lib/Target/AArch64/AArch64SchedExynosM3.td

   1 //=- AArch64SchedExynosM3.td - Samsung Exynos M3 Sched Defs --*- 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-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9 // This file defines the machine model for the Samsung Exynos M3 to support
  10 // instruction scheduling and other instruction cost heuristics.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 //===----------------------------------------------------------------------===//
  15 // The Exynos-M3 is an advanced superscalar microprocessor with a 6-wide
  16 // in-order stage for decode and dispatch and a wider issue stage.
  17 // The execution units and loads and stores are out-of-order.
  18
  19 def ExynosM3Model : SchedMachineModel {
  20   let IssueWidth            =   6; // Up to 6 uops per cycle.
  21   let MicroOpBufferSize     = 228; // ROB size.
  22   let LoopMicroOpBufferSize =  40; // Based on the instruction queue size.
  23   let LoadLatency           =   4; // Optimistic load cases.
  24   let MispredictPenalty     =  16; // Minimum branch misprediction penalty.
  25   let CompleteModel         =   1; // Use the default model otherwise.
  26
  27   list<Predicate> UnsupportedFeatures = SVEUnsupported.F;
  28 }
  29
  30 //===----------------------------------------------------------------------===//
  31 // Define each kind of processor resource and number available on the Exynos-M3,
  32 // which has 12 pipelines, each with its own queue with out-of-order dispatch.
  33
  34 let SchedModel = ExynosM3Model in {
  35
  36 def M3UnitA  : ProcResource<2>; // Simple integer
  37 def M3UnitC  : ProcResource<2>; // Simple and complex integer
  38 def M3UnitD  : ProcResource<1>; // Integer division (inside C0, serialized)
  39 def M3UnitB  : ProcResource<2>; // Branch
  40 def M3UnitL  : ProcResource<2>; // Load
  41 def M3UnitS  : ProcResource<1>; // Store
  42 def M3PipeF0 : ProcResource<1>; // FP #0
  43 let Super = M3PipeF0 in {
  44   def M3UnitFMAC0 : ProcResource<1>; // FP multiplication
  45   def M3UnitFADD0 : ProcResource<1>; // Simple FP
  46   def M3UnitFCVT0 : ProcResource<1>; // FP conversion
  47   def M3UnitFSQR  : ProcResource<2>; // FP square root (serialized)
  48   def M3UnitNALU0 : ProcResource<1>; // Simple vector
  49   def M3UnitNMSC  : ProcResource<1>; // FP and vector miscellanea
  50   def M3UnitNSHT0 : ProcResource<1>; // Vector shifting
  51   def M3UnitNSHF0 : ProcResource<1>; // Vector shuffling
  52 }
  53 def M3PipeF1 : ProcResource<1>; // FP #1
  54 let Super = M3PipeF1 in {
  55   def M3UnitFMAC1 : ProcResource<1>; // FP multiplication
  56   def M3UnitFADD1 : ProcResource<1>; // Simple FP
  57   def M3UnitFDIV0 : ProcResource<2>; // FP division (serialized)
  58   def M3UnitFCVT1 : ProcResource<1>; // FP conversion
  59   def M3UnitFST0  : ProcResource<1>; // FP store
  60   def M3UnitNALU1 : ProcResource<1>; // Simple vector
  61   def M3UnitNCRY0 : ProcResource<1>; // Cryptographic
  62   def M3UnitNMUL  : ProcResource<1>; // Vector multiplication
  63   def M3UnitNSHT1 : ProcResource<1>; // Vector shifting
  64   def M3UnitNSHF1 : ProcResource<1>; // Vector shuffling
  65 }
  66 def M3PipeF2 : ProcResource<1>; // FP #2
  67 let Super = M3PipeF2 in {
  68   def M3UnitFMAC2 : ProcResource<1>; // FP multiplication
  69   def M3UnitFADD2 : ProcResource<1>; // Simple FP
  70   def M3UnitFDIV1 : ProcResource<2>; // FP division (serialized)
  71   def M3UnitFST1  : ProcResource<1>; // FP store
  72   def M3UnitNALU2 : ProcResource<1>; // Simple vector
  73   def M3UnitNCRY1 : ProcResource<1>; // Cryptographic
  74   def M3UnitNSHT2 : ProcResource<1>; // Vector shifting
  75   def M3UnitNSHF2 : ProcResource<1>; // Vector shuffling
  76 }
  77
  78
  79 def M3UnitALU  : ProcResGroup<[M3UnitA,
  80                                M3UnitC]>;
  81 def M3UnitFMAC : ProcResGroup<[M3UnitFMAC0,
  82                                M3UnitFMAC1,
  83                                M3UnitFMAC2]>;
  84 def M3UnitFADD : ProcResGroup<[M3UnitFADD0,
  85                                M3UnitFADD1,
  86                                M3UnitFADD2]>;
  87 def M3UnitFDIV : ProcResGroup<[M3UnitFDIV0,
  88                                M3UnitFDIV1]>;
  89 def M3UnitFCVT : ProcResGroup<[M3UnitFCVT0,
  90                                M3UnitFCVT1]>;
  91 def M3UnitFST  : ProcResGroup<[M3UnitFST0,
  92                                M3UnitFST1]>;
  93 def M3UnitNALU : ProcResGroup<[M3UnitNALU0,
  94                                M3UnitNALU1,
  95                                M3UnitNALU2]>;
  96 def M3UnitNCRY : ProcResGroup<[M3UnitNCRY0,
  97                                M3UnitNCRY1]>;
  98 def M3UnitNSHT : ProcResGroup<[M3UnitNSHT0,
  99                                M3UnitNSHT1,
 100                                M3UnitNSHT2]>;
 101 def M3UnitNSHF : ProcResGroup<[M3UnitNSHF0,
 102                                M3UnitNSHF1,
 103                                M3UnitNSHF2]>;
 104
 105 //===----------------------------------------------------------------------===//
 106 // Coarse scheduling model.
