lib/Target/ARM/ARMScheduleA9.td

   1 //=- ARMScheduleA9.td - ARM Cortex-A9 Scheduling Definitions -*- tablegen -*-=//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This file defines the itinerary class data for the ARM Cortex A9 processors.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 //
  15 // Ad-hoc scheduling information derived from pretty vague "Cortex-A9 Technical
  16 // Reference Manual".
  17 //
  18 // Functional units
  19 def A9_Issue0  : FuncUnit; // Issue 0
  20 def A9_Issue1  : FuncUnit; // Issue 1
  21 def A9_Branch  : FuncUnit; // Branch
  22 def A9_ALU0    : FuncUnit; // ALU / MUL pipeline 0
  23 def A9_ALU1    : FuncUnit; // ALU pipeline 1
  24 def A9_AGU     : FuncUnit; // Address generation unit for ld / st
  25 def A9_NPipe   : FuncUnit; // NEON pipeline
  26 def A9_MUX0    : FuncUnit; // AGU + NEON/FPU multiplexer
  27 def A9_LSUnit  : FuncUnit; // L/S Unit
  28 def A9_DRegsVFP: FuncUnit; // FP register set, VFP side
  29 def A9_DRegsN  : FuncUnit; // FP register set, NEON side
  30
  31 // Bypasses
  32 def A9_LdBypass : Bypass;
  33
  34 def CortexA9Itineraries : ProcessorItineraries<
  35   [A9_Issue0, A9_Issue1, A9_Branch, A9_ALU0, A9_ALU1, A9_AGU, A9_NPipe, A9_MUX0,
  36    A9_LSUnit, A9_DRegsVFP, A9_DRegsN],
  37   [A9_LdBypass], [
  38   // Two fully-pipelined integer ALU pipelines
  39
  40   //
  41   // Move instructions, unconditional
  42   InstrItinData<IIC_iMOVi   , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  43                                InstrStage<1, [A9_ALU0, A9_ALU1]>], [1]>,
  44   InstrItinData<IIC_iMOVr   , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  45                                InstrStage<1, [A9_ALU0, A9_ALU1]>], [1, 1]>,
  46   InstrItinData<IIC_iMOVsi  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  47                                InstrStage<1, [A9_ALU0, A9_ALU1]>], [1, 1]>,
  48   InstrItinData<IIC_iMOVsr  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  49                                InstrStage<2, [A9_ALU0, A9_ALU1]>], [2, 1, 1]>,
  50   InstrItinData<IIC_iMOVix2 , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  51                                InstrStage<1, [A9_ALU0, A9_ALU1]>,
  52                                InstrStage<1, [A9_ALU0, A9_ALU1]>], [2]>,
  53   InstrItinData<IIC_iMOVix2addpc,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  54                                   InstrStage<1, [A9_ALU0, A9_ALU1]>,
  55                                   InstrStage<1, [A9_ALU0, A9_ALU1]>,
  56                                   InstrStage<1, [A9_ALU0, A9_ALU1]>], [3]>,
  57   InstrItinData<IIC_iMOVix2ld,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  58                                InstrStage<1, [A9_ALU0, A9_ALU1]>,
  59                                InstrStage<1, [A9_ALU0, A9_ALU1]>,
  60                                InstrStage<1, [A9_MUX0], 0>,
  61                                InstrStage<1, [A9_AGU], 0>,
  62                                InstrStage<1, [A9_LSUnit]>], [5]>,
  63   //
  64   // MVN instructions
  65   InstrItinData<IIC_iMVNi   , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  66                                InstrStage<1, [A9_ALU0, A9_ALU1]>],
  67                               [1]>,
  68   InstrItinData<IIC_iMVNr   , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  69                                InstrStage<1, [A9_ALU0, A9_ALU1]>],
  70                               [1, 1], [NoBypass, A9_LdBypass]>,
  71   InstrItinData<IIC_iMVNsi  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  72                                InstrStage<2, [A9_ALU0, A9_ALU1]>],
  73                               [2, 1]>,
  74   InstrItinData<IIC_iMVNsr  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  75                                InstrStage<3, [A9_ALU0, A9_ALU1]>],
  76                               [3, 1, 1]>,
  77   //
  78   // No operand cycles
  79   InstrItinData<IIC_iALUx   , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  80                                InstrStage<1, [A9_ALU0, A9_ALU1]>]>,
  81   //
  82   // Binary Instructions that produce a result
  83   InstrItinData<IIC_iALUi , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  84                              InstrStage<1, [A9_ALU0, A9_ALU1]>],
  85                             [1, 1], [NoBypass, A9_LdBypass]>,
  86   InstrItinData<IIC_iALUr , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  87                              InstrStage<1, [A9_ALU0, A9_ALU1]>],
  88                             [1, 1, 1], [NoBypass, A9_LdBypass, A9_LdBypass]>,
  89   InstrItinData<IIC_iALUsi, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  90                              InstrStage<2, [A9_ALU0, A9_ALU1]>],
  91                             [2, 1, 1], [NoBypass, A9_LdBypass, NoBypass]>,
  92   InstrItinData<IIC_iALUsir,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  93                              InstrStage<2, [A9_ALU0, A9_ALU1]>],
  94                             [2, 1, 1], [NoBypass, NoBypass, A9_LdBypass]>,
  95   InstrItinData<IIC_iALUsr, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
  96                              InstrStage<3, [A9_ALU0, A9_ALU1]>],
  97                             [3, 1, 1, 1],
  98                             [NoBypass, A9_LdBypass, NoBypass, NoBypass]>,
  99   //
 100   // Bitwise Instructions that produce a result
 101   InstrItinData<IIC_iBITi , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 102                              InstrStage<1, [A9_ALU0, A9_ALU1]>], [1, 1]>,
 103   InstrItinData<IIC_iBITr , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 104                              InstrStage<1, [A9_ALU0, A9_ALU1]>], [1, 1, 1]>,
 105   InstrItinData<IIC_iBITsi, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 106                              InstrStage<2, [A9_ALU0, A9_ALU1]>], [2, 1, 1]>,
 107   InstrItinData<IIC_iBITsr, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 108                              InstrStage<3, [A9_ALU0, A9_ALU1]>], [3, 1, 1, 1]>,
 109   //
 110   // Unary Instructions that produce a result
 111
 112   // CLZ, RBIT, etc.
 113   InstrItinData<IIC_iUNAr , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 114                              InstrStage<1, [A9_ALU0, A9_ALU1]>], [1, 1]>,
 115
 116   // BFC, BFI, UBFX, SBFX
 117   InstrItinData<IIC_iUNAsi, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 118                              InstrStage<2, [A9_ALU0, A9_ALU1]>], [2, 1]>,
 119
 120   //
 121   // Zero and sign extension instructions
 122   InstrItinData<IIC_iEXTr , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 123                              InstrStage<1, [A9_ALU0, A9_ALU1]>], [2, 1]>,
 124   InstrItinData<IIC_iEXTAr, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 125                              InstrStage<2, [A9_ALU0, A9_ALU1]>], [3, 1, 1]>,
 126   InstrItinData<IIC_iEXTAsr,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 127                              InstrStage<3, [A9_ALU0, A9_ALU1]>], [3, 1, 1, 1]>,
 128   //
 129   // Compare instructions
 130   InstrItinData<IIC_iCMPi   , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 131                                InstrStage<1, [A9_ALU0, A9_ALU1]>],
 132                                [1], [A9_LdBypass]>,
 133   InstrItinData<IIC_iCMPr   , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 134                                InstrStage<1, [A9_ALU0, A9_ALU1]>],
 135                                [1, 1], [A9_LdBypass, A9_LdBypass]>,
 136   InstrItinData<IIC_iCMPsi  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 137                                InstrStage<2, [A9_ALU0, A9_ALU1]>],
 138                                 [1, 1], [A9_LdBypass, NoBypass]>,
 139   InstrItinData<IIC_iCMPsr  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 140                                InstrStage<3, [A9_ALU0, A9_ALU1]>],
 141                               [1, 1, 1], [A9_LdBypass, NoBypass, NoBypass]>,
 142   //
 143   // Test instructions
 144   InstrItinData<IIC_iTSTi   , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 145                                InstrStage<1, [A9_ALU0, A9_ALU1]>], [1]>,
 146   InstrItinData<IIC_iTSTr   , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 147                                InstrStage<1, [A9_ALU0, A9_ALU1]>], [1, 1]>,
 148   InstrItinData<IIC_iTSTsi  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 149                                InstrStage<2, [A9_ALU0, A9_ALU1]>], [1, 1]>,
 150   InstrItinData<IIC_iTSTsr  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 151                                InstrStage<3, [A9_ALU0, A9_ALU1]>], [1, 1, 1]>,
 152   //
 153   // Move instructions, conditional
 154   // FIXME: Correctly model the extra input dep on the destination.
