot_mla_add_sub_array_vg4_multi<"bfmla", 0b1100001, MatrixOp16, ZZZZ_h_mul_r, nxv8bf16, int_aarch64_sme_fmla_vg1x4>;1006 1007defm BFMLS_VG2_M2ZZI : sme2p1_multi_vec_array_vg2_index_16b<"bfmls", 0b00, 0b111, ZZ_h_mul_r, ZPR4b16, nxv8bf16, int_aarch64_sme_fmls_lane_vg1x2>;1008defm BFMLS_VG4_M4ZZI : sme2p1_multi_vec_array_vg4_index_16b<"bfmls", 0b011, ZZZZ_h_mul_r, ZPR4b16, nxv8bf16, int_aarch64_sme_fmls_lane_vg1x4>;1009defm BFMLS_VG2_M2ZZ : sme2_dot_mla_add_sub_array_vg2_single<"bfmls", 0b1011101, MatrixOp16, ZZ_h, ZPR4b16, nxv8bf16, int_aarch64_sme_fmls_single_vg1x2>;1010defm BFMLS_VG4_M4ZZ : sme2_dot_mla_add_sub_array_vg4_single<"bfmls", 0b1111101, MatrixOp16, ZZZZ_h, ZPR4b16, nxv8bf16, int_aarch64_sme_fmls_single_vg1x4>;1011defm BFMLS_VG2_M2Z2Z : sme2_dot_mla_add_sub_array_vg2_multi<"bfmls", 0b1100011, MatrixOp16, ZZ_h_mul_r, nxv8bf16, int_aarch64_sme_fmls_vg1x2>;1012defm BFMLS_VG4_M4Z4Z : sme2_dot_mla_add_sub_array_vg4_multi<"bfmls", 0b1100011, MatrixOp16, ZZZZ_h_mul_r, nxv8bf16, int_aarch64_sme_fmls_vg1x4>;1013 1014defm BFMOPA_MPPZZ_H : sme2p1_fmop_tile_fp16<"bfmopa", 0b1, 0b0, nxv8bf16, int_aarch64_sme_mopa>;1015defm BFMOPS_MPPZZ_H : sme2p1_fmop_tile_fp16<"bfmops", 0b1, 0b1, nxv8bf16, int_aarch64_sme_mops>;1016}1017 1018// SME2 Z-targeting non-widening BFloat16 instructions1019let Predicates = [HasSME2, HasSVEB16B16] in {1020defm BFMAX_VG2_2ZZ : sme2p1_bf_max_min_vector_vg2_single<"bfmax", 0b0010000>;1021defm BFMAX_VG4_4ZZ : sme2p1_bf_max_min_vector_vg4_single<"bfmax", 0b0010000>;1022defm BFMAX_VG2_2Z2Z : sme2p1_bf_max_min_vector_vg2_multi<"bfmax", 0b0010000>;1023defm BFMAX_VG4_4Z2Z : sme2p1_bf_max_min_vector_vg4_multi<"bfmax", 0b0010000>;1024 1025defm BFMIN_VG2_2ZZ : sme2p1_bf_max_min_vector_vg2_single<"bfmin", 0b0010001>;1026defm BFMIN_VG4_4ZZ : sme2p1_bf_max_min_vector_vg4_single<"bfmin", 0b0010001>;1027defm BFMIN_VG2_2Z2Z : sme2p1_bf_max_min_vector_vg2_multi<"bfmin", 0b0010001>;1028defm BFMIN_VG4_4Z2Z : sme2p1_bf_max_min_vector_vg4_multi<"bfmin", 0b0010001>;1029 1030defm BFMAXNM_VG2_2ZZ : sme2p1_bf_max_min_vector_vg2_single<"bfmaxnm", 0b0010010>;1031defm BFMAXNM_VG4_4ZZ : sme2p1_bf_max_min_vector_vg4_single<"bfmaxnm", 0b0010010>;1032defm BFMAXNM_VG2_2Z2Z : sme2p1_bf_max_min_vector_vg2_multi<"bfmaxnm", 0b0010010>;1033defm BFMAXNM_VG4_4Z2Z : sme2p1_bf_max_min_vector_vg4_multi<"bfmaxnm", 0b0010010>;1034 1035defm BFMINNM_VG2_2ZZ : sme2p1_bf_max_min_vector_vg2_single<"bfminnm", 0b0010011>;1036defm BFMINNM_VG4_4ZZ : sme2p1_bf_max_min_vector_vg4_single<"bfminnm", 