 107
 108 def M3WriteZ0 : SchedWriteRes<[]> { let Latency = 0;
 109                                     let NumMicroOps = 1; }
 110 def M3WriteZ1 : SchedWriteRes<[]> { let Latency = 1;
 111                                     let NumMicroOps = 0; }
 112
 113 def M3WriteA1 : SchedWriteRes<[M3UnitALU]> { let Latency = 1; }
 114 def M3WriteAA : SchedWriteRes<[M3UnitALU]> { let Latency = 2;
 115                                              let ResourceCycles = [2]; }
 116 def M3WriteAB : SchedWriteRes<[M3UnitALU,
 117                                M3UnitC]>   { let Latency = 1;
 118                                              let NumMicroOps = 2; }
 119 def M3WriteAC : SchedWriteRes<[M3UnitALU,
 120                                M3UnitALU,
 121                                M3UnitC]>   { let Latency = 2;
 122                                              let NumMicroOps = 3; }
 123 def M3WriteAD : SchedWriteRes<[M3UnitALU,
 124                                M3UnitC]>   { let Latency = 2;
 125                                              let NumMicroOps = 2; }
 126 def M3WriteC1 : SchedWriteRes<[M3UnitC]>   { let Latency = 1; }
 127 def M3WriteC2 : SchedWriteRes<[M3UnitC]>   { let Latency = 2; }
 128 def M3WriteAU : SchedWriteVariant<[SchedVar<IsCopyIdiomPred, [M3WriteZ0]>,
 129                                    SchedVar<ExynosArithPred, [M3WriteA1]>,
 130                                    SchedVar<ExynosLogicPred, [M3WriteA1]>,
 131                                    SchedVar<NoSchedPred,     [M3WriteAA]>]>;
 132 def M3WriteAV : SchedWriteVariant<[SchedVar<IsCopyIdiomPred, [M3WriteZ0]>,
 133                                    SchedVar<NoSchedPred,     [M3WriteAA]>]>;
 134 def M3WriteAW : SchedWriteVariant<[SchedVar<IsZeroIdiomPred, [M3WriteZ0]>,
 135                                    SchedVar<NoSchedPred,     [M3WriteAA]>]>;
 136 def M3WriteAX : SchedWriteVariant<[SchedVar<ExynosArithPred, [M3WriteA1]>,
 137                                    SchedVar<ExynosLogicPred, [M3WriteA1]>,
 138                                    SchedVar<NoSchedPred,     [M3WriteAA]>]>;
 139 def M3WriteAY : SchedWriteVariant<[SchedVar<ExynosRotateRightImmPred, [M3WriteA1]>,
 140                                    SchedVar<NoSchedPred,              [M3WriteAA]>]>;
 141
 142 def M3WriteB1 : SchedWriteRes<[M3UnitB]> { let Latency = 1; }
 143 def M3WriteBX : SchedWriteVariant<[SchedVar<ExynosBranchLinkLRPred, [M3WriteAC]>,
 144                                    SchedVar<NoSchedPred,            [M3WriteAB]>]>;
 145
 146 def M3WriteL4 : SchedWriteRes<[M3UnitL]> { let Latency = 4; }
 147 def M3WriteL5 : SchedWriteRes<[M3UnitL]> { let Latency = 5; }
 148 def M3WriteLA : SchedWriteRes<[M3UnitL,
 149                                M3UnitL]> { let Latency = 5;
 150                                            let NumMicroOps = 1; }
 151 def M3WriteLB : SchedWriteRes<[M3UnitA,
 152                                M3UnitL]> { let Latency = 5;
 153                                            let NumMicroOps = 2; }
 154 def M3WriteLC : SchedWriteRes<[M3UnitA,
 155                                M3UnitL,
 156                                M3UnitL]> { let Latency = 5;
 157                                            let NumMicroOps = 2; }
 158 def M3WriteLD : SchedWriteRes<[M3UnitA,
 159                                M3UnitL]> { let Latency = 4;
 160                                            let NumMicroOps = 2; }
 161 def M3WriteLE : SchedWriteRes<[M3UnitA,
 162                                M3UnitL]> { let Latency = 6;
 163                                            let NumMicroOps = 2; }
 164 def M3WriteLH : SchedWriteRes<[]>        { let Latency = 5;
 165                                            let NumMicroOps = 0; }
 166 def M3WriteLX : SchedWriteVariant<[SchedVar<ExynosScaledIdxPred, [M3WriteL5]>,
 167                                    SchedVar<NoSchedPred,         [M3WriteL4]>]>;
 168
 169 def M3WriteS1 : SchedWriteRes<[M3UnitS]>   { let Latency = 1; }
 170 def M3WriteSA : SchedWriteRes<[M3UnitA,
 171                                M3UnitS,
 172                                M3UnitFST]> { let Latency = 3;
 173                                              let NumMicroOps = 2; }
 174 def M3WriteSB : SchedWriteRes<[M3UnitA,
 175                                M3UnitS]>   { let Latency = 2;
 176                                              let NumMicroOps = 2; }
 177
 178 def M3ReadAdrBase : SchedReadVariant<[SchedVar<ExynosScaledIdxPred, [ReadDefault]>,
 179                                       SchedVar<NoSchedPred,         [ReadDefault]>]>;
 180
 181 // Branch instructions.
 182 def : SchedAlias<WriteBr, M3WriteZ0>;
 183 def : SchedAlias<WriteBrReg, M3WriteC1>;
 184
 185 // Arithmetic and logical integer instructions.
 186 def : SchedAlias<WriteI,     M3WriteA1>;
 187 def : SchedAlias<WriteISReg, M3WriteA1>;
 188 def : SchedAlias<WriteIEReg, M3WriteA1>;
 189 def : SchedAlias<WriteIS,    M3WriteA1>;
 190
 191 // Move instructions.
 192 def : SchedAlias<WriteImm, M3WriteA1>;
 193
 194 // Divide and multiply instructions.
 195 def : WriteRes<WriteID32, [M3UnitC,
 196                            M3UnitD]>  { let Latency = 12;
 197                                         let ResourceCycles = [1, 12]; }
 198 def : WriteRes<WriteID64, [M3UnitC,
 199                            M3UnitD]>  { let Latency = 21;
 200                                         let ResourceCycles = [1, 21]; }
 201 def : WriteRes<WriteIM32, [M3UnitC]>  { let Latency = 3; }
 202 def : WriteRes<WriteIM64, [M3UnitC]>  { let Latency = 4;
 203                                         let ResourceCycles = [2]; }
 204
 205 // Miscellaneous instructions.
 206 def : SchedAlias<WriteExtr, M3WriteAY>;
 207
 208 // Addressing modes.
 209 def : SchedAlias<WriteAdr,    M3WriteZ1>;
 210 def : SchedAlias<ReadAdrBase, M3ReadAdrBase>;
 211
 212 // Load instructions.
 213 def : SchedAlias<WriteLD, M3WriteL4>;
 214 def : WriteRes<WriteLDHi, []> { let Latency = 4;
 215                                 let NumMicroOps = 0; }
 216 def : SchedAlias<WriteLDIdx, M3WriteLB>;
 217
 218 // Store instructions.
 219 def : SchedAlias<WriteST,    M3WriteS1>;
 220 def : SchedAlias<WriteSTP,   M3WriteS1>;
 221 def : SchedAlias<WriteSTX,   M3WriteS1>;
 222 def : SchedAlias<WriteSTIdx, M3WriteSB>;
 223
 224 // FP data instructions.
 225 def : WriteRes<WriteF,    [M3UnitFADD]>  { let Latency = 2; }
 226 def : WriteRes<WriteFCmp, [M3UnitNMSC]>  { let Latency = 2; }
 227 def : WriteRes<WriteFDiv, [M3UnitFDIV]>  { let Latency = 12;
 228                                            let ResourceCycles = [12]; }
 229 def : WriteRes<WriteFMul, [M3UnitFMAC]>  { let Latency = 4; }
 230
 231 // FP miscellaneous instructions.
 232 def : WriteRes<WriteFCvt,  [M3UnitFCVT]> { let Latency = 3; }
 233 def : WriteRes<WriteFImm,  [M3UnitNALU]> { let Latency = 1; }
 234 def : WriteRes<WriteFCopy, [M3UnitNALU]> { let Latency = 1; }
 235
 236 // FP load instructions.
 237 def : SchedAlias<WriteVLD, M3WriteL5>;
 238
 239 // FP store instructions.
 240 def : WriteRes<WriteVST, [M3UnitS,
 241                           M3UnitFST]> { let Latency = 1;
 242                                         let NumMicroOps = 1; }
 243
 244 // ASIMD FP instructions.
 245 def : WriteRes<WriteV, [M3UnitNALU]> { let Latency = 3; }
 246
 247 // Other miscellaneous instructions.
 248 def : WriteRes<WriteAtomic,  []> { let Unsupported = 1; }
 249 def : WriteRes<WriteBarrier, []> { let Latency = 1; }
 250 def : WriteRes<WriteHint,    []> { let Latency = 1; }
 251 def : WriteRes<WriteSys,     []> { let Latency = 1; }
 252
 253 //===----------------------------------------------------------------------===//
 254 // Generic fast forwarding.