 155   InstrItinData<IIC_iCMOVi  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 156                                InstrStage<1, [A9_ALU0, A9_ALU1]>], [1]>,
 157   InstrItinData<IIC_iCMOVr  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 158                                InstrStage<1, [A9_ALU0, A9_ALU1]>], [1, 1]>,
 159   InstrItinData<IIC_iCMOVsi , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 160                                InstrStage<1, [A9_ALU0, A9_ALU1]>], [1, 1]>,
 161   InstrItinData<IIC_iCMOVsr , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 162                                InstrStage<2, [A9_ALU0, A9_ALU1]>], [2, 1, 1]>,
 163   InstrItinData<IIC_iCMOVix2, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 164                                InstrStage<1, [A9_ALU0, A9_ALU1]>,
 165                                InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 166                                InstrStage<1, [A9_ALU0, A9_ALU1]>], [2]>,
 167
 168   // Integer multiply pipeline
 169   //
 170   InstrItinData<IIC_iMUL16  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 171                                InstrStage<2, [A9_ALU0]>], [3, 1, 1]>,
 172   InstrItinData<IIC_iMAC16  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 173                                InstrStage<2, [A9_ALU0]>],
 174                               [3, 1, 1, 1]>,
 175   InstrItinData<IIC_iMUL32  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 176                                InstrStage<2, [A9_ALU0]>], [4, 1, 1]>,
 177   InstrItinData<IIC_iMAC32  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 178                                InstrStage<2, [A9_ALU0]>],
 179                               [4, 1, 1, 1]>,
 180   InstrItinData<IIC_iMUL64  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 181                                InstrStage<3, [A9_ALU0]>], [4, 5, 1, 1]>,
 182   InstrItinData<IIC_iMAC64  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 183                                InstrStage<3, [A9_ALU0]>],
 184                               [4, 5, 1, 1]>,
 185   // Integer load pipeline
 186   // FIXME: The timings are some rough approximations
 187   //
 188   // Immediate offset
 189   InstrItinData<IIC_iLoad_i   , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 190                                  InstrStage<1, [A9_MUX0], 0>,
 191                                  InstrStage<1, [A9_AGU], 0>,
 192                                  InstrStage<1, [A9_LSUnit]>],
 193                                 [3, 1], [A9_LdBypass]>,
 194   InstrItinData<IIC_iLoad_bh_i, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 195                                  InstrStage<1, [A9_MUX0], 0>,
 196                                  InstrStage<2, [A9_AGU], 0>,
 197                                  InstrStage<1, [A9_LSUnit]>],
 198                                 [4, 1], [A9_LdBypass]>,
 199   // FIXME: If address is 64-bit aligned, AGU cycles is 1.
 200   InstrItinData<IIC_iLoad_d_i , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 201                                  InstrStage<1, [A9_MUX0], 0>,
 202                                  InstrStage<2, [A9_AGU], 0>,
 203                                  InstrStage<1, [A9_LSUnit]>],
 204                                 [3, 3, 1], [A9_LdBypass]>,
 205   //
 206   // Register offset
 207   InstrItinData<IIC_iLoad_r   , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 208                                  InstrStage<1, [A9_MUX0], 0>,
 209                                  InstrStage<1, [A9_AGU], 0>,
 210                                  InstrStage<1, [A9_LSUnit]>],
 211                                 [3, 1, 1], [A9_LdBypass]>,
 212   InstrItinData<IIC_iLoad_bh_r, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 213                                  InstrStage<1, [A9_MUX0], 0>,
 214                                  InstrStage<2, [A9_AGU], 0>,
 215                                  InstrStage<1, [A9_LSUnit]>],
 216                                 [4, 1, 1], [A9_LdBypass]>,
 217   InstrItinData<IIC_iLoad_d_r , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 218                                  InstrStage<1, [A9_MUX0], 0>,
 219                                  InstrStage<2, [A9_AGU], 0>,
 220                                  InstrStage<1, [A9_LSUnit]>],
 221                                 [3, 3, 1, 1], [A9_LdBypass]>,
 222   //
 223   // Scaled register offset
 224   InstrItinData<IIC_iLoad_si  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 225                                  InstrStage<1, [A9_MUX0], 0>,
 226                                  InstrStage<1, [A9_AGU], 0>,
 227                                  InstrStage<1, [A9_LSUnit], 0>],
 228                                 [4, 1, 1], [A9_LdBypass]>,
 229   InstrItinData<IIC_iLoad_bh_si,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 230                                  InstrStage<1, [A9_MUX0], 0>,
 231                                  InstrStage<2, [A9_AGU], 0>,
 232                                  InstrStage<1, [A9_LSUnit]>],
 233                                 [5, 1, 1], [A9_LdBypass]>,
 234   //
 235   // Immediate offset with update
 236   InstrItinData<IIC_iLoad_iu  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 237                                  InstrStage<1, [A9_MUX0], 0>,
 238                                  InstrStage<1, [A9_AGU], 0>,
 239                                  InstrStage<1, [A9_LSUnit]>],
 240                                 [3, 2, 1], [A9_LdBypass]>,
 241   InstrItinData<IIC_iLoad_bh_iu,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 242                                  InstrStage<1, [A9_MUX0], 0>,
 243                                  InstrStage<2, [A9_AGU], 0>,
 244                                  InstrStage<1, [A9_LSUnit]>],
 245                                 [4, 3, 1], [A9_LdBypass]>,
 246   //
 247   // Register offset with update
 248   InstrItinData<IIC_iLoad_ru  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 249                                  InstrStage<1, [A9_MUX0], 0>,
 250                                  InstrStage<1, [A9_AGU], 0>,
 251                                  InstrStage<1, [A9_LSUnit]>],
 252                                 [3, 2, 1, 1], [A9_LdBypass]>,
 253   InstrItinData<IIC_iLoad_bh_ru,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 254                                  InstrStage<1, [A9_MUX0], 0>,
 255                                  InstrStage<2, [A9_AGU], 0>,
 256                                  InstrStage<1, [A9_LSUnit]>],
 257                                 [4, 3, 1, 1], [A9_LdBypass]>,
 258   InstrItinData<IIC_iLoad_d_ru, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 259                                  InstrStage<1, [A9_MUX0], 0>,
 260                                  InstrStage<2, [A9_AGU], 0>,
 261                                  InstrStage<1, [A9_LSUnit]>],
 262                                 [3, 3, 1, 1], [A9_LdBypass]>,
 263   //
 264   // Scaled register offset with update
 265   InstrItinData<IIC_iLoad_siu , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 266                                  InstrStage<1, [A9_MUX0], 0>,
 267                                  InstrStage<1, [A9_AGU], 0>,
 268                                  InstrStage<1, [A9_LSUnit]>],
 269                                 [4, 3, 1, 1], [A9_LdBypass]>,
 270   InstrItinData<IIC_iLoad_bh_siu,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 271                                   InstrStage<1, [A9_MUX0], 0>,
 272                                   InstrStage<2, [A9_AGU], 0>,
 273                                   InstrStage<1, [A9_LSUnit]>],
 274                                  [5, 4, 1, 1], [A9_LdBypass]>,
 275   //
 276   // Load multiple, def is the 5th operand.
 277   // FIXME: This assumes 3 to 4 registers.
 278   InstrItinData<IIC_iLoad_m  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 279                                 InstrStage<1, [A9_MUX0], 0>,
 280                                 InstrStage<2, [A9_AGU], 1>,
 281                                 InstrStage<2, [A9_LSUnit]>],
 282                                [1, 1, 1, 1, 3],
 283                          [NoBypass, NoBypass, NoBypass, NoBypass, A9_LdBypass]>,
 284   //
 285   // Load multiple + update, defs are the 1st and 5th operands.
 286   InstrItinData<IIC_iLoad_mu , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 287                                 InstrStage<1, [A9_MUX0], 0>,
 288                                 InstrStage<2, [A9_AGU], 1>,
 289                                 InstrStage<2, [A9_LSUnit]>],
 290                                [2, 1, 1, 1, 3],
 291                          [NoBypass, NoBypass, NoBypass, NoBypass, A9_LdBypass]>,
 292   //
 293   // Load multiple plus branch
 294   InstrItinData<IIC_iLoad_mBr, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 295                                 InstrStage<1, [A9_MUX0], 0>,
 296                                 InstrStage<1, [A9_AGU], 1>,
 297                                 InstrStage<2, [A9_LSUnit]>,
 298                                 InstrStage<1, [A9_Branch]>],
 299                                [1, 2, 1, 1, 3],
 300                          [NoBypass, NoBypass, NoBypass, NoBypass, A9_LdBypass]>,
 301   //
 302   // Pop, def is the 3rd operand.
 303   InstrItinData<IIC_iPop  ,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 304                                 InstrStage<1, [A9_MUX0], 0>,
 305                                 InstrStage<2, [A9_AGU], 1>,
 306                                 InstrStage<2, [A9_LSUnit]>],
 307                                [1, 1, 3],
 308                                [NoBypass, NoBypass, A9_LdBypass]>,
 309   //
 310   // Pop + branch, def is the 3rd operand.
 311   InstrItinData<IIC_iPop_Br,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 312                                 InstrStage<1, [A9_MUX0], 0>,
 313                                 InstrStage<2, [A9_AGU], 1>,
 314                                 InstrStage<2, [A9_LSUnit]>,
 315                                 InstrStage<1, [A9_Branch]>],
 316                                [1, 1, 3],
 317                                [NoBypass, NoBypass, A9_LdBypass]>,
 318
 319   //
 320   // iLoadi + iALUr for t2LDRpci_pic.
 321   InstrItinData<IIC_iLoadiALU, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 322                                 InstrStage<1, [A9_MUX0], 0>,
 323                                 InstrStage<1, [A9_AGU], 0>,
 324                                 InstrStage<1, [A9_LSUnit]>,
 325                                 InstrStage<1, [A9_ALU0, A9_ALU1]>],
 326                                [2, 1]>,
 327
 328   // Integer store pipeline
 329   ///
 330   // Immediate offset
 331   InstrItinData<IIC_iStore_i  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 332                                  InstrStage<1, [A9_MUX0], 0>,
 333                                  InstrStage<1, [A9_AGU], 0>,
 334                                  InstrStage<1, [A9_LSUnit]>], [1, 1]>,
 335   InstrItinData<IIC_iStore_bh_i,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 336                                  InstrStage<1, [A9_MUX0], 0>,
 337                                  InstrStage<2, [A9_AGU], 1>,
 338                                  InstrStage<1, [A9_LSUnit]>], [1, 1]>,
 339   // FIXME: If address is 64-bit aligned, AGU cycles is 1.