0b0010011>;1037defm BFMINNM_VG2_2Z2Z : sme2p1_bf_max_min_vector_vg2_multi<"bfminnm", 0b0010011>;1038defm BFMINNM_VG4_4Z2Z : sme2p1_bf_max_min_vector_vg4_multi<"bfminnm", 0b0010011>;1039 1040defm BFCLAMP_VG2_2ZZZ: sme2p1_bfclamp_vector_vg2_multi<"bfclamp">;1041defm BFCLAMP_VG4_4ZZZ: sme2p1_bfclamp_vector_vg4_multi<"bfclamp">;1042}1043 1044let Predicates = [HasSME2, HasFP8] in {1045defm F1CVT_2ZZ_BtoH : sme2p1_fp8_cvt_vector_vg2_single<"f1cvt", 0b00, 0b0>;1046defm F1CVTL_2ZZ_BtoH : sme2p1_fp8_cvt_vector_vg2_single<"f1cvtl", 0b00, 0b1>;1047defm BF1CVT_2ZZ_BtoH : sme2p1_fp8_cvt_vector_vg2_single<"bf1cvt", 0b01, 0b0>;1048defm BF1CVTL_2ZZ_BtoH : sme2p1_fp8_cvt_vector_vg2_single<"bf1cvtl", 0b01, 0b1>;1049defm F2CVT_2ZZ_BtoH : sme2p1_fp8_cvt_vector_vg2_single<"f2cvt", 0b10, 0b0>;1050defm F2CVTL_2ZZ_BtoH : sme2p1_fp8_cvt_vector_vg2_single<"f2cvtl", 0b10, 0b1>;1051defm BF2CVT_2ZZ_BtoH : sme2p1_fp8_cvt_vector_vg2_single<"bf2cvt", 0b11, 0b0>;1052defm BF2CVTL_2ZZ_BtoH : sme2p1_fp8_cvt_vector_vg2_single<"bf2cvtl", 0b11, 0b1>;1053 1054defm FCVT_Z2Z_HtoB : sme2_fp8_cvt_vg2_single<"fcvt", 0b0, nxv8f16, int_aarch64_sve_fp8_cvt_x2>;1055defm BFCVT_Z2Z_HtoB : sme2_fp8_cvt_vg2_single<"bfcvt", 0b1, nxv8bf16, int_aarch64_sve_fp8_cvt_x2>;1056defm FCVT_Z4Z_StoB : sme2_fp8_cvt_vg4_single<"fcvt", 0b0, int_aarch64_sve_fp8_cvt_x4>;1057defm FCVTN_Z4Z_StoB : sme2_fp8_cvt_vg4_single<"fcvtn", 0b1, int_aarch64_sve_fp8_cvtn_x4>;1058 1059defm FSCALE_2ZZ : sme2_fp_sve_destructive_vector_vg2_single<"fscale", 0b0011000>;1060defm FSCALE_4ZZ : sme2_fp_sve_destructive_vector_vg4_single<"fscale", 0b0011000>;1061defm FSCALE_2Z2Z : sme2_fp_sve_destructive_vector_vg2_multi<"fscale", 0b0011000>;1062defm FSCALE_4Z4Z : sme2_fp_sve_destructive_vector_vg4_multi<"fscale", 0b0011000>;1063}1064 1065let Predicates = [HasSME2, HasFAMINMAX] in {1066defm FAMAX_2Z2Z : sme2_fp_sve_destructive_vector_vg2_multi<"famax", 0b0010100>;1067defm FAMIN_2Z2Z : sme2_fp_sve_destructive_vector_vg2_multi<"famin", 0b0010101>;1068 1069defm FAMAX_4Z4Z : sme2_fp_sve_destructive_vector_vg4_multi<"famax", 0b0010100>;1070defm FAMIN_4Z4Z : sme2_fp_sve_destructive_vector_vg4_multi<"famin", 0b0010101>;1071}1072 1073let Predicates = [HasSME_LUTv2] in {1074defm MOVT_TIZ : sme2_movt_zt_to_zt<"movt", 0b0011111, int_aarch64_sme_write_lane_zt, int_aarch64_sme_write_zt>;1075def LUTI4_4ZZT2Z : sme2_luti4_vector_vg4<0b00, 0b00,"luti4">;1076}1077 1078let Predicates = [HasSME2p1, HasSME_LUTv2] in {1079def LUTI4_S_4ZZT2Z : sme2_luti4_vector_vg4_strided<0b00, 0b00, "luti4">;1080}1081 1082let Predicates = [HasSMEF8F16] in {1083defm FVDOT_VG2_M2ZZI_BtoH : sme2_fp8_fdot_index_za16_vg1x2<"fvdot", 0b110, int_aarch64_sme_fp8_fvdot_lane_za16_vg1x2>;1084defm FDOT_VG2_M2ZZI_BtoH : sme2_fp8_fdot_index_za16_vg1x2<"fdot", 0b010, int_aarch64_sme_fp8_fdot_lane_za16_vg1x2>;1085defm FDOT_VG4_M4ZZI_BtoH : sme2_fp8_fdot_index_za16_vg1x4<"fdot", int_aarch64_sme_fp8_fdot_lane_za16_vg1x4>;1086 1087defm FDOT_VG2_M2ZZ_BtoH : sme2_fp8_fdot_single_vg1x2<"fdot", 0b0010001, MatrixOp16, int_aarch64_sme_fp8_fdot_single_za16_vg1x2>;1088defm FDOT_VG4_M4ZZ_BtoH : sme2_fp8_fdot_single_vg1x4<"fdot", 0b0110001, MatrixOp16, int_aarch64_sme_fp8_fdot_single_za16_vg1x4>;1089defm FDOT_VG2_M2Z2Z_BtoH : sme2_fp8_fdot_multi_vg1x2 <"fdot", 0b0100100, MatrixOp16, int_aarch64_sme_fp8_fdot_multi_za16_vg1x2>;1090defm FDOT_VG4_M4Z4Z_BtoH : sme2_fp8_fdot_multi_vg1x4 <"fdot", 0b0100100, MatrixOp16, int_aarch64_sme_fp8_fdot_multi_za16_vg1x4>;1091 1092defm FMLAL_MZZI_BtoH : sme2_fp8_fmlal_index_za16<"fmlal", int_aarch64_sme_fp8_fmlal_lane_za16_vg2x1>;1093defm FMLAL_VG2_M2ZZI_BtoH : sme2_fp8_fmlal_index_za16_vgx2<"fmlal", int_aarch64_sme_fp8_fmlal_lane_za16_vg2x2>;1094defm FMLAL_VG4_M4ZZI_BtoH : sme2_fp8_fmlal_index_za16_vgx4<"fmlal", int_aarch64_sme_fp8_fmlal_lane_za16_vg2x4>;1095 1096defm FMLAL_VG2_MZZ_BtoH : sme2_fp8_fmlal_single_za16<"fmlal", int_aarch64_sme_fp8_fmlal_single_za16_vg2x1>;1097defm FMLAL_VG2_M2ZZ_BtoH : sme2_fp_mla_long_array_vg2_single<"fmlal", 0b001, MatrixOp16, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_fp8_fmlal_single_za16_vg2x2, [FPMR, FPCR]>;1098defm FMLAL_VG4_M4ZZ_BtoH : sme2_fp_mla_long_array_vg4_single<"fmlal", 0b001, MatrixOp16, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_fp8_fmlal_single_za16_vg2x4, [FPMR, FPCR]>;1099 1100defm FMLAL_VG2_M2Z2Z_BtoH : sme2_fp_mla_long_array_vg2_multi<"fmlal", 0b100, MatrixOp16, ZZ_b_mul_r, nxv16i8, int_aarch64_sme_fp8_fmlal_multi_za16_vg2x2, [FPMR, FPCR]>;1101defm FMLAL_VG4_M4Z4Z_BtoH : sme2_fp_mla_long_array_vg4_multi<"fmlal", 0b100, MatrixOp16, ZZZZ_b_mul_r, nxv16i8, int_aarch64_sme_fp8_fmlal_multi_za16_vg2x4, [FPMR, FPCR]>;1102 1103defm FMOPA_MPPZZ_BtoH : sme2_fp8_fmopa_za16<"fmopa", int_aarch64_sme_fp8_fmopa_za16>;1104}1105 1106let Predicates = [HasSMEF8F32] in {1107defm FDOT_VG2_M2ZZI_BtoS : sme2_fp8_fdot_index_za32_vg1x2<"fdot", int_aarch64_sme_fp8_fdot_lane_za32_vg1x2>;1108defm FDOT_VG4_M4ZZI_BtoS : sme2_fp8_fdot_index_za32_vg1x4<"fdot", int_aarch64_sme_fp8_fdot_lane_za32_vg1x4>;1109 