 255
 256 // TODO: Add FP register forwarding rules.
 257
 258 def : ReadAdvance<ReadI,       0>;
 259 def : ReadAdvance<ReadISReg,   0>;
 260 def : ReadAdvance<ReadIEReg,   0>;
 261 def : ReadAdvance<ReadIM,      0>;
 262 // TODO: The forwarding for 32 bits actually saves 2 cycles.
 263 def : ReadAdvance<ReadIMA,     3, [WriteIM32, WriteIM64]>;
 264 def : ReadAdvance<ReadID,      0>;
 265 def : ReadAdvance<ReadExtrHi,  0>;
 266 def : ReadAdvance<ReadAdrBase, 0>;
 267 def : ReadAdvance<ReadVLD,     0>;
 268
 269 //===----------------------------------------------------------------------===//
 270 // Finer scheduling model.
 271
 272 def M3WriteNEONA   : SchedWriteRes<[M3UnitNSHF,
 273                                     M3UnitFADD]>  { let Latency = 3;
 274                                                     let NumMicroOps = 2; }
 275 def M3WriteNEONB   : SchedWriteRes<[M3UnitNALU,
 276                                     M3UnitFST]>   { let Latency = 10;
 277                                                     let NumMicroOps = 2; }
 278 def M3WriteNEOND   : SchedWriteRes<[M3UnitNSHF,
 279                                     M3UnitFST]>   { let Latency = 6;
 280                                                     let NumMicroOps = 2; }
 281 def M3WriteNEONH   : SchedWriteRes<[M3UnitNALU,
 282                                     M3UnitS]>     { let Latency = 5;
 283                                                     let NumMicroOps = 2; }
 284 def M3WriteNEONI   : SchedWriteRes<[M3UnitNSHF,
 285                                     M3UnitS]>     { let Latency = 5;
 286                                                     let NumMicroOps = 2; }
 287 def M3WriteNEONV   : SchedWriteRes<[M3UnitFDIV0,
 288                                     M3UnitFDIV1]>  { let Latency = 7;
 289                                                      let NumMicroOps = 2;
 290                                                      let ResourceCycles = [8, 8]; }
 291 def M3WriteNEONW   : SchedWriteRes<[M3UnitFDIV0,
 292                                     M3UnitFDIV1]>  { let Latency = 12;
 293                                                      let NumMicroOps = 2;
 294                                                      let ResourceCycles = [13, 13]; }
 295 def M3WriteNEONX   : SchedWriteRes<[M3UnitFSQR,
 296                                     M3UnitFSQR]>  { let Latency = 18;
 297                                                     let NumMicroOps = 2;
 298                                                     let ResourceCycles = [19, 19]; }
 299 def M3WriteNEONY   : SchedWriteRes<[M3UnitFSQR,
 300                                     M3UnitFSQR]>  { let Latency = 25;
 301                                                     let NumMicroOps = 2;
 302                                                     let ResourceCycles = [26, 26]; }
 303 def M3WriteNEONZ   : SchedWriteRes<[M3UnitNMSC,
 304                                     M3UnitNMSC]>  { let Latency = 5;
 305                                                     let NumMicroOps = 2; }
 306 def M3WriteFADD2   : SchedWriteRes<[M3UnitFADD]>  { let Latency = 2; }
 307 def M3WriteFCVT2   : SchedWriteRes<[M3UnitFCVT]>  { let Latency = 2; }
 308 def M3WriteFCVT3   : SchedWriteRes<[M3UnitFCVT]>  { let Latency = 3; }
 309 def M3WriteFCVT3A  : SchedWriteRes<[M3UnitFCVT0]> { let Latency = 3; }
 310 def M3WriteFCVT4A  : SchedWriteRes<[M3UnitFCVT0]> { let Latency = 4; }
 311 def M3WriteFCVT4   : SchedWriteRes<[M3UnitFCVT]>  { let Latency = 4; }
 312 def M3WriteFDIV10  : SchedWriteRes<[M3UnitFDIV]>  { let Latency = 7;
 313                                                     let ResourceCycles = [8]; }
 314 def M3WriteFDIV12  : SchedWriteRes<[M3UnitFDIV]>  { let Latency = 12;
 315                                                     let ResourceCycles = [13]; }
 316 def M3WriteFMAC3   : SchedWriteRes<[M3UnitFMAC]>  { let Latency = 3; }
 317 def M3WriteFMAC4   : SchedWriteRes<[M3UnitFMAC]>  { let Latency = 4; }
 318 def M3WriteFMAC5   : SchedWriteRes<[M3UnitFMAC]>  { let Latency = 5; }
 319 def M3WriteFSQR17  : SchedWriteRes<[M3UnitFSQR]>  { let Latency = 18;
 320                                                     let ResourceCycles = [19]; }
 321 def M3WriteFSQR25  : SchedWriteRes<[M3UnitFSQR]>  { let Latency = 25;
 322                                                     let ResourceCycles = [26]; }
 323 def M3WriteNALU1   : SchedWriteRes<[M3UnitNALU]>  { let Latency = 1; }
 324 def M3WriteNCRY1A  : SchedWriteRes<[M3UnitNCRY0]> { let Latency = 1; }
 325 def M3WriteNCRY3A  : SchedWriteRes<[M3UnitNCRY0]> { let Latency = 3; }
 326 def M3WriteNCRY5A  : SchedWriteRes<[M3UnitNCRY]>  { let Latency = 5; }
 327 def M3WriteNMSC1   : SchedWriteRes<[M3UnitNMSC]>  { let Latency = 1; }
 328 def M3WriteNMSC2   : SchedWriteRes<[M3UnitNMSC]>  { let Latency = 2; }
 329 def M3WriteNMSC3   : SchedWriteRes<[M3UnitNMSC]>  { let Latency = 3; }
 330 def M3WriteNMUL3   : SchedWriteRes<[M3UnitNMUL]>  { let Latency = 3; }
 331 def M3WriteNSHF1   : SchedWriteRes<[M3UnitNSHF]>  { let Latency = 1; }
 332 def M3WriteNSHF3   : SchedWriteRes<[M3UnitNSHF]>  { let Latency = 3; }
 333 def M3WriteNSHT1   : SchedWriteRes<[M3UnitNSHT]>  { let Latency = 1; }
 334 def M3WriteNSHT2   : SchedWriteRes<[M3UnitNSHT]>  { let Latency = 2; }
 335 def M3WriteNSHT3   : SchedWriteRes<[M3UnitNSHT]>  { let Latency = 3; }
 336 def M3WriteVLDA    : SchedWriteRes<[M3UnitL,
 337                                     M3UnitL]>     { let Latency = 5;
 338                                                     let NumMicroOps = 2; }
 339 def M3WriteVLDB    : SchedWriteRes<[M3UnitL,
 340                                     M3UnitL,
 341                                     M3UnitL]>     { let Latency = 6;
 342                                                     let NumMicroOps = 3; }
 343 def M3WriteVLDC    : SchedWriteRes<[M3UnitL,
 344                                     M3UnitL,
 345                                     M3UnitL,
 346                                     M3UnitL]>     { let Latency = 6;
 347                                                     let NumMicroOps = 4; }
 348 def M3WriteVLDD    : SchedWriteRes<[M3UnitL,
 349                                     M3UnitNALU]>  { let Latency = 7;
 350                                                     let NumMicroOps = 2;
 351                                                     let ResourceCycles = [2, 1]; }
 352 def M3WriteVLDE    : SchedWriteRes<[M3UnitL,
 353                                     M3UnitNALU]>  { let Latency = 6;
 354                                                     let NumMicroOps = 2;
 355                                                     let ResourceCycles = [2, 1]; }
 356 def M3WriteVLDF    : SchedWriteRes<[M3UnitL,
 357                                     M3UnitL]>     { let Latency = 10;
 358                                                     let NumMicroOps = 2;
 359                                                     let ResourceCycles = [5, 5]; }
 360 def M3WriteVLDG    : SchedWriteRes<[M3UnitL,
 361                                     M3UnitNALU,
 362                                     M3UnitNALU]>  { let Latency = 7;
 363                                                     let NumMicroOps = 3;
 364                                                     let ResourceCycles = [2, 1, 1]; }
 365 def M3WriteVLDH    : SchedWriteRes<[M3UnitL,