 340   InstrItinData<IIC_iStore_d_i, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 341                                  InstrStage<1, [A9_MUX0], 0>,
 342                                  InstrStage<2, [A9_AGU], 1>,
 343                                  InstrStage<1, [A9_LSUnit]>], [1, 1]>,
 344   //
 345   // Register offset
 346   InstrItinData<IIC_iStore_r  , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 347                                  InstrStage<1, [A9_MUX0], 0>,
 348                                  InstrStage<1, [A9_AGU], 0>,
 349                                  InstrStage<1, [A9_LSUnit]>], [1, 1, 1]>,
 350   InstrItinData<IIC_iStore_bh_r,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 351                                  InstrStage<1, [A9_MUX0], 0>,
 352                                  InstrStage<2, [A9_AGU], 1>,
 353                                  InstrStage<1, [A9_LSUnit]>], [1, 1, 1]>,
 354   InstrItinData<IIC_iStore_d_r, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 355                                  InstrStage<1, [A9_MUX0], 0>,
 356                                  InstrStage<2, [A9_AGU], 1>,
 357                                  InstrStage<1, [A9_LSUnit]>], [1, 1, 1]>,
 358   //
 359   // Scaled register offset
 360   InstrItinData<IIC_iStore_si ,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 361                                   InstrStage<1, [A9_MUX0], 0>,
 362                                   InstrStage<1, [A9_AGU], 0>,
 363                                   InstrStage<1, [A9_LSUnit]>], [1, 1, 1]>,
 364   InstrItinData<IIC_iStore_bh_si,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 365                                   InstrStage<1, [A9_MUX0], 0>,
 366                                   InstrStage<2, [A9_AGU], 1>,
 367                                   InstrStage<1, [A9_LSUnit]>], [1, 1, 1]>,
 368   //
 369   // Immediate offset with update
 370   InstrItinData<IIC_iStore_iu ,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 371                                   InstrStage<1, [A9_MUX0], 0>,
 372                                   InstrStage<1, [A9_AGU], 0>,
 373                                   InstrStage<1, [A9_LSUnit]>], [2, 1, 1]>,
 374   InstrItinData<IIC_iStore_bh_iu,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 375                                   InstrStage<1, [A9_MUX0], 0>,
 376                                   InstrStage<2, [A9_AGU], 1>,
 377                                   InstrStage<1, [A9_LSUnit]>], [3, 1, 1]>,
 378   //
 379   // Register offset with update
 380   InstrItinData<IIC_iStore_ru ,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 381                                   InstrStage<1, [A9_MUX0], 0>,
 382                                   InstrStage<1, [A9_AGU], 0>,
 383                                   InstrStage<1, [A9_LSUnit]>],
 384                                  [2, 1, 1, 1]>,
 385   InstrItinData<IIC_iStore_bh_ru,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 386                                   InstrStage<1, [A9_MUX0], 0>,
 387                                   InstrStage<2, [A9_AGU], 1>,
 388                                   InstrStage<1, [A9_LSUnit]>],
 389                                  [3, 1, 1, 1]>,
 390   InstrItinData<IIC_iStore_d_ru, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 391                                   InstrStage<1, [A9_MUX0], 0>,
 392                                   InstrStage<2, [A9_AGU], 1>,
 393                                   InstrStage<1, [A9_LSUnit]>],
 394                                  [3, 1, 1, 1]>,
 395   //
 396   // Scaled register offset with update
 397   InstrItinData<IIC_iStore_siu,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 398                                     InstrStage<1, [A9_MUX0], 0>,
 399                                     InstrStage<1, [A9_AGU], 0>,
 400                                     InstrStage<1, [A9_LSUnit]>],
 401                                    [2, 1, 1, 1]>,
 402   InstrItinData<IIC_iStore_bh_siu, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 403                                     InstrStage<1, [A9_MUX0], 0>,
 404                                     InstrStage<2, [A9_AGU], 1>,
 405                                     InstrStage<1, [A9_LSUnit]>],
 406                                    [3, 1, 1, 1]>,
 407   //
 408   // Store multiple
 409   InstrItinData<IIC_iStore_m , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 410                                 InstrStage<1, [A9_MUX0], 0>,
 411                                 InstrStage<1, [A9_AGU], 0>,
 412                                 InstrStage<2, [A9_LSUnit]>]>,
 413   //
 414   // Store multiple + update
 415   InstrItinData<IIC_iStore_mu, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 416                                 InstrStage<1, [A9_MUX0], 0>,
 417                                 InstrStage<1, [A9_AGU], 0>,
 418                                 InstrStage<2, [A9_LSUnit]>], [2]>,
 419
 420   //
 421   // Preload
 422   InstrItinData<IIC_Preload,   [InstrStage<1, [A9_Issue0, A9_Issue1]>], [1, 1]>,
 423
 424   // Branch
 425   //
 426   // no delay slots, so the latency of a branch is unimportant
 427   InstrItinData<IIC_Br       , [InstrStage<1, [A9_Issue0], 0>,
 428                                 InstrStage<1, [A9_Issue1], 0>,
 429                                 InstrStage<1, [A9_Branch]>]>,
 430
 431   // VFP and NEON shares the same register file. This means that every VFP
 432   // instruction should wait for full completion of the consecutive NEON
 433   // instruction and vice-versa. We model this behavior with two artificial FUs:
 434   // DRegsVFP and DRegsVFP.
 435   //
 436   // Every VFP instruction:
 437   //  - Acquires DRegsVFP resource for 1 cycle
 438   //  - Reserves DRegsN resource for the whole duration (including time to
 439   //    register file writeback!).
 440   // Every NEON instruction does the same but with FUs swapped.
 441   //
 442   // Since the reserved FU cannot be acquired, this models precisely
 443   // "cross-domain" stalls.
 444
 445   // VFP
 446   // Issue through integer pipeline, and execute in NEON unit.
 447
 448   // FP Special Register to Integer Register File Move
 449   InstrItinData<IIC_fpSTAT , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 450                               InstrStage<1, [A9_MUX0], 0>,
 451                               InstrStage<1, [A9_DRegsVFP], 0, Required>,
 452                               InstrStage<2, [A9_DRegsN],   0, Reserved>,
 453                               InstrStage<1, [A9_NPipe]>],
 454                              [1]>,
 455   //
 456   // Single-precision FP Unary
 457   InstrItinData<IIC_fpUNA32 , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 458                                InstrStage<1, [A9_MUX0], 0>,
 459                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 460                                // Extra latency cycles since wbck is 2 cycles
 461                                InstrStage<3, [A9_DRegsN],   0, Reserved>,
 462                                InstrStage<1, [A9_NPipe]>],
 463                               [1, 1]>,
 464   //
 465   // Double-precision FP Unary
 466   InstrItinData<IIC_fpUNA64 , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 467                                InstrStage<1, [A9_MUX0], 0>,
 468                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 469                                // Extra latency cycles since wbck is 2 cycles
 470                                InstrStage<3, [A9_DRegsN],   0, Reserved>,
 471                                InstrStage<1, [A9_NPipe]>],
 472                               [1, 1]>,
 473
 474   //
 475   // Single-precision FP Compare
 476   InstrItinData<IIC_fpCMP32 , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 477                                InstrStage<1, [A9_MUX0], 0>,
 478                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 479                                // Extra latency cycles since wbck is 4 cycles
 480                                InstrStage<5, [A9_DRegsN],   0, Reserved>,
 481                                InstrStage<1, [A9_NPipe]>],
 482                               [1, 1]>,
 483   //
 484   // Double-precision FP Compare
 485   InstrItinData<IIC_fpCMP64 , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 486                                InstrStage<1, [A9_MUX0], 0>,
 487                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 488                                // Extra latency cycles since wbck is 4 cycles
 489                                InstrStage<5, [A9_DRegsN],   0, Reserved>,
 490                                InstrStage<1, [A9_NPipe]>],
 491                               [1, 1]>,
 492   //
 493   // Single to Double FP Convert
 494   InstrItinData<IIC_fpCVTSD , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 495                                InstrStage<1, [A9_MUX0], 0>,
 496                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 497                                InstrStage<5, [A9_DRegsN],   0, Reserved>,
 498                                InstrStage<1, [A9_NPipe]>],
 499                               [4, 1]>,
 500   //
 501   // Double to Single FP Convert
 502   InstrItinData<IIC_fpCVTDS , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 503                                InstrStage<1, [A9_MUX0], 0>,
 504                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 505                                InstrStage<5, [A9_DRegsN],   0, Reserved>,
 506                                InstrStage<1, [A9_NPipe]>],
 507                               [4, 1]>,
 508
 509   //
 510   // Single to Half FP Convert
 511   InstrItinData<IIC_fpCVTSH , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 512                                InstrStage<1, [A9_MUX0], 0>,
 513                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 514                                InstrStage<5, [A9_DRegsN],   0, Reserved>,
 515                                InstrStage<1, [A9_NPipe]>],
 516                               [4, 1]>,
 517   //
 518   // Half to Single FP Convert
 519   InstrItinData<IIC_fpCVTHS , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 520                                InstrStage<1, [A9_MUX0], 0>,
 521                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 522                                InstrStage<3, [A9_DRegsN],   0, Reserved>,
 523                                InstrStage<1, [A9_NPipe]>],
 524                               [2, 1]>,
 525
 526   //
 527   // Single-Precision FP to Integer Convert
 528   InstrItinData<IIC_fpCVTSI , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 529                                InstrStage<1, [A9_MUX0], 0>,
 530                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 531                                InstrStage<5, [A9_DRegsN],   0, Reserved>,
 532                                InstrStage<1, [A9_NPipe]>],
 533                               [4, 1]>,
 534   //
 535   // Double-Precision FP to Integer Convert
 536   InstrItinData<IIC_fpCVTDI , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 537                                InstrStage<1, [A9_MUX0], 0>,
 538                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 539                                InstrStage<5, [A9_DRegsN],   0, Reserved>,
 540                                InstrStage<1, [A9_NPipe]>],
 541                               [4, 1]>,
 542   //
 543   // Integer to Single-Precision FP Convert