1110defm FDOT_VG2_M2ZZ_BtoS : sme2_fp8_fdot_single_vg1x2<"fdot", 0b0010011, MatrixOp32, int_aarch64_sme_fp8_fdot_single_za32_vg1x2>;1111defm FDOT_VG4_M4ZZ_BtoS : sme2_fp8_fdot_single_vg1x4<"fdot", 0b0110011, MatrixOp32, int_aarch64_sme_fp8_fdot_single_za32_vg1x4>;1112defm FDOT_VG2_M2Z2Z_BtoS : sme2_fp8_fdot_multi_vg1x2 <"fdot", 0b0100110, MatrixOp32, int_aarch64_sme_fp8_fdot_multi_za32_vg1x2>;1113defm FDOT_VG4_M4Z4Z_BtoS : sme2_fp8_fdot_multi_vg1x4 <"fdot", 0b0100110, MatrixOp32, int_aarch64_sme_fp8_fdot_multi_za32_vg1x4>;1114 1115defm FVDOTB_VG4_M2ZZI_BtoS : sme2_fp8_fdotv_index_za32_vg1x4<"fvdotb", 0b0, int_aarch64_sme_fp8_fvdotb_lane_za32_vg1x4>;1116defm FVDOTT_VG4_M2ZZI_BtoS : sme2_fp8_fdotv_index_za32_vg1x4<"fvdott", 0b1, int_aarch64_sme_fp8_fvdott_lane_za32_vg1x4>;1117 1118defm FMLALL_MZZI_BtoS : sme2_mla_ll_array_index_32b<"fmlall", 0b01, 0b000, int_aarch64_sme_fp8_fmlall_lane_za32_vg4x1, [FPMR, FPCR]>;1119defm FMLALL_VG2_M2ZZI_BtoS : sme2_mla_ll_array_vg2_index_32b<"fmlall", 0b10, 0b100, int_aarch64_sme_fp8_fmlall_lane_za32_vg4x2, [FPMR, FPCR]>;1120defm FMLALL_VG4_M4ZZI_BtoS : sme2_mla_ll_array_vg4_index_32b<"fmlall", 0b00, 0b1000, int_aarch64_sme_fp8_fmlall_lane_za32_vg4x4, [FPMR, FPCR]>;1121 1122defm FMLALL_MZZ_BtoS : sme2_mla_ll_array_single<"fmlall", 0b01000, MatrixOp32, ZPR8, ZPR4b8, nxv16i8, int_aarch64_sme_fp8_fmlall_single_za32_vg4x1, [FPMR, FPCR]>;1123defm FMLALL_VG2_M2ZZ_BtoS : sme2_mla_ll_array_vg2_single<"fmlall", 0b000001, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_fp8_fmlall_single_za32_vg4x2, [FPMR, FPCR]>;1124defm FMLALL_VG4_M4ZZ_BtoS : sme2_mla_ll_array_vg4_single<"fmlall", 0b010001, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_fp8_fmlall_single_za32_vg4x4, [FPMR, FPCR]>;1125 1126defm FMLALL_VG2_M2Z2Z_BtoS : sme2_mla_ll_array_vg2_multi<"fmlall", 0b01000, MatrixOp32, ZZ_b_mul_r, nxv16i8, int_aarch64_sme_fp8_fmlall_multi_za32_vg4x2, [FPMR, FPCR]>;1127defm FMLALL_VG4_M4Z4Z_BtoS : sme2_mla_ll_array_vg4_multi<"fmlall", 0b01000, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, int_aarch64_sme_fp8_fmlall_multi_za32_vg4x4, [FPMR, FPCR]>;1128 1129defm FMOPA_MPPZZ_BtoS : sme2_fp8_fmopa_za32<"fmopa", int_aarch64_sme_fp8_fmopa_za32>;1130}1131 1132let Predicates = [HasSME2, HasSVEBFSCALE] in {1133 defm BFSCALE : sme2_bfscale_single<"bfscale">;1134 defm BFSCALE : sme2_bfscale_multi<"bfscale">;1135}1136 1137let Predicates = [HasSME_MOP4] in {1138 defm BFMOP4A : sme2_bfmop4as_widening<0, "bfmop4a", "int_aarch64_sme_mop4a_wide">;1139 defm BFMOP4S : sme2_bfmop4as_widening<1, "bfmop4s", "int_aarch64_sme_mop4s_wide">;1140 1141 defm FMOP4A : sme2_fmop4as_fp16_fp32_widening<0, "fmop4a", "int_aarch64_sme_mop4a_wide">;1142 defm FMOP4S : sme2_fmop4as_fp16_fp32_widening<1, "fmop4s", "int_aarch64_sme_mop4s_wide">;1143 1144 defm FMOP4A : sme2_fmop4as_fp32_non_widening<0, "fmop4a", "int_aarch64_sme_mop4a">;1145 defm FMOP4S : sme2_fmop4as_fp32_non_widening<1, "fmop4s", "int_aarch64_sme_mop4s">;1146}1147 1148let Predicates = [HasSME2p2] in {1149 defm FMUL_2ZZ : sme2_multi2_fmul_sm<"fmul">;1150 defm FMUL_2Z2Z : sme2_multi2_fmul_mm< "fmul">;1151 defm FMUL_4ZZ : sme2_multi4_fmul_sm<"fmul">;1152 defm FMUL_4Z4Z : sme2_multi4_fmul_mm< "fmul">;1153 1154} // [HasSME2p2]1155 1156let Predicates = [HasSME_MOP4, HasSMEF8F16] in {1157 defm FMOP4A : sme2_fmop4a_fp8_fp16_2way<"fmop4a", "int_aarch64_sme_fp8_fmop4a_za16">;1158}1159 1160let Predicates = [HasSME_MOP4, HasSMEF16F16] in {1161 defm FMOP4A : sme2_fmop4as_fp16_non_widening<0, "fmop4a", "int_aarch64_sme_mop4a">;1162 defm FMOP4S : sme2_fmop4as_fp16_non_widening<1, "fmop4s", "int_aarch64_sme_mop4s">;1163}1164 1165let Predicates = [HasSME2, HasSVEBFSCALE] in {1166 defm BFMUL : sme2_bfmul_single<"bfmul">;1167 defm BFMUL : sme2_bfmul_multi<"bfmul">;1168}1169 1170let Predicates = [HasSME_MOP4, HasSMEF8F32] in {1171 defm FMOP4A : sme2_fmop4a_fp8_fp32_4way<"fmop4a", "int_aarch64_sme_fp8_fmop4a_za32">;1172}1173 1174let Predicates = [HasSME_MOP4, HasSMEB16B16] in {1175 defm BFMOP4A : sme2_bfmop4as_non_widening<0, "bfmop4a", "int_aarch64_sme_mop4a">;1176 defm BFMOP4S : sme2_bfmop4as_non_widening<1, "bfmop4s", "int_aarch64_sme_mop4s">;1177}1178 1179let Predicates = [HasSME_MOP4, HasSMEF64F64] in {1180 defm FMOP4A : sme2_fmop4as_fp64_non_widening<0, "fmop4a", "int_aarch64_sme_mop4a">;1181 defm FMOP4S : sme2_fmop4as_fp64_non_widening<1, "fmop4s", "int_aarch64_sme_mop4s">;1182}1183 1184//===----------------------------------------------------------------------===//1185// SME2.3 instructions1186//===----------------------------------------------------------------------===//1187let Predicates = [HasSME2p3] in {1188 def LUTI6_ZTZ : sme2_lut_single<"luti6">;1189 def LUTI6_4ZT3Z : sme2_luti6_zt_consecutive<"luti6">;1190 def LUTI6_S_4ZT3Z : sme2_luti6_zt_strided<"luti6">;1191 def LUTI6_4Z2Z2ZI : sme2_luti6_vector_vg4_consecutive<"luti6">;1192 def LUTI6_S_4Z2Z2ZI : sme2_luti6_vector_vg4_strided<"luti6">;1193} // [HasSME2p3]1194