 366                                     M3UnitNALU,
 367                                     M3UnitNALU]>  { let Latency = 6;
 368                                                     let NumMicroOps = 3;
 369                                                     let ResourceCycles = [2, 1, 1]; }
 370 def M3WriteVLDI    : SchedWriteRes<[M3UnitL,
 371                                     M3UnitL,
 372                                     M3UnitL]>     { let Latency = 12;
 373                                                     let NumMicroOps = 3;
 374                                                     let ResourceCycles = [6, 6, 6]; }
 375 def M3WriteVLDJ    : SchedWriteRes<[M3UnitL,
 376                                     M3UnitNALU,
 377                                     M3UnitNALU,
 378                                     M3UnitNALU]>  { let Latency = 7;
 379                                                     let NumMicroOps = 4;
 380                                                     let ResourceCycles = [2, 1, 1, 1]; }
 381 def M3WriteVLDK    : SchedWriteRes<[M3UnitL,
 382                                     M3UnitNALU,
 383                                     M3UnitNALU,
 384                                     M3UnitNALU,
 385                                     M3UnitNALU]>  { let Latency = 9;
 386                                                     let NumMicroOps = 5;
 387                                                     let ResourceCycles = [4, 1, 1, 1, 1]; }
 388 def M3WriteVLDL    : SchedWriteRes<[M3UnitL,
 389                                     M3UnitNALU,
 390                                     M3UnitNALU,
 391                                     M3UnitL,
 392                                     M3UnitNALU]>  { let Latency = 6;
 393                                                     let NumMicroOps = 5;
 394                                                     let ResourceCycles = [6, 1, 1, 6, 1]; }
 395 def M3WriteVLDM    : SchedWriteRes<[M3UnitL,
 396                                     M3UnitNALU,
 397                                     M3UnitNALU,
 398                                     M3UnitL,
 399                                     M3UnitNALU,
 400                                     M3UnitNALU]>  { let Latency = 7;
 401                                                     let NumMicroOps = 6;
 402                                                     let ResourceCycles = [6, 1, 1, 6, 1, 1]; }
 403 def M3WriteVLDN    : SchedWriteRes<[M3UnitL,
 404                                     M3UnitL,
 405                                     M3UnitL,
 406                                     M3UnitL]>     { let Latency = 14;
 407                                                     let NumMicroOps = 4;
 408                                                     let ResourceCycles = [6, 6, 6, 6]; }
 409 def M3WriteVSTA    : WriteSequence<[WriteVST], 2>;
 410 def M3WriteVSTB    : WriteSequence<[WriteVST], 3>;
 411 def M3WriteVSTC    : WriteSequence<[WriteVST], 4>;
 412 def M3WriteVSTD    : SchedWriteRes<[M3UnitS,
 413                                     M3UnitFST,
 414                                     M3UnitS,
 415                                     M3UnitFST]>   { let Latency = 7;
 416                                                     let NumMicroOps = 4;
 417                                                     let ResourceCycles = [1, 3, 1, 3]; }
 418 def M3WriteVSTE    : SchedWriteRes<[M3UnitS,
 419                                     M3UnitFST,
 420                                     M3UnitS,
 421                                     M3UnitFST,
 422                                     M3UnitS,
 423                                     M3UnitFST]>   { let Latency = 8;
 424                                                     let NumMicroOps = 6;
 425                                                     let ResourceCycles = [1, 3, 1, 3, 1, 3]; }
 426 def M3WriteVSTF    : SchedWriteRes<[M3UnitNALU,
 427                                     M3UnitFST,
 428                                     M3UnitFST,
 429                                     M3UnitS,
 430                                     M3UnitFST,
 431                                     M3UnitS,
 432                                     M3UnitFST]>   { let Latency = 15;
 433                                                     let NumMicroOps = 7;
 434                                                     let ResourceCycles = [1, 3, 3, 1, 3, 1, 3]; }
 435 def M3WriteVSTG    : SchedWriteRes<[M3UnitNALU,
 436                                     M3UnitFST,
 437                                     M3UnitFST,
 438                                     M3UnitS,
 439                                     M3UnitFST,
 440                                     M3UnitS,
 441                                     M3UnitFST,
 442                                     M3UnitS,
 443                                     M3UnitFST]>   { let Latency = 16;
 444                                                     let NumMicroOps = 9;
 445                                                     let ResourceCycles = [1, 3, 3, 1, 3, 1, 3, 1, 3]; }
 446 def M3WriteVSTH    : SchedWriteRes<[M3UnitNALU,
 447                                     M3UnitFST,
 448                                     M3UnitFST,
 449                                     M3UnitS,
 450                                     M3UnitFST]>   { let Latency = 14;
 451                                                     let NumMicroOps = 5;
 452                                                     let ResourceCycles = [1, 3, 3, 1, 3]; }
 453 def M3WriteVSTI    : SchedWriteRes<[M3UnitNALU,
 454                                     M3UnitFST,
 455                                     M3UnitFST,
 456                                     M3UnitS,
 457                                     M3UnitFST,
 458                                     M3UnitS,
 459                                     M3UnitFST,
 460                                     M3UnitS,
 461                                     M3UnitFST]>   { let Latency = 17;
 462                                                     let NumMicroOps = 9;
 463                                                     let ResourceCycles = [1, 3, 3, 1, 3, 1, 3, 1, 3]; }
 464
 465 // Special cases.
 466 def M3WriteAES     : SchedWriteRes<[M3UnitNCRY]>  { let Latency = 1; }
 467 def M3WriteCOPY    : SchedWriteVariant<[SchedVar<ExynosFPPred, [M3WriteNALU1]>,
 468                                         SchedVar<NoSchedPred,  [M3WriteZ0]>]>;
 469 def M3WriteMOVI    : SchedWriteVariant<[SchedVar<IsZeroFPIdiomPred, [M3WriteZ0]>,
 470                                         SchedVar<NoSchedPred,       [M3WriteNALU1]>]>;
 471
 472 // Fast forwarding.
 473 def M3ReadAES      : SchedReadAdvance<1, [M3WriteAES]>;
 474 def M3ReadFMAC     : SchedReadAdvance<1, [M3WriteFMAC4,
 475                                           M3WriteFMAC5]>;
 476 def M3ReadNMUL     : SchedReadAdvance<1, [M3WriteNMUL3]>;
 477
 478 // Branch instructions
 479 def : InstRW<[M3WriteB1], (instrs Bcc)>;
 480 def : InstRW<[M3WriteA1], (instrs BL)>;
 481 def : InstRW<[M3WriteBX], (instrs BLR)>;
 482 def : InstRW<[M3WriteC1], (instregex "^CBN?Z[WX]")>;
 483 def : InstRW<[M3WriteAD], (instregex "^TBN?Z[WX]")>;
 484
 485 // Arithmetic and logical integer instructions.