 544   InstrItinData<IIC_fpCVTIS , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 545                                InstrStage<1, [A9_MUX0], 0>,
 546                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 547                                InstrStage<5, [A9_DRegsN],   0, Reserved>,
 548                                InstrStage<1, [A9_NPipe]>],
 549                               [4, 1]>,
 550   //
 551   // Integer to Double-Precision FP Convert
 552   InstrItinData<IIC_fpCVTID , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 553                                InstrStage<1, [A9_MUX0], 0>,
 554                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 555                                InstrStage<5, [A9_DRegsN],   0, Reserved>,
 556                                InstrStage<1, [A9_NPipe]>],
 557                               [4, 1]>,
 558   //
 559   // Single-precision FP ALU
 560   InstrItinData<IIC_fpALU32 , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 561                                InstrStage<1, [A9_MUX0], 0>,
 562                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 563                                InstrStage<5, [A9_DRegsN],   0, Reserved>,
 564                                InstrStage<1, [A9_NPipe]>],
 565                               [4, 1, 1]>,
 566   //
 567   // Double-precision FP ALU
 568   InstrItinData<IIC_fpALU64 , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 569                                InstrStage<1, [A9_MUX0], 0>,
 570                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 571                                InstrStage<5, [A9_DRegsN],   0, Reserved>,
 572                                InstrStage<1, [A9_NPipe]>],
 573                               [4, 1, 1]>,
 574   //
 575   // Single-precision FP Multiply
 576   InstrItinData<IIC_fpMUL32 , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 577                                InstrStage<1, [A9_MUX0], 0>,
 578                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 579                                InstrStage<6, [A9_DRegsN],   0, Reserved>,
 580                                InstrStage<1, [A9_NPipe]>],
 581                               [5, 1, 1]>,
 582   //
 583   // Double-precision FP Multiply
 584   InstrItinData<IIC_fpMUL64 , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 585                                InstrStage<1, [A9_MUX0], 0>,
 586                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 587                                InstrStage<7, [A9_DRegsN],   0, Reserved>,
 588                                InstrStage<2, [A9_NPipe]>],
 589                               [6, 1, 1]>,
 590   //
 591   // Single-precision FP MAC
 592   InstrItinData<IIC_fpMAC32 , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 593                                InstrStage<1, [A9_MUX0], 0>,
 594                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 595                                InstrStage<9, [A9_DRegsN],   0, Reserved>,
 596                                InstrStage<1, [A9_NPipe]>],
 597                               [8, 1, 1, 1]>,
 598   //
 599   // Double-precision FP MAC
 600   InstrItinData<IIC_fpMAC64 , [InstrStage<1,  [A9_Issue0, A9_Issue1], 0>,
 601                                InstrStage<1,  [A9_MUX0], 0>,
 602                                InstrStage<1,  [A9_DRegsVFP], 0, Required>,
 603                                InstrStage<10, [A9_DRegsN],  0, Reserved>,
 604                                InstrStage<2,  [A9_NPipe]>],
 605                               [9, 1, 1, 1]>,
 606   //
 607   // Single-precision FP DIV
 608   InstrItinData<IIC_fpDIV32 , [InstrStage<1,  [A9_Issue0, A9_Issue1], 0>,
 609                                InstrStage<1,  [A9_MUX0], 0>,
 610                                InstrStage<1,  [A9_DRegsVFP], 0, Required>,
 611                                InstrStage<16, [A9_DRegsN],  0, Reserved>,
 612                                InstrStage<10, [A9_NPipe]>],
 613                               [15, 1, 1]>,
 614   //
 615   // Double-precision FP DIV
 616   InstrItinData<IIC_fpDIV64 , [InstrStage<1,  [A9_Issue0, A9_Issue1], 0>,
 617                                InstrStage<1,  [A9_MUX0], 0>,
 618                                InstrStage<1,  [A9_DRegsVFP], 0, Required>,
 619                                InstrStage<26, [A9_DRegsN],  0, Reserved>,
 620                                InstrStage<20, [A9_NPipe]>],
 621                               [25, 1, 1]>,
 622   //
 623   // Single-precision FP SQRT
 624   InstrItinData<IIC_fpSQRT32, [InstrStage<1,  [A9_Issue0, A9_Issue1], 0>,
 625                                InstrStage<1,  [A9_MUX0], 0>,
 626                                InstrStage<1,  [A9_DRegsVFP], 0, Required>,
 627                                InstrStage<18, [A9_DRegsN],   0, Reserved>,
 628                                InstrStage<13, [A9_NPipe]>],
 629                               [17, 1]>,
 630   //
 631   // Double-precision FP SQRT
 632   InstrItinData<IIC_fpSQRT64, [InstrStage<1,  [A9_Issue0, A9_Issue1], 0>,
 633                                InstrStage<1,  [A9_MUX0], 0>,
 634                                InstrStage<1,  [A9_DRegsVFP], 0, Required>,
 635                                InstrStage<33, [A9_DRegsN],   0, Reserved>,
 636                                InstrStage<28, [A9_NPipe]>],
 637                               [32, 1]>,
 638
 639   //
 640   // Integer to Single-precision Move
 641   InstrItinData<IIC_fpMOVIS,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 642                                InstrStage<1, [A9_MUX0], 0>,
 643                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 644                                // Extra 1 latency cycle since wbck is 2 cycles
 645                                InstrStage<3, [A9_DRegsN],   0, Reserved>,
 646                                InstrStage<1, [A9_NPipe]>],
 647                               [1, 1]>,
 648   //
 649   // Integer to Double-precision Move
 650   InstrItinData<IIC_fpMOVID,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 651                                InstrStage<1, [A9_MUX0], 0>,
 652                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 653                                // Extra 1 latency cycle since wbck is 2 cycles
 654                                InstrStage<3, [A9_DRegsN],   0, Reserved>,
 655                                InstrStage<1, [A9_NPipe]>],
 656                               [1, 1, 1]>,
 657   //
 658   // Single-precision to Integer Move
 659   InstrItinData<IIC_fpMOVSI,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 660                                InstrStage<1, [A9_MUX0], 0>,
 661                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 662                                InstrStage<2, [A9_DRegsN],   0, Reserved>,
 663                                InstrStage<1, [A9_NPipe]>],
 664                               [2, 1]>,
 665   //
 666   // Double-precision to Integer Move
 667   InstrItinData<IIC_fpMOVDI,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 668                                InstrStage<1, [A9_MUX0], 0>,
 669                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 670                                InstrStage<2, [A9_DRegsN],   0, Reserved>,
 671                                InstrStage<1, [A9_NPipe]>],
 672                               [2, 1, 1]>,
 673   //
 674   // Single-precision FP Load
 675   InstrItinData<IIC_fpLoad32, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 676                                InstrStage<1, [A9_MUX0], 0>,
 677                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 678                                InstrStage<2, [A9_DRegsN],   0, Reserved>,
 679                                InstrStage<1, [A9_NPipe], 0>,
 680                                InstrStage<1, [A9_LSUnit]>],
 681                               [1, 1]>,
 682   //
 683   // Double-precision FP Load
 684   // FIXME: Result latency is 1 if address is 64-bit aligned.
 685   InstrItinData<IIC_fpLoad64, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 686                                InstrStage<1, [A9_MUX0], 0>,
 687                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 688                                InstrStage<2, [A9_DRegsN],   0, Reserved>,
 689                                InstrStage<1, [A9_NPipe], 0>,
 690                                InstrStage<1, [A9_LSUnit]>],
 691                               [2, 1]>,
 692   //
 693   // FP Load Multiple
 694   InstrItinData<IIC_fpLoad_m, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 695                                InstrStage<1, [A9_MUX0], 0>,
 696                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 697                                InstrStage<2, [A9_DRegsN],   0, Reserved>,
 698                                InstrStage<1, [A9_NPipe], 0>,
 699                                InstrStage<1, [A9_LSUnit]>], [1, 1, 1, 1]>,
 700   //
 701   // FP Load Multiple + update
 702   InstrItinData<IIC_fpLoad_mu,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 703                                InstrStage<1, [A9_MUX0], 0>,
 704                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 705                                InstrStage<2, [A9_DRegsN],   0, Reserved>,
 706                                InstrStage<1, [A9_NPipe], 0>,
 707                                InstrStage<1, [A9_LSUnit]>], [2, 1, 1, 1]>,
 708   //
 709   // Single-precision FP Store
 710   InstrItinData<IIC_fpStore32,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 711                                InstrStage<1, [A9_MUX0], 0>,
 712                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 713                                InstrStage<2, [A9_DRegsN],   0, Reserved>,
 714                                InstrStage<1, [A9_NPipe], 0>,
 715                                InstrStage<1, [A9_LSUnit]>],
 716                               [1, 1]>,
 717   //
 718   // Double-precision FP Store
 719   InstrItinData<IIC_fpStore64,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 720                                InstrStage<1, [A9_MUX0], 0>,
 721                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 722                                InstrStage<2, [A9_DRegsN],   0, Reserved>,
 723                                InstrStage<1, [A9_NPipe], 0>,
 724                                InstrStage<1, [A9_LSUnit]>],
 725                               [1, 1]>,
 726   //
 727   // FP Store Multiple
 728   InstrItinData<IIC_fpStore_m,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 729                                InstrStage<1, [A9_MUX0], 0>,
 730                                InstrStage<1, [A9_DRegsVFP], 0, Required>,
 731                                InstrStage<2, [A9_DRegsN],   0, Reserved>,
 732                                InstrStage<1, [A9_NPipe], 0>,
 733                                InstrStage<1, [A9_LSUnit]>], [1, 1, 1, 1]>,
 734   //
 735   // FP Store Multiple + update
 736   InstrItinData<IIC_fpStore_mu,[InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 737                                 InstrStage<1, [A9_MUX0], 0>,
 738                                 InstrStage<1, [A9_DRegsVFP], 0, Required>,
 739                                 InstrStage<2, [A9_DRegsN],   0, Reserved>,
 740                                 InstrStage<1, [A9_NPipe], 0>,
 741                                 InstrStage<1, [A9_LSUnit]>], [2, 1, 1, 1]>,
 742   // NEON
 743   // VLD1
 744   // FIXME: Conservatively assume insufficent alignment.