 486 def : InstRW<[M3WriteAX], (instregex "^(ADD|AND|BIC|EON|EOR|ORN|SUB)[WX]rs$")>;
 487 def : InstRW<[M3WriteAU], (instrs ORRWrs, ORRXrs)>;
 488 def : InstRW<[M3WriteAX], (instregex "^(ADD|SUB)S?[WX]rx(64)?$")>;
 489 def : InstRW<[M3WriteAX], (instregex "^(ADD|AND|BIC|SUB)S[WX]rs$")>;
 490 def : InstRW<[M3WriteAV], (instrs ADDWri, ADDXri)>;
 491 def : InstRW<[M3WriteAW], (instrs ORRWri, ORRXri)>;
 492
 493 // Move instructions.
 494 def : InstRW<[M3WriteCOPY], (instrs COPY)>;
 495 def : InstRW<[M3WriteZ0],   (instrs ADR, ADRP)>;
 496 def : InstRW<[M3WriteZ0],   (instregex "^MOV[NZ][WX]i")>;
 497
 498 // Divide and multiply instructions.
 499
 500 // Miscellaneous instructions.
 501
 502 // Load instructions.
 503 def : InstRW<[M3WriteLD,
 504               WriteLDHi,
 505               WriteAdr],    (instregex "^LDP(SW|W|X)(post|pre)")>;
 506 def : InstRW<[M3WriteLB,
 507               ReadAdrBase], (instregex "^LDR(BB|SBW|SBX|HH|SHW|SHX|SW|W|X)roW")>;
 508 def : InstRW<[M3WriteLX,
 509               ReadAdrBase], (instregex "^LDR(BB|SBW|SBX|HH|SHW|SHX|SW|W|X)roX")>;
 510 def : InstRW<[M3WriteLB,
 511               ReadAdrBase], (instrs PRFMroW)>;
 512 def : InstRW<[M3WriteLX,
 513               ReadAdrBase], (instrs PRFMroX)>;
 514
 515 // Store instructions.
 516 def : InstRW<[M3WriteSB,
 517               ReadAdrBase], (instregex "^STR(BB|HH|W|X)roW")>;
 518 def : InstRW<[WriteST,
 519               ReadAdrBase], (instregex "^STR(BB|HH|W|X)roX")>;
 520
 521 // FP data instructions.
 522 def : InstRW<[M3WriteNSHF1],  (instregex "^FABS[DS]r")>;
 523 def : InstRW<[M3WriteFADD2],  (instregex "^F(ADD|SUB)[DS]rr")>;
 524 def : InstRW<[M3WriteFDIV10], (instrs FDIVSrr)>;
 525 def : InstRW<[M3WriteFDIV12], (instrs FDIVDrr)>;
 526 def : InstRW<[M3WriteNMSC1],  (instregex "^F(MAX|MIN).+rr")>;
 527 def : InstRW<[M3WriteFMAC3],  (instregex "^FN?MUL[DS]rr")>;
 528 def : InstRW<[M3WriteFMAC4,
 529               M3ReadFMAC],    (instregex "^FN?M(ADD|SUB)[DS]rrr")>;
 530 def : InstRW<[M3WriteNALU1],  (instregex "^FNEG[DS]r")>;
 531 def : InstRW<[M3WriteFCVT3A], (instregex "^FRINT.+r")>;
 532 def : InstRW<[M3WriteNEONH],  (instregex "^FCSEL[DS]rrr")>;
 533 def : InstRW<[M3WriteFSQR17], (instrs FSQRTSr)>;
 534 def : InstRW<[M3WriteFSQR25], (instrs FSQRTDr)>;
 535
 536 // FP miscellaneous instructions.
 537 def : InstRW<[M3WriteFCVT3],  (instregex "^FCVT[DHS][DHS]r")>;
 538 def : InstRW<[M3WriteFCVT4A], (instregex "^[SU]CVTF[SU][XW][DHS]ri")>;
 539 def : InstRW<[M3WriteFCVT3A], (instregex "^FCVT[AMNPZ][SU]U[XW][DHS]r")>;
 540 def : InstRW<[M3WriteFCVT3A], (instregex "^FCVTZ[SU][dhs]")>;
 541 def : InstRW<[M3WriteNALU1],  (instregex "^FMOV[DS][ir]")>;
 542 def : InstRW<[M3WriteFCVT4],  (instregex "^[FU](RECP|RSQRT)Ev1")>;
 543 def : InstRW<[M3WriteNMSC1],  (instregex "^FRECPXv1")>;
 544 def : InstRW<[M3WriteFMAC4,
 545               M3ReadFMAC],    (instregex "^F(RECP|RSQRT)S(16|32|64)")>;
 546 def : InstRW<[M3WriteNALU1],  (instregex "^FMOV[WX][DS]r")>;
 547 def : InstRW<[M3WriteNALU1],  (instregex "^FMOV[DS][WX]r")>;
 548 def : InstRW<[M3WriteNEONI],  (instregex "^FMOV(DX|XD)Highr")>;
 549
 550 // FP load instructions.
 551 def : InstRW<[WriteVLD],    (instregex "^LDR[DSQ]l")>;
 552 def : InstRW<[WriteVLD],    (instregex "^LDUR[BDHSQ]i")>;
 553 def : InstRW<[WriteVLD,
 554               WriteAdr],    (instregex "^LDR[BDHSQ](post|pre)")>;
 555 def : InstRW<[WriteVLD],    (instregex "^LDR[BDHSQ]ui")>;
 556 def : InstRW<[M3WriteLE,
 557               ReadAdrBase], (instregex "^LDR[BDHS]roW")>;
 558 def : InstRW<[WriteVLD,
 559               ReadAdrBase], (instregex "^LDR[BDHS]roX")>;
 560 def : InstRW<[M3WriteLE,
 561               ReadAdrBase], (instregex "^LDRQro[WX]")>;
 562 def : InstRW<[WriteVLD,
 563               M3WriteLH],   (instregex "^LDN?P[DS]i")>;
 564 def : InstRW<[M3WriteLA,
 565               M3WriteLH],   (instregex "^LDN?PQi")>;
 566 def : InstRW<[M3WriteLB,
 567               M3WriteLH,
 568               WriteAdr],    (instregex "^LDP[DS](post|pre)")>;
 569 def : InstRW<[M3WriteLC,
 570               M3WriteLH,
 571               WriteAdr],    (instregex "^LDPQ(post|pre)")>;
 572
 573 // FP store instructions.
 574 def : InstRW<[WriteVST],    (instregex "^STUR[BDHSQ]i")>;
 575 def : InstRW<[WriteVST,
 576               WriteAdr],    (instregex "^STR[BDHSQ](post|pre)")>;
 577 def : InstRW<[WriteVST],    (instregex "^STR[BDHSQ]ui")>;
 578 def : InstRW<[M3WriteSA,
 579               ReadAdrBase], (instregex "^STR[BDHS]roW")>;
 580 def : InstRW<[WriteVST,
 581               ReadAdrBase], (instregex "^STR[BDHS]roX")>;
 582 def : InstRW<[M3WriteSA,
 583               ReadAdrBase], (instregex "^STRQro[WX]")>;
 584 def : InstRW<[WriteVST],    (instregex "^STN?P[DSQ]i")>;
 585 def : InstRW<[WriteVST,
 586               WriteAdr],    (instregex "^STP[DS](post|pre)")>;
 587 def : InstRW<[M3WriteSA,
 588               WriteAdr],    (instregex "^STPQ(post|pre)")>;
 589
 590 // ASIMD instructions.
 591 def : InstRW<[M3WriteNMSC3], (instregex "^[SU]ABAL?v")>;
 592 def : InstRW<[M3WriteNMSC1], (instregex "^[SU]ABDL?v")>;
 593 def : InstRW<[M3WriteNMSC1], (instregex "^((SQ)?ABS|SQNEG)v")>;
 594 def : InstRW<[M3WriteNALU1], (instregex "^(ADD|NEG|SUB)v")>;
 595 def : InstRW<[M3WriteNMSC3], (instregex "^[SU]?ADDL?Pv")>;
 596 def : InstRW<[M3WriteNMSC3], (instregex "^[SU]H(ADD|SUB)v")>;
 597 def : InstRW<[M3WriteNMSC3], (instregex "^[SU](ADD|SUB)[LW]V?v")>;
 598 def : InstRW<[M3WriteNMSC3], (instregex "^R?(ADD|SUB)HN2?v")>;
 599 def : InstRW<[M3WriteNMSC3], (instregex "^[SU]Q(ADD|SUB)v")>;
 600 def : InstRW<[M3WriteNMSC3], (instregex "^(SU|US)QADDv")>;
 601 def : InstRW<[M3WriteNMSC3], (instregex "^[SU]RHADDv")>;
 602 def : InstRW<[M3WriteNMSC1], (instregex "^CM(EQ|GE|GT|HI|HS|LE|LT)v")>;
 603 def : InstRW<[M3WriteNALU1], (instregex "^CMTSTv")>;
 604 def : InstRW<[M3WriteNALU1], (instregex "^(AND|BIC|EOR|MVNI|NOT|ORN|ORR)v")>;
 605 def : InstRW<[M3WriteNMSC1], (instregex "^[SU](MIN|MAX)v")>;
 606 def : InstRW<[M3WriteNMSC2], (instregex "^[SU](MIN|MAX)Pv")>;
 607 def : InstRW<[M3WriteNMSC3], (instregex "^[SU](MIN|MAX)Vv")>;
 608 def : InstRW<[M3WriteNMUL3], (instregex "^(MUL|SQR?DMULH)v")>;
 609 def : InstRW<[M3WriteNMUL3,
 610               M3ReadNMUL],   (instregex "^ML[AS]v")>;
 611 def : InstRW<[M3WriteNMUL3], (instregex "^[SU]ML[AS]Lv")>;
 612 def : InstRW<[M3WriteNMUL3], (instregex "^SQDML[AS]L")>;
 613 def : InstRW<[M3WriteNMUL3], (instregex "^(S|U|SQD)MULLv")>;
 614 def : InstRW<[M3WriteNMSC3], (instregex "^[SU]ADALPv")>;
 615 def : InstRW<[M3WriteNSHT3], (instregex "^[SU]R?SRAv")>;
 616 def : InstRW<[M3WriteNSHT1], (instregex "^SHL[dv]")>;
 617 def : InstRW<[M3WriteNSHT1], (instregex "^[SU]SH[LR][dv]")>;
 618 def : InstRW<[M3WriteNSHT1], (instregex "^S[RS]I[dv]")>;
 619 def : InstRW<[M3WriteNSHT2], (instregex "^[SU]?SHLLv")>;
 620 def : InstRW<[M3WriteNSHT3], (instregex "^(([SU]Q)?R)?SHRU?N[bhsv]")>;
 621 def : InstRW<[M3WriteNSHT3], (instregex "^[SU]RSH[LR][dv]")>;
 622 def : InstRW<[M3WriteNSHT3], (instregex "^[SU]QR?SHLU?[bdhsv]")>;
 623
 624 // ASIMD FP instructions.
 625 def : InstRW<[M3WriteNSHF1],  (instregex "^FABSv")>;
 626 def : InstRW<[M3WriteFADD2],  (instregex "^F(ABD|ADD|SUB)v")>;
 627 def : InstRW<[M3WriteNEONA],  (instregex "^FADDP")>;
 628 def : InstRW<[M3WriteNMSC1],  (instregex "^F(AC|CM)(EQ|GE|GT|LE|LT)v[^1]")>;
 629 def : InstRW<[M3WriteFCVT3],  (instregex "^FCVT(L|N|XN)v")>;
 630 def : InstRW<[M3WriteFCVT2],  (instregex "^FCVT[AMNPZ][SU]v")>;
 631 def : InstRW<[M3WriteFCVT2],  (instregex "^[SU]CVTFv")>;
 632 def : InstRW<[M3WriteFDIV10], (instrs FDIVv2f32)>;
 633 def : InstRW<[M3WriteNEONV],  (instrs FDIVv4f32)>;
 634 def : InstRW<[M3WriteNEONW],  (instrs FDIVv2f64)>;
 635 def : InstRW<[M3WriteNMSC1],  (instregex "^F(MAX|MIN)(NM)?v")>;
 636 def : InstRW<[M3WriteNMSC2],  (instregex "^F(MAX|MIN)(NM)?Pv")>;
 637 def : InstRW<[M3WriteNEONZ],  (instregex "^F(MAX|MIN)(NM)?Vv")>;
 638 def : InstRW<[M3WriteFMAC3],  (instregex "^FMULX?v.[fi]")>;
 639 def : InstRW<[M3WriteFMAC4,
 640               M3ReadFMAC],    (instregex "^FML[AS]v.f")>;
 641 def : InstRW<[M3WriteFMAC5,
 642               M3ReadFMAC],    (instregex "^FML[AS]v.i")>;
 643 def : InstRW<[M3WriteNALU1],  (instregex "^FNEGv")>;
 644 def : InstRW<[M3WriteFCVT3A], (instregex "^FRINT[AIMNPXZ]v")>;
 645 def : InstRW<[M3WriteFSQR17], (instrs FSQRTv2f32)>;
 646 def : InstRW<[M3WriteNEONX],  (instrs FSQRTv4f32)>;
 647 def : InstRW<[M3WriteNEONY],  (instrs FSQRTv2f64)>;
 648
 649 // ASIMD miscellaneous instructions.
 650 def : InstRW<[M3WriteNALU1], (instregex "^RBITv")>;
 651 def : InstRW<[M3WriteNALU1], (instregex "^(BIF|BIT|BSL)v")>;
 652 def : InstRW<[M3WriteNEONB], (instregex "^DUPv.+gpr")>;
 653 def : InstRW<[M3WriteNSHF1], (instregex "^DUPv.+lane")>;
 654 def : InstRW<[M3WriteNSHF1], (instregex "^EXTv")>;
 655 def : InstRW<[M3WriteNSHF1], (instregex "^[SU]?Q?XTU?Nv")>;
 656 def : InstRW<[M3WriteNSHF1], (instregex "^CPY")>;
 657 def : InstRW<[M3WriteNSHF1], (instregex "^INSv.+lane")>;
 658 def : InstRW<[M3WriteMOVI],  (instregex "^MOVI")>;
 659 def : InstRW<[M3WriteNALU1], (instregex "^FMOVv")>;
 660 def : InstRW<[M3WriteFCVT4], (instregex "^[FU](RECP|RSQRT)Ev[248]")>;
 661 def : InstRW<[M3WriteFMAC4,
 662               M3ReadFMAC],   (instregex "^F(RECP|RSQRT)Sv")>;
 663 def : InstRW<[M3WriteNSHF1], (instregex "^REV(16|32|64)v")>;
 664 def : InstRW<[M3WriteNSHF1], (instregex "^TB[LX]v")>;
 665 def : InstRW<[M3WriteNEOND], (instregex "^[SU]MOVv")>;
 666 def : InstRW<[M3WriteNSHF3], (instregex "^INSv.+gpr")>;
 667 def : InstRW<[M3WriteNSHF1], (instregex "^(TRN|UZP|ZIP)[12]v")>;
 668
 669 // ASIMD load instructions.