 745   InstrItinData<IIC_VLD1,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 746                                InstrStage<1, [A9_MUX0], 0>,
 747                                InstrStage<1, [A9_DRegsN],   0, Required>,
 748                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
 749                                InstrStage<2, [A9_NPipe], 0>,
 750                                InstrStage<2, [A9_LSUnit]>],
 751                               [2, 1]>,
 752   // VLD1x2
 753   InstrItinData<IIC_VLD1x2,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 754                                InstrStage<1, [A9_MUX0], 0>,
 755                                InstrStage<1, [A9_DRegsN],   0, Required>,
 756                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
 757                                InstrStage<2, [A9_NPipe], 0>,
 758                                InstrStage<2, [A9_LSUnit]>],
 759                               [2, 2, 1]>,
 760   // VLD1x3
 761   InstrItinData<IIC_VLD1x3,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 762                                InstrStage<1, [A9_MUX0], 0>,
 763                                InstrStage<1, [A9_DRegsN],   0, Required>,
 764                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
 765                                InstrStage<3, [A9_NPipe], 0>,
 766                                InstrStage<3, [A9_LSUnit]>],
 767                               [2, 2, 3, 1]>,
 768   // VLD1x4
 769   InstrItinData<IIC_VLD1x4,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 770                                InstrStage<1, [A9_MUX0], 0>,
 771                                InstrStage<1, [A9_DRegsN],   0, Required>,
 772                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
 773                                InstrStage<3, [A9_NPipe], 0>,
 774                                InstrStage<3, [A9_LSUnit]>],
 775                               [2, 2, 3, 3, 1]>,
 776   // VLD1u
 777   InstrItinData<IIC_VLD1u,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 778                                InstrStage<1, [A9_MUX0], 0>,
 779                                InstrStage<1, [A9_DRegsN],   0, Required>,
 780                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
 781                                InstrStage<2, [A9_NPipe], 0>,
 782                                InstrStage<2, [A9_LSUnit]>],
 783                               [2, 2, 1]>,
 784   // VLD1x2u
 785   InstrItinData<IIC_VLD1x2u,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 786                                InstrStage<1, [A9_MUX0], 0>,
 787                                InstrStage<1, [A9_DRegsN],   0, Required>,
 788                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
 789                                InstrStage<2, [A9_NPipe], 0>,
 790                                InstrStage<2, [A9_LSUnit]>],
 791                               [2, 2, 2, 1]>,
 792   // VLD1x3u
 793   InstrItinData<IIC_VLD1x3u,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 794                                InstrStage<1, [A9_MUX0], 0>,
 795                                InstrStage<1, [A9_DRegsN],   0, Required>,
 796                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
 797                                InstrStage<3, [A9_NPipe], 0>,
 798                                InstrStage<3, [A9_LSUnit]>],
 799                               [2, 2, 3, 2, 1]>,
 800   // VLD1x4u
 801   InstrItinData<IIC_VLD1x4u,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 802                                InstrStage<1, [A9_MUX0], 0>,
 803                                InstrStage<1, [A9_DRegsN],   0, Required>,
 804                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
 805                                InstrStage<3, [A9_NPipe], 0>,
 806                                InstrStage<3, [A9_LSUnit]>],
 807                               [2, 2, 3, 3, 2, 1]>,
 808   //
 809   // VLD1ln
 810   InstrItinData<IIC_VLD1ln,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 811                                InstrStage<1, [A9_MUX0], 0>,
 812                                InstrStage<1, [A9_DRegsN],   0, Required>,
 813                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
 814                                InstrStage<3, [A9_NPipe], 0>,
 815                                InstrStage<3, [A9_LSUnit]>],
 816                               [4, 1, 1, 1]>,
 817   //
 818   // VLD1lnu
 819   InstrItinData<IIC_VLD1lnu,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 820                                InstrStage<1, [A9_MUX0], 0>,
 821                                InstrStage<1, [A9_DRegsN],   0, Required>,
 822                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
 823                                InstrStage<3, [A9_NPipe], 0>,
 824                                InstrStage<3, [A9_LSUnit]>],
 825                               [4, 2, 1, 1, 1, 1]>,
 826   //
 827   // VLD1dup
 828   InstrItinData<IIC_VLD1dup,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 829                                InstrStage<1, [A9_MUX0], 0>,
 830                                InstrStage<1, [A9_DRegsN],   0, Required>,
 831                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
 832                                InstrStage<2, [A9_NPipe], 0>,
 833                                InstrStage<2, [A9_LSUnit]>],
 834                               [3, 1]>,
 835   //
 836   // VLD1dupu
 837   InstrItinData<IIC_VLD1dupu, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 838                                InstrStage<1, [A9_MUX0], 0>,
 839                                InstrStage<1, [A9_DRegsN],   0, Required>,
 840                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
 841                                InstrStage<2, [A9_NPipe], 0>,
 842                                InstrStage<2, [A9_LSUnit]>],
 843                               [3, 2, 1, 1]>,
 844   //
 845   // VLD2
 846   InstrItinData<IIC_VLD2,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 847                                InstrStage<1, [A9_MUX0], 0>,
 848                                InstrStage<1, [A9_DRegsN],   0, Required>,
 849                                // Extra latency cycles since wbck is 7 cycles
 850                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
 851                                InstrStage<2, [A9_NPipe], 0>,
 852                                InstrStage<2, [A9_LSUnit]>],
 853                               [3, 3, 1]>,
 854   //
 855   // VLD2x2
 856   InstrItinData<IIC_VLD2x2,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 857                                InstrStage<1, [A9_MUX0], 0>,
 858                                InstrStage<1, [A9_DRegsN],   0, Required>,
 859                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
 860                                InstrStage<3, [A9_NPipe], 0>,
 861                                InstrStage<3, [A9_LSUnit]>],
 862                               [3, 4, 3, 4, 1]>,
 863   //
 864   // VLD2ln
 865   InstrItinData<IIC_VLD2ln,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 866                                InstrStage<1, [A9_MUX0], 0>,
 867                                InstrStage<1, [A9_DRegsN],   0, Required>,
 868                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
 869                                InstrStage<3, [A9_NPipe], 0>,
 870                                InstrStage<3, [A9_LSUnit]>],
 871                               [4, 4, 1, 1, 1, 1]>,
 872   //
 873   // VLD2u
 874   InstrItinData<IIC_VLD2u,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 875                                InstrStage<1, [A9_MUX0], 0>,
 876                                InstrStage<1, [A9_DRegsN],   0, Required>,
 877                                // Extra latency cycles since wbck is 7 cycles
 878                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
 879                                InstrStage<2, [A9_NPipe], 0>,
 880                                InstrStage<2, [A9_LSUnit]>],
 881                               [3, 3, 2, 1, 1, 1]>,
 882   //
 883   // VLD2x2u
 884   InstrItinData<IIC_VLD2x2u,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 885                                InstrStage<1, [A9_MUX0], 0>,
 886                                InstrStage<1, [A9_DRegsN],   0, Required>,
 887                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
 888                                InstrStage<3, [A9_NPipe], 0>,
 889                                InstrStage<3, [A9_LSUnit]>],
 890                               [3, 4, 3, 4, 2, 1]>,
 891   //
 892   // VLD2lnu
 893   InstrItinData<IIC_VLD2lnu,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 894                                InstrStage<1, [A9_MUX0], 0>,
 895                                InstrStage<1, [A9_DRegsN],   0, Required>,
 896                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
 897                                InstrStage<3, [A9_NPipe], 0>,
 898                                InstrStage<3, [A9_LSUnit]>],
 899                               [4, 4, 2, 1, 1, 1, 1, 1]>,
 900   //
 901   // VLD2dup
 902   InstrItinData<IIC_VLD2dup,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 903                                InstrStage<1, [A9_MUX0], 0>,
 904                                InstrStage<1, [A9_DRegsN],   0, Required>,
 905                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
 906                                InstrStage<2, [A9_NPipe], 0>,
 907                                InstrStage<2, [A9_LSUnit]>],
 908                               [3, 3, 1]>,
 909   //
 910   // VLD2dupu
 911   InstrItinData<IIC_VLD2dupu, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 912                                InstrStage<1, [A9_MUX0], 0>,
 913                                InstrStage<1, [A9_DRegsN],   0, Required>,
 914                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
 915                                InstrStage<2, [A9_NPipe], 0>,
 916                                InstrStage<2, [A9_LSUnit]>],
 917                               [3, 3, 2, 1, 1]>,
 918   //
 919   // VLD3
 920   InstrItinData<IIC_VLD3,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 921                                InstrStage<1, [A9_MUX0], 0>,
 922                                InstrStage<1, [A9_DRegsN],   0, Required>,
 923                                InstrStage<10,[A9_DRegsVFP], 0, Reserved>,
 924                                InstrStage<4, [A9_NPipe], 0>,
 925                                InstrStage<4, [A9_LSUnit]>],
 926                               [4, 4, 5, 1]>,
 927   //
 928   // VLD3ln
 929   InstrItinData<IIC_VLD3ln,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 930                                InstrStage<1, [A9_MUX0], 0>,
 931                                InstrStage<1, [A9_DRegsN],   0, Required>,
 932                                InstrStage<11,[A9_DRegsVFP], 0, Reserved>,
 933                                InstrStage<5, [A9_NPipe], 0>,
 934                                InstrStage<5, [A9_LSUnit]>],
 935                               [5, 5, 6, 1, 1, 1, 1, 2]>,
 936   //
 937   // VLD3u
 938   InstrItinData<IIC_VLD3u,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 939                                InstrStage<1, [A9_MUX0], 0>,
 940                                InstrStage<1, [A9_DRegsN],   0, Required>,
 941                                InstrStage<10,[A9_DRegsVFP], 0, Reserved>,
 942                                InstrStage<4, [A9_NPipe], 0>,
 943                                InstrStage<4, [A9_LSUnit]>],
 944                               [4, 4, 5, 2, 1]>,
 945   //
 946   // VLD3lnu
 947   InstrItinData<IIC_VLD3lnu,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 948                                InstrStage<1, [A9_MUX0], 0>,
 949                                InstrStage<1, [A9_DRegsN],   0, Required>,
 950                                InstrStage<11,[A9_DRegsVFP], 0, Reserved>,
 951                                InstrStage<5, [A9_NPipe], 0>,
 952                                InstrStage<5, [A9_LSUnit]>],
 953                               [5, 5, 6, 2, 1, 1, 1, 1, 1, 2]>,
 954   //