 670 def : InstRW<[M3WriteL5],   (instregex "LD1Onev(8b|4h|2s|1d)$")>;
 671 def : InstRW<[M3WriteL5,
 672               M3WriteA1],   (instregex "LD1Onev(8b|4h|2s|1d)_POST")>;
 673 def : InstRW<[M3WriteL5],   (instregex "LD1Onev(16b|8h|4s|2d)$")>;
 674 def : InstRW<[M3WriteL5,
 675               M3WriteA1],   (instregex "LD1Onev(16b|8h|4s|2d)_POST")>;
 676
 677 def : InstRW<[M3WriteVLDA], (instregex "LD1Twov(8b|4h|2s|1d)$")>;
 678 def : InstRW<[M3WriteVLDA,
 679               M3WriteA1],   (instregex "LD1Twov(8b|4h|2s|1d)_POST")>;
 680 def : InstRW<[M3WriteVLDA], (instregex "LD1Twov(16b|8h|4s|2d)$")>;
 681 def : InstRW<[M3WriteVLDA,
 682               M3WriteA1],   (instregex "LD1Twov(16b|8h|4s|2d)_POST")>;
 683
 684 def : InstRW<[M3WriteVLDB], (instregex "LD1Threev(8b|4h|2s|1d)$")>;
 685 def : InstRW<[M3WriteVLDB,
 686               M3WriteA1],   (instregex "LD1Threev(8b|4h|2s|1d)_POST")>;
 687 def : InstRW<[M3WriteVLDB], (instregex "LD1Threev(16b|8h|4s|2d)$")>;
 688 def : InstRW<[M3WriteVLDB,
 689               M3WriteA1],   (instregex "LD1Threev(16b|8h|4s|2d)_POST")>;
 690
 691 def : InstRW<[M3WriteVLDC], (instregex "LD1Fourv(8b|4h|2s|1d)$")>;
 692 def : InstRW<[M3WriteVLDC,
 693               M3WriteA1],   (instregex "LD1Fourv(8b|4h|2s|1d)_POST")>;
 694 def : InstRW<[M3WriteVLDC], (instregex "LD1Fourv(16b|8h|4s|2d)$")>;
 695 def : InstRW<[M3WriteVLDC,
 696               M3WriteA1],   (instregex "LD1Fourv(16b|8h|4s|2d)_POST")>;
 697
 698 def : InstRW<[M3WriteVLDD], (instregex "LD1i(8|16|32)$")>;
 699 def : InstRW<[M3WriteVLDD,
 700               M3WriteA1],   (instregex "LD1i(8|16|32)_POST")>;
 701 def : InstRW<[M3WriteVLDE], (instregex "LD1i(64)$")>;
 702 def : InstRW<[M3WriteVLDE,
 703               M3WriteA1],   (instregex "LD1i(64)_POST")>;
 704
 705 def : InstRW<[M3WriteL5],   (instregex "LD1Rv(8b|4h|2s|1d)$")>;
 706 def : InstRW<[M3WriteL5,
 707               M3WriteA1],   (instregex "LD1Rv(8b|4h|2s|1d)_POST")>;
 708 def : InstRW<[M3WriteL5],   (instregex "LD1Rv(16b|8h|4s|2d)$")>;
 709 def : InstRW<[M3WriteL5,
 710               M3WriteA1],   (instregex "LD1Rv(16b|8h|4s|2d)_POST")>;
 711
 712 def : InstRW<[M3WriteVLDF], (instregex "LD2Twov(8b|4h|2s)$")>;
 713 def : InstRW<[M3WriteVLDF,
 714               M3WriteA1],   (instregex "LD2Twov(8b|4h|2s)_POST")>;
 715 def : InstRW<[M3WriteVLDF], (instregex "LD2Twov(16b|8h|4s|2d)$")>;
 716 def : InstRW<[M3WriteVLDF,
 717               M3WriteA1],   (instregex "LD2Twov(16b|8h|4s|2d)_POST")>;
 718
 719 def : InstRW<[M3WriteVLDG], (instregex "LD2i(8|16|32)$")>;
 720 def : InstRW<[M3WriteVLDG,
 721               M3WriteA1],   (instregex "LD2i(8|16|32)_POST")>;
 722 def : InstRW<[M3WriteVLDH], (instregex "LD2i(64)$")>;
 723 def : InstRW<[M3WriteVLDH,
 724               M3WriteA1],   (instregex "LD2i(64)_POST")>;
 725
 726 def : InstRW<[M3WriteVLDA], (instregex "LD2Rv(8b|4h|2s|1d)$")>;
 727 def : InstRW<[M3WriteVLDA,
 728               M3WriteA1],   (instregex "LD2Rv(8b|4h|2s|1d)_POST")>;
 729 def : InstRW<[M3WriteVLDA], (instregex "LD2Rv(16b|8h|4s|2d)$")>;
 730 def : InstRW<[M3WriteVLDA,
 731               M3WriteA1],   (instregex "LD2Rv(16b|8h|4s|2d)_POST")>;
 732
 733 def : InstRW<[M3WriteVLDI], (instregex "LD3Threev(8b|4h|2s)$")>;
 734 def : InstRW<[M3WriteVLDI,
 735               M3WriteA1],   (instregex "LD3Threev(8b|4h|2s)_POST")>;
 736 def : InstRW<[M3WriteVLDI], (instregex "LD3Threev(16b|8h|4s|2d)$")>;
 737 def : InstRW<[M3WriteVLDI,
 738               M3WriteA1],   (instregex "LD3Threev(16b|8h|4s|2d)_POST")>;
 739
 740 def : InstRW<[M3WriteVLDJ], (instregex "LD3i(8|16|32)$")>;
 741 def : InstRW<[M3WriteVLDJ,
 742               M3WriteA1],   (instregex "LD3i(8|16|32)_POST")>;
 743 def : InstRW<[M3WriteVLDL], (instregex "LD3i(64)$")>;
 744 def : InstRW<[M3WriteVLDL,
 745               M3WriteA1],   (instregex "LD3i(64)_POST")>;
 746
 747 def : InstRW<[M3WriteVLDB], (instregex "LD3Rv(8b|4h|2s|1d)$")>;
 748 def : InstRW<[M3WriteVLDB,
 749               M3WriteA1],   (instregex "LD3Rv(8b|4h|2s|1d)_POST")>;
 750 def : InstRW<[M3WriteVLDB], (instregex "LD3Rv(16b|8h|4s|2d)$")>;
 751 def : InstRW<[M3WriteVLDB,
 752               M3WriteA1],   (instregex "LD3Rv(16b|8h|4s|2d)_POST")>;
 753
 754 def : InstRW<[M3WriteVLDN], (instregex "LD4Fourv(8b|4h|2s)$")>;
 755 def : InstRW<[M3WriteVLDN,
 756               M3WriteA1],   (instregex "LD4Fourv(8b|4h|2s)_POST")>;
 757 def : InstRW<[M3WriteVLDN], (instregex "LD4Fourv(16b|8h|4s|2d)$")>;
 758 def : InstRW<[M3WriteVLDN,
 759               M3WriteA1],   (instregex "LD4Fourv(16b|8h|4s|2d)_POST")>;
 760
 761 def : InstRW<[M3WriteVLDK], (instregex "LD4i(8|16|32)$")>;
 762 def : InstRW<[M3WriteVLDK,
 763               M3WriteA1],   (instregex "LD4i(8|16|32)_POST")>;
 764 def : InstRW<[M3WriteVLDM], (instregex "LD4i(64)$")>;
 765 def : InstRW<[M3WriteVLDM,
 766               M3WriteA1],   (instregex "LD4i(64)_POST")>;
 767
 768 def : InstRW<[M3WriteVLDC], (instregex "LD4Rv(8b|4h|2s|1d)$")>;
 769 def : InstRW<[M3WriteVLDC,
 770               M3WriteA1],   (instregex "LD4Rv(8b|4h|2s|1d)_POST")>;
 771 def : InstRW<[M3WriteVLDC], (instregex "LD4Rv(16b|8h|4s|2d)$")>;
 772 def : InstRW<[M3WriteVLDC,
 773               M3WriteA1],   (instregex "LD4Rv(16b|8h|4s|2d)_POST")>;
 774
 775 // ASIMD store instructions.
 776 def : InstRW<[WriteVST],    (instregex "ST1Onev(8b|4h|2s|1d)$")>;
 777 def : InstRW<[WriteVST,
 778               WriteAdr],    (instregex "ST1Onev(8b|4h|2s|1d)_POST")>;
 779 def : InstRW<[WriteVST],    (instregex "ST1Onev(16b|8h|4s|2d)$")>;
 780 def : InstRW<[WriteVST,
 781               WriteAdr],    (instregex "ST1Onev(16b|8h|4s|2d)_POST")>;
 782
 783 def : InstRW<[M3WriteVSTA], (instregex "ST1Twov(8b|4h|2s|1d)$")>;
 784 def : InstRW<[M3WriteVSTA,
 785               WriteAdr],    (instregex "ST1Twov(8b|4h|2s|1d)_POST")>;
 786 def : InstRW<[M3WriteVSTA], (instregex "ST1Twov(16b|8h|4s|2d)$")>;
 787 def : InstRW<[M3WriteVSTA,
 788               WriteAdr],    (instregex "ST1Twov(16b|8h|4s|2d)_POST")>;
 789
 790 def : InstRW<[M3WriteVSTB], (instregex "ST1Threev(8b|4h|2s|1d)$")>;
 791 def : InstRW<[M3WriteVSTB,
 792               WriteAdr],    (instregex "ST1Threev(8b|4h|2s|1d)_POST")>;
 793 def : InstRW<[M3WriteVSTB], (instregex "ST1Threev(16b|8h|4s|2d)$")>;
 794 def : InstRW<[M3WriteVSTB,
 795               WriteAdr],    (instregex "ST1Threev(16b|8h|4s|2d)_POST")>;
 796
 797 def : InstRW<[M3WriteVSTC], (instregex "ST1Fourv(8b|4h|2s|1d)$")>;
 798 def : InstRW<[M3WriteVSTC,
 799               WriteAdr],    (instregex "ST1Fourv(8b|4h|2s|1d)_POST")>;
 800 def : InstRW<[M3WriteVSTC], (instregex "ST1Fourv(16b|8h|4s|2d)$")>;
 801 def : InstRW<[M3WriteVSTC,
 802               WriteAdr],    (instregex "ST1Fourv(16b|8h|4s|2d)_POST")>;
 803
 804 def : InstRW<[M3WriteVSTD], (instregex "ST1i(8|16|32|64)$")>;
 805 def : InstRW<[M3WriteVSTD,
 806               WriteAdr],    (instregex "ST1i(8|16|32|64)_POST")>;
 807
 808 def : InstRW<[M3WriteVSTD], (instregex "ST2Twov(8b|4h|2s)$")>;
 809 def : InstRW<[M3WriteVSTD,
 810               WriteAdr],    (instregex "ST2Twov(8b|4h|2s)_POST")>;
 811 def : InstRW<[M3WriteVSTE], (instregex "ST2Twov(16b|8h|4s|2d)$")>;
 812 def : InstRW<[M3WriteVSTE,
 813               WriteAdr],    (instregex "ST2Twov(16b|8h|4s|2d)_POST")>;
 814
 815 def : InstRW<[M3WriteVSTD], (instregex "ST2i(8|16|32)$")>;
 816 def : InstRW<[M3WriteVSTD,
 817               WriteAdr],    (instregex "ST2i(8|16|32)_POST")>;
 818 def : InstRW<[M3WriteVSTD], (instregex "ST2i(64)$")>;
 819 def : InstRW<[M3WriteVSTD,
 820               WriteAdr],    (instregex "ST2i(64)_POST")>;
 821
 822 def : InstRW<[M3WriteVSTF], (instregex "ST3Threev(8b|4h|2s)$")>;
 823 def : InstRW<[M3WriteVSTF,
 824               WriteAdr],    (instregex "ST3Threev(8b|4h|2s)_POST")>;
 825 def : InstRW<[M3WriteVSTG], (instregex "ST3Threev(16b|8h|4s|2d)$")>;
 826 def : InstRW<[M3WriteVSTG,
 827               WriteAdr],    (instregex "ST3Threev(16b|8h|4s|2d)_POST")>;
 828
 829 def : InstRW<[M3WriteVSTH], (instregex "ST3i(8|16|32)$")>;
 830 def : InstRW<[M3WriteVSTH,
 831               WriteAdr],    (instregex "ST3i(8|16|32)_POST")>;
 832 def : InstRW<[M3WriteVSTF], (instregex "ST3i(64)$")>;
 833 def : InstRW<[M3WriteVSTF,
 834               WriteAdr],    (instregex "ST3i(64)_POST")>;
 835
 836 def : InstRW<[M3WriteVSTF], (instregex "ST4Fourv(8b|4h|2s)$")>;
 837 def : InstRW<[M3WriteVSTF,
 838               WriteAdr],    (instregex "ST4Fourv(8b|4h|2s)_POST")>;
 839 def : InstRW<[M3WriteVSTI], (instregex "ST4Fourv(16b|8h|4s|2d)$")>;
 840 def : InstRW<[M3WriteVSTI,
 841               WriteAdr],    (instregex "ST4Fourv(16b|8h|4s|2d)_POST")>;
 842
 843 def : InstRW<[M3WriteVSTF], (instregex "ST4i(8|16|32|64)$")>;
 844 def : InstRW<[M3WriteVSTF,
 845               WriteAdr],    (instregex "ST4i(8|16|32|64)_POST")>;
 846
 847 // Cryptography instructions.
 848 def : InstRW<[M3WriteAES],    (instregex "^AES[DE]")>;
 849 def : InstRW<[M3WriteAES,
 850               M3ReadAES],     (instregex "^AESI?MC")>;
 851
 852 def : InstRW<[M3WriteNCRY3A], (instregex "^PMULL?v")>;
 853
 854 def : InstRW<[M3WriteNCRY1A], (instregex "^SHA1([CHMP]|SU[01])")>;
 855 def : InstRW<[M3WriteNCRY1A], (instregex "^SHA256SU0")>;
 856 def : InstRW<[M3WriteNCRY5A], (instregex "^SHA256(H2?|SU1)")>;
 857
 858 // CRC instructions.
 859 def : InstRW<[M3WriteC2], (instregex "^CRC32")>;
 860
 861 } // SchedModel = ExynosM3Model