 955   // VLD3dup
 956   InstrItinData<IIC_VLD3dup,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 957                                InstrStage<1, [A9_MUX0], 0>,
 958                                InstrStage<1, [A9_DRegsN],   0, Required>,
 959                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
 960                                InstrStage<3, [A9_NPipe], 0>,
 961                                InstrStage<3, [A9_LSUnit]>],
 962                               [3, 3, 4, 1]>,
 963   //
 964   // VLD3dupu
 965   InstrItinData<IIC_VLD3dupu, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 966                                InstrStage<1, [A9_MUX0], 0>,
 967                                InstrStage<1, [A9_DRegsN],   0, Required>,
 968                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
 969                                InstrStage<3, [A9_NPipe], 0>,
 970                                InstrStage<3, [A9_LSUnit]>],
 971                               [3, 3, 4, 2, 1, 1]>,
 972   //
 973   // VLD4
 974   InstrItinData<IIC_VLD4,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 975                                InstrStage<1, [A9_MUX0], 0>,
 976                                InstrStage<1, [A9_DRegsN],   0, Required>,
 977                                InstrStage<10,[A9_DRegsVFP], 0, Reserved>,
 978                                InstrStage<4, [A9_NPipe], 0>,
 979                                InstrStage<4, [A9_LSUnit]>],
 980                               [4, 4, 5, 5, 1]>,
 981   //
 982   // VLD4ln
 983   InstrItinData<IIC_VLD4ln,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 984                                InstrStage<1, [A9_MUX0], 0>,
 985                                InstrStage<1, [A9_DRegsN],   0, Required>,
 986                                InstrStage<11,[A9_DRegsVFP], 0, Reserved>,
 987                                InstrStage<5, [A9_NPipe], 0>,
 988                                InstrStage<5, [A9_LSUnit]>],
 989                               [5, 5, 6, 6, 1, 1, 1, 1, 2, 2]>,
 990   //
 991   // VLD4u
 992   InstrItinData<IIC_VLD4u,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
 993                                InstrStage<1, [A9_MUX0], 0>,
 994                                InstrStage<1, [A9_DRegsN],   0, Required>,
 995                                InstrStage<10,[A9_DRegsVFP], 0, Reserved>,
 996                                InstrStage<4, [A9_NPipe], 0>,
 997                                InstrStage<4, [A9_LSUnit]>],
 998                               [4, 4, 5, 5, 2, 1]>,
 999   //
1000   // VLD4lnu
1001   InstrItinData<IIC_VLD4lnu,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1002                                InstrStage<1, [A9_MUX0], 0>,
1003                                InstrStage<1, [A9_DRegsN],   0, Required>,
1004                                InstrStage<11,[A9_DRegsVFP], 0, Reserved>,
1005                                InstrStage<5, [A9_NPipe], 0>,
1006                                InstrStage<5, [A9_LSUnit]>],
1007                               [5, 5, 6, 6, 2, 1, 1, 1, 1, 1, 2, 2]>,
1008   //
1009   // VLD4dup
1010   InstrItinData<IIC_VLD4dup,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1011                                InstrStage<1, [A9_MUX0], 0>,
1012                                InstrStage<1, [A9_DRegsN],   0, Required>,
1013                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
1014                                InstrStage<3, [A9_NPipe], 0>,
1015                                InstrStage<3, [A9_LSUnit]>],
1016                               [3, 3, 4, 4, 1]>,
1017   //
1018   // VLD4dupu
1019   InstrItinData<IIC_VLD4dupu, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1020                                InstrStage<1, [A9_MUX0], 0>,
1021                                InstrStage<1, [A9_DRegsN],   0, Required>,
1022                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
1023                                InstrStage<3, [A9_NPipe], 0>,
1024                                InstrStage<3, [A9_LSUnit]>],
1025                               [3, 3, 4, 4, 2, 1, 1]>,
1026   //
1027   // VST1
1028   InstrItinData<IIC_VST1,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1029                                InstrStage<1, [A9_MUX0], 0>,
1030                                InstrStage<1, [A9_DRegsN],   0, Required>,
1031                                InstrStage<2, [A9_DRegsVFP], 0, Reserved>,
1032                                InstrStage<2, [A9_NPipe], 0>,
1033                                InstrStage<2, [A9_LSUnit]>],
1034                               [1, 1, 1]>,
1035   //
1036   // VST1x2
1037   InstrItinData<IIC_VST1x2,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1038                                InstrStage<1, [A9_MUX0], 0>,
1039                                InstrStage<1, [A9_DRegsN],   0, Required>,
1040                                InstrStage<2, [A9_DRegsVFP], 0, Reserved>,
1041                                InstrStage<2, [A9_NPipe], 0>,
1042                                InstrStage<2, [A9_LSUnit]>],
1043                               [1, 1, 1, 1]>,
1044   //
1045   // VST1x3
1046   InstrItinData<IIC_VST1x3,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1047                                InstrStage<1, [A9_MUX0], 0>,
1048                                InstrStage<1, [A9_DRegsN],   0, Required>,
1049                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1050                                InstrStage<3, [A9_NPipe], 0>,
1051                                InstrStage<3, [A9_LSUnit]>],
1052                               [1, 1, 1, 1, 2]>,
1053   //
1054   // VST1x4
1055   InstrItinData<IIC_VST1x4,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1056                                InstrStage<1, [A9_MUX0], 0>,
1057                                InstrStage<1, [A9_DRegsN],   0, Required>,
1058                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1059                                InstrStage<3, [A9_NPipe], 0>,
1060                                InstrStage<3, [A9_LSUnit]>],
1061                               [1, 1, 1, 1, 2, 2]>,
1062   //
1063   // VST1u
1064   InstrItinData<IIC_VST1u,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1065                                InstrStage<1, [A9_MUX0], 0>,
1066                                InstrStage<1, [A9_DRegsN],   0, Required>,
1067                                InstrStage<2, [A9_DRegsVFP], 0, Reserved>,
1068                                InstrStage<2, [A9_NPipe], 0>,
1069                                InstrStage<2, [A9_LSUnit]>],
1070                               [2, 1, 1, 1, 1]>,
1071   //
1072   // VST1x2u
1073   InstrItinData<IIC_VST1x2u,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1074                                InstrStage<1, [A9_MUX0], 0>,
1075                                InstrStage<1, [A9_DRegsN],   0, Required>,
1076                                InstrStage<2, [A9_DRegsVFP], 0, Reserved>,
1077                                InstrStage<2, [A9_NPipe], 0>,
1078                                InstrStage<2, [A9_LSUnit]>],
1079                               [2, 1, 1, 1, 1, 1]>,
1080   //
1081   // VST1x3u
1082   InstrItinData<IIC_VST1x3u,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1083                                InstrStage<1, [A9_MUX0], 0>,
1084                                InstrStage<1, [A9_DRegsN],   0, Required>,
1085                                InstrStage<2, [A9_DRegsVFP], 0, Reserved>,
1086                                InstrStage<3, [A9_NPipe], 0>,
1087                                InstrStage<3, [A9_LSUnit]>],
1088                               [2, 1, 1, 1, 1, 1, 2]>,
1089   //
1090   // VST1x4u
1091   InstrItinData<IIC_VST1x4u,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1092                                InstrStage<1, [A9_MUX0], 0>,
1093                                InstrStage<1, [A9_DRegsN],   0, Required>,
1094                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1095                                InstrStage<3, [A9_NPipe], 0>,
1096                                InstrStage<3, [A9_LSUnit]>],
1097                               [2, 1, 1, 1, 1, 1, 2, 2]>,
1098   //
1099   // VST1ln
1100   InstrItinData<IIC_VST1ln,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1101                                InstrStage<1, [A9_MUX0], 0>,
1102                                InstrStage<1, [A9_DRegsN],   0, Required>,
1103                                InstrStage<2, [A9_DRegsVFP], 0, Reserved>,
1104                                InstrStage<2, [A9_NPipe], 0>,
1105                                InstrStage<2, [A9_LSUnit]>],
1106                               [1, 1, 1]>,
1107   //
1108   // VST1lnu
1109   InstrItinData<IIC_VST1lnu,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1110                                InstrStage<1, [A9_MUX0], 0>,
1111                                InstrStage<1, [A9_DRegsN],   0, Required>,
1112                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1113                                InstrStage<3, [A9_NPipe], 0>,
1114                                InstrStage<3, [A9_LSUnit]>],
1115                               [2, 1, 1, 1, 1]>,
1116   //
1117   // VST2
1118   InstrItinData<IIC_VST2,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1119                                InstrStage<1, [A9_MUX0], 0>,
1120                                InstrStage<1, [A9_DRegsN],   0, Required>,
1121                                InstrStage<2, [A9_DRegsVFP], 0, Reserved>,
1122                                InstrStage<2, [A9_NPipe], 0>,
1123                                InstrStage<2, [A9_LSUnit]>],
1124                               [1, 1, 1, 1]>,
1125   //
1126   // VST2x2
1127   InstrItinData<IIC_VST2x2,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1128                                InstrStage<1, [A9_MUX0], 0>,
1129                                InstrStage<1, [A9_DRegsN],   0, Required>,
1130                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1131                                InstrStage<3, [A9_NPipe], 0>,
1132                                InstrStage<3, [A9_LSUnit]>],
1133                               [1, 1, 1, 1, 2, 2]>,
1134   //
1135   // VST2u
1136   InstrItinData<IIC_VST2u,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1137                                InstrStage<1, [A9_MUX0], 0>,
1138                                InstrStage<1, [A9_DRegsN],   0, Required>,
1139                                InstrStage<2, [A9_DRegsVFP], 0, Reserved>,
1140                                InstrStage<2, [A9_NPipe], 0>,
1141                                InstrStage<2, [A9_LSUnit]>],
1142                               [2, 1, 1, 1, 1, 1]>,
1143   //
1144   // VST2x2u
1145   InstrItinData<IIC_VST2x2u,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1146                                InstrStage<1, [A9_MUX0], 0>,
1147                                InstrStage<1, [A9_DRegsN],   0, Required>,
1148                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1149                                InstrStage<3, [A9_NPipe], 0>,
1150                                InstrStage<3, [A9_LSUnit]>],
1151                               [2, 1, 1, 1, 1, 1, 2, 2]>,
1152   //
1153   // VST2ln
1154   InstrItinData<IIC_VST2ln,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1155                                InstrStage<1, [A9_MUX0], 0>,
1156                                InstrStage<1, [A9_DRegsN],   0, Required>,
1157                                InstrStage<2, [A9_DRegsVFP], 0, Reserved>,
1158                                InstrStage<2, [A9_NPipe], 0>,
1159                                InstrStage<2, [A9_LSUnit]>],
1160                               [1, 1, 1, 1]>,
1161   //
1162   // VST2lnu
1163   InstrItinData<IIC_VST2lnu,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1164                                InstrStage<1, [A9_MUX0], 0>,
1165                                InstrStage<1, [A9_DRegsN],   0, Required>,
1166                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1167                                InstrStage<3, [A9_NPipe], 0>,
1168                                InstrStage<3, [A9_LSUnit]>],
1169                               [2, 1, 1, 1, 1, 1]>,
1170   //
1171   // VST3
1172   InstrItinData<IIC_VST3,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1173                                InstrStage<1, [A9_MUX0], 0>,
1174                                InstrStage<1, [A9_DRegsN],   0, Required>,
1175                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1176                                InstrStage<3, [A9_NPipe], 0>,
1177                                InstrStage<3, [A9_LSUnit]>],
1178                               [1, 1, 1, 1, 2]>,
1179   //
1180   // VST3u
1181   InstrItinData<IIC_VST3u,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1182                                InstrStage<1, [A9_MUX0], 0>,
1183                                InstrStage<1, [A9_DRegsN],   0, Required>,
1184                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1185                                InstrStage<3, [A9_NPipe], 0>,
1186                                InstrStage<3, [A9_LSUnit]>],
1187                               [2, 1, 1, 1, 1, 1, 2]>,
1188   //
1189   // VST3ln
1190   InstrItinData<IIC_VST3ln,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1191                                InstrStage<1, [A9_MUX0], 0>,
1192                                InstrStage<1, [A9_DRegsN],   0, Required>,
1193                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1194                                InstrStage<3, [A9_NPipe], 0>,
1195                                InstrStage<3, [A9_LSUnit]>],
1196                               [1, 1, 1, 1, 2]>,
1197   //
1198   // VST3lnu
1199   InstrItinData<IIC_VST3lnu,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1200                                InstrStage<1, [A9_MUX0], 0>,
1201                                InstrStage<1, [A9_DRegsN],   0, Required>,
1202                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1203                                InstrStage<3, [A9_NPipe], 0>,
1204                                InstrStage<3, [A9_LSUnit]>],
1205                               [2, 1, 1, 1, 1, 1, 2]>,
1206   //
1207   // VST4
1208   InstrItinData<IIC_VST4,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1209                                InstrStage<1, [A9_MUX0], 0>,
1210                                InstrStage<1, [A9_DRegsN],   0, Required>,
1211                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1212                                InstrStage<3, [A9_NPipe], 0>,
1213                                InstrStage<3, [A9_LSUnit]>],
1214                               [1, 1, 1, 1, 2, 2]>,
1215   //
1216   // VST4u
1217   InstrItinData<IIC_VST4u,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1218                                InstrStage<1, [A9_MUX0], 0>,
1219                                InstrStage<1, [A9_DRegsN],   0, Required>,
1220                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1221                                InstrStage<3, [A9_NPipe], 0>,
1222                                InstrStage<3, [A9_LSUnit]>],
1223                               [2, 1, 1, 1, 1, 1, 2, 2]>,
1224   //
1225   // VST4ln
1226   InstrItinData<IIC_VST4ln,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1227                                InstrStage<1, [A9_MUX0], 0>,
1228                                InstrStage<1, [A9_DRegsN],   0, Required>,
1229                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1230                                InstrStage<3, [A9_NPipe], 0>,
1231                                InstrStage<3, [A9_LSUnit]>],
1232                               [1, 1, 1, 1, 2, 2]>,
1233   //
1234   // VST4lnu
1235   InstrItinData<IIC_VST4lnu,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1236                                InstrStage<1, [A9_MUX0], 0>,
1237                                InstrStage<1, [A9_DRegsN],   0, Required>,
1238                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1239                                InstrStage<3, [A9_NPipe], 0>,
1240                                InstrStage<3, [A9_LSUnit]>],
1241                               [2, 1, 1, 1, 1, 1, 2, 2]>,
1242
1243   //
1244   // Double-register Integer Unary
1245   InstrItinData<IIC_VUNAiD,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1246                                InstrStage<1, [A9_MUX0], 0>,
1247                                InstrStage<1, [A9_DRegsN],   0, Required>,
1248                                // Extra latency cycles since wbck is 6 cycles
1249                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1250                                InstrStage<1, [A9_NPipe]>],
1251                               [4, 2]>,
1252   //
1253   // Quad-register Integer Unary
1254   InstrItinData<IIC_VUNAiQ,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1255                                InstrStage<1, [A9_MUX0], 0>,
1256                                InstrStage<1, [A9_DRegsN],   0, Required>,
1257                                // Extra latency cycles since wbck is 6 cycles
1258                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1259                                InstrStage<1, [A9_NPipe]>],
1260                               [4, 2]>,
1261   //
1262   // Double-register Integer Q-Unary
1263   InstrItinData<IIC_VQUNAiD,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1264                                InstrStage<1, [A9_MUX0], 0>,
1265                                InstrStage<1, [A9_DRegsN],   0, Required>,
1266                                // Extra latency cycles since wbck is 6 cycles
1267                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1268                                InstrStage<1, [A9_NPipe]>],
1269                               [4, 1]>,
1270   //
1271   // Quad-register Integer CountQ-Unary
1272   InstrItinData<IIC_VQUNAiQ,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1273                                InstrStage<1, [A9_MUX0], 0>,
1274                                InstrStage<1, [A9_DRegsN],   0, Required>,
1275                                // Extra latency cycles since wbck is 6 cycles
1276                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1277                                InstrStage<1, [A9_NPipe]>],
1278                               [4, 1]>,
1279   //
1280   // Double-register Integer Binary
1281   InstrItinData<IIC_VBINiD,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1282                                InstrStage<1, [A9_MUX0], 0>,
1283                                InstrStage<1, [A9_DRegsN],   0, Required>,
1284                                // Extra latency cycles since wbck is 6 cycles
1285                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1286                                InstrStage<1, [A9_NPipe]>],
1287                               [3, 2, 2]>,
1288   //
1289   // Quad-register Integer Binary
1290   InstrItinData<IIC_VBINiQ,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1291                                InstrStage<1, [A9_MUX0], 0>,
1292                                InstrStage<1, [A9_DRegsN],   0, Required>,
1293                                // Extra latency cycles since wbck is 6 cycles
1294                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1295                                InstrStage<1, [A9_NPipe]>],
1296                               [3, 2, 2]>,
1297   //
1298   // Double-register Integer Subtract
1299   InstrItinData<IIC_VSUBiD,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1300                                InstrStage<1, [A9_MUX0], 0>,
1301                                InstrStage<1, [A9_DRegsN],   0, Required>,
1302                                // Extra latency cycles since wbck is 6 cycles
1303                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1304                                InstrStage<1, [A9_NPipe]>],
1305                               [3, 2, 1]>,
1306   //
1307   // Quad-register Integer Subtract
1308   InstrItinData<IIC_VSUBiQ,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1309                                InstrStage<1, [A9_MUX0], 0>,
1310                                InstrStage<1, [A9_DRegsN],   0, Required>,
1311                                // Extra latency cycles since wbck is 6 cycles
1312                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1313                                InstrStage<1, [A9_NPipe]>],
1314                               [3, 2, 1]>,
1315   //
1316   // Double-register Integer Shift
1317   InstrItinData<IIC_VSHLiD,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1318                                InstrStage<1, [A9_MUX0], 0>,
1319                                InstrStage<1, [A9_DRegsN],   0, Required>,
1320                                // Extra latency cycles since wbck is 6 cycles
1321                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1322                                InstrStage<1, [A9_NPipe]>],
1323                               [3, 1, 1]>,
1324   //
1325   // Quad-register Integer Shift
1326   InstrItinData<IIC_VSHLiQ,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1327                                InstrStage<1, [A9_MUX0], 0>,
1328                                InstrStage<1, [A9_DRegsN],   0, Required>,
1329                                // Extra latency cycles since wbck is 6 cycles
1330                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1331                                InstrStage<1, [A9_NPipe]>],
1332                               [3, 1, 1]>,
1333   //
1334   // Double-register Integer Shift (4 cycle)
1335   InstrItinData<IIC_VSHLi4D,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1336                                InstrStage<1, [A9_MUX0], 0>,
1337                                InstrStage<1, [A9_DRegsN],   0, Required>,
1338                                // Extra latency cycles since wbck is 6 cycles
1339                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1340                                InstrStage<1, [A9_NPipe]>],
1341                               [4, 1, 1]>,
1342   //
1343   // Quad-register Integer Shift (4 cycle)
1344   InstrItinData<IIC_VSHLi4Q,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1345                                InstrStage<1, [A9_MUX0], 0>,
1346                                InstrStage<1, [A9_DRegsN],   0, Required>,
1347                                // Extra latency cycles since wbck is 6 cycles
1348                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1349                                InstrStage<1, [A9_NPipe]>],
1350                               [4, 1, 1]>,
1351   //
1352   // Double-register Integer Binary (4 cycle)
1353   InstrItinData<IIC_VBINi4D,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1354                                InstrStage<1, [A9_MUX0], 0>,
1355                                InstrStage<1, [A9_DRegsN],   0, Required>,
1356                                // Extra latency cycles since wbck is 6 cycles
1357                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1358                                InstrStage<1, [A9_NPipe]>],
1359                               [4, 2, 2]>,
1360   //
1361   // Quad-register Integer Binary (4 cycle)
1362   InstrItinData<IIC_VBINi4Q,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1363                                InstrStage<1, [A9_MUX0], 0>,
1364                                InstrStage<1, [A9_DRegsN],   0, Required>,
1365                                // Extra latency cycles since wbck is 6 cycles
1366                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1367                                InstrStage<1, [A9_NPipe]>],
1368                               [4, 2, 2]>,
1369   //
1370   // Double-register Integer Subtract (4 cycle)
1371   InstrItinData<IIC_VSUBi4D,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1372                                InstrStage<1, [A9_MUX0], 0>,
1373                                InstrStage<1, [A9_DRegsN],   0, Required>,
1374                                // Extra latency cycles since wbck is 6 cycles
1375                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1376                                InstrStage<1, [A9_NPipe]>],
1377                               [4, 2, 1]>,
1378   //
1379   // Quad-register Integer Subtract (4 cycle)
1380   InstrItinData<IIC_VSUBi4Q,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1381                                InstrStage<1, [A9_MUX0], 0>,
1382                                InstrStage<1, [A9_DRegsN],   0, Required>,
1383                                // Extra latency cycles since wbck is 6 cycles
1384                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1385                                InstrStage<1, [A9_NPipe]>],
1386                               [4, 2, 1]>,
1387
1388   //
1389   // Double-register Integer Count
1390   InstrItinData<IIC_VCNTiD,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1391                                InstrStage<1, [A9_MUX0], 0>,
1392                                InstrStage<1, [A9_DRegsN],   0, Required>,
1393                                // Extra latency cycles since wbck is 6 cycles
1394                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1395                                InstrStage<1, [A9_NPipe]>],
1396                               [3, 2, 2]>,
1397   //
1398   // Quad-register Integer Count
1399   // Result written in N3, but that is relative to the last cycle of multicycle,
1400   // so we use 4 for those cases
1401   InstrItinData<IIC_VCNTiQ,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1402                                InstrStage<1, [A9_MUX0], 0>,
1403                                InstrStage<1, [A9_DRegsN],   0, Required>,
1404                                // Extra latency cycles since wbck is 7 cycles
1405                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1406                                InstrStage<2, [A9_NPipe]>],
1407                               [4, 2, 2]>,
1408   //
1409   // Double-register Absolute Difference and Accumulate
1410   InstrItinData<IIC_VABAD,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1411                                InstrStage<1, [A9_MUX0], 0>,
1412                                InstrStage<1, [A9_DRegsN],   0, Required>,
1413                                // Extra latency cycles since wbck is 6 cycles
1414                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1415                                InstrStage<1, [A9_NPipe]>],
1416                               [6, 3, 2, 1]>,
1417   //
1418   // Quad-register Absolute Difference and Accumulate
1419   InstrItinData<IIC_VABAQ,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1420                                InstrStage<1, [A9_MUX0], 0>,
1421                                InstrStage<1, [A9_DRegsN],   0, Required>,
1422                                // Extra latency cycles since wbck is 6 cycles
1423                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1424                                InstrStage<2, [A9_NPipe]>],
1425                               [6, 3, 2, 1]>,
1426   //
1427   // Double-register Integer Pair Add Long
1428   InstrItinData<IIC_VPALiD,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1429                                InstrStage<1, [A9_MUX0], 0>,
1430                                InstrStage<1, [A9_DRegsN],   0, Required>,
1431                                // Extra latency cycles since wbck is 6 cycles
1432                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1433                                InstrStage<1, [A9_NPipe]>],
1434                               [6, 3, 1]>,
1435   //
1436   // Quad-register Integer Pair Add Long
1437   InstrItinData<IIC_VPALiQ,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1438                                InstrStage<1, [A9_MUX0], 0>,
1439                                InstrStage<1, [A9_DRegsN],   0, Required>,
1440                                // Extra latency cycles since wbck is 6 cycles
1441                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1442                                InstrStage<2, [A9_NPipe]>],
1443                               [6, 3, 1]>,
1444
1445   //
1446   // Double-register Integer Multiply (.8, .16)
1447   InstrItinData<IIC_VMULi16D, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1448                                InstrStage<1, [A9_MUX0], 0>,
1449                                InstrStage<1, [A9_DRegsN],   0, Required>,
1450                                // Extra latency cycles since wbck is 6 cycles
1451                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1452                                InstrStage<1, [A9_NPipe]>],
1453                               [6, 2, 2]>,
1454   //
1455   // Quad-register Integer Multiply (.8, .16)
1456   InstrItinData<IIC_VMULi16Q, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1457                                InstrStage<1, [A9_MUX0], 0>,
1458                                InstrStage<1, [A9_DRegsN],   0, Required>,
1459                                // Extra latency cycles since wbck is 7 cycles
1460                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1461                                InstrStage<2, [A9_NPipe]>],
1462                               [7, 2, 2]>,
1463
1464   //
1465   // Double-register Integer Multiply (.32)
1466   InstrItinData<IIC_VMULi32D, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1467                                InstrStage<1, [A9_MUX0], 0>,
1468                                InstrStage<1, [A9_DRegsN],   0, Required>,
1469                                // Extra latency cycles since wbck is 7 cycles
1470                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1471                                InstrStage<2, [A9_NPipe]>],
1472                               [7, 2, 1]>,
1473   //
1474   // Quad-register Integer Multiply (.32)
1475   InstrItinData<IIC_VMULi32Q, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1476                                InstrStage<1, [A9_MUX0], 0>,
1477                                InstrStage<1, [A9_DRegsN],   0, Required>,
1478                                // Extra latency cycles since wbck is 9 cycles
1479                                InstrStage<10, [A9_DRegsVFP], 0, Reserved>,
1480                                InstrStage<4, [A9_NPipe]>],
1481                               [9, 2, 1]>,
1482   //
1483   // Double-register Integer Multiply-Accumulate (.8, .16)
1484   InstrItinData<IIC_VMACi16D, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1485                                InstrStage<1, [A9_MUX0], 0>,
1486                                InstrStage<1, [A9_DRegsN],   0, Required>,
1487                                // Extra latency cycles since wbck is 6 cycles
1488                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1489                                InstrStage<1, [A9_NPipe]>],
1490                               [6, 3, 2, 2]>,
1491   //
1492   // Double-register Integer Multiply-Accumulate (.32)
1493   InstrItinData<IIC_VMACi32D, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1494                                InstrStage<1, [A9_MUX0], 0>,
1495                                InstrStage<1, [A9_DRegsN],   0, Required>,
1496                                // Extra latency cycles since wbck is 7 cycles
1497                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1498                                InstrStage<2, [A9_NPipe]>],
1499                               [7, 3, 2, 1]>,
1500   //
1501   // Quad-register Integer Multiply-Accumulate (.8, .16)
1502   InstrItinData<IIC_VMACi16Q, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1503                                InstrStage<1, [A9_MUX0], 0>,
1504                                InstrStage<1, [A9_DRegsN],   0, Required>,
1505                                // Extra latency cycles since wbck is 7 cycles
1506                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1507                                InstrStage<2, [A9_NPipe]>],
1508                               [7, 3, 2, 2]>,
1509   //
1510   // Quad-register Integer Multiply-Accumulate (.32)
1511   InstrItinData<IIC_VMACi32Q, [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1512                                InstrStage<1, [A9_MUX0], 0>,
1513                                InstrStage<1, [A9_DRegsN],   0, Required>,
1514                                // Extra latency cycles since wbck is 9 cycles
1515                                InstrStage<10, [A9_DRegsVFP], 0, Reserved>,
1516                                InstrStage<4, [A9_NPipe]>],
1517                               [9, 3, 2, 1]>,
1518
1519   //
1520   // Move
1521   InstrItinData<IIC_VMOV,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1522                                InstrStage<1, [A9_MUX0], 0>,
1523                                InstrStage<1, [A9_DRegsN],   0, Required>,
1524                                InstrStage<1, [A9_DRegsVFP], 0, Reserved>,
1525                                InstrStage<1, [A9_NPipe]>],
1526                               [1,1]>,
1527   //
1528   // Move Immediate
1529   InstrItinData<IIC_VMOVImm,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1530                                InstrStage<1, [A9_MUX0], 0>,
1531                                InstrStage<1, [A9_DRegsN],   0, Required>,
1532                                // Extra latency cycles since wbck is 6 cycles
1533                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1534                                InstrStage<1, [A9_NPipe]>],
1535                               [3]>,
1536   //
1537   // Double-register Permute Move
1538   InstrItinData<IIC_VMOVD,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1539                                InstrStage<1, [A9_MUX0], 0>,
1540                                InstrStage<1, [A9_DRegsN],   0, Required>,
1541                                // Extra latency cycles since wbck is 6 cycles
1542                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1543                                InstrStage<1, [A9_NPipe]>],
1544                               [2, 1]>,
1545   //
1546   // Quad-register Permute Move
1547   InstrItinData<IIC_VMOVQ,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1548                                InstrStage<1, [A9_MUX0], 0>,
1549                                InstrStage<1, [A9_DRegsN],   0, Required>,
1550                                // Extra latency cycles since wbck is 6 cycles
1551                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1552                                InstrStage<1, [A9_NPipe]>],
1553                               [2, 1]>,
1554   //
1555   // Integer to Single-precision Move
1556   InstrItinData<IIC_VMOVIS ,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1557                                InstrStage<1, [A9_MUX0], 0>,
1558                                InstrStage<1, [A9_DRegsN],   0, Required>,
1559                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1560                                InstrStage<1, [A9_NPipe]>],
1561                               [1, 1]>,
1562   //
1563   // Integer to Double-precision Move
1564   InstrItinData<IIC_VMOVID ,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1565                                InstrStage<1, [A9_MUX0], 0>,
1566                                InstrStage<1, [A9_DRegsN],   0, Required>,
1567                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1568                                InstrStage<1, [A9_NPipe]>],
1569                               [1, 1, 1]>,
1570   //
1571   // Single-precision to Integer Move
1572   InstrItinData<IIC_VMOVSI ,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1573                                InstrStage<1, [A9_MUX0], 0>,
1574                                InstrStage<1, [A9_DRegsN],   0, Required>,
1575                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1576                                InstrStage<1, [A9_NPipe]>],
1577                               [2, 1]>,
1578   //
1579   // Double-precision to Integer Move
1580   InstrItinData<IIC_VMOVDI ,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1581                                InstrStage<1, [A9_MUX0], 0>,
1582                                InstrStage<1, [A9_DRegsN],   0, Required>,
1583                                InstrStage<3, [A9_DRegsVFP], 0, Reserved>,
1584                                InstrStage<1, [A9_NPipe]>],
1585                               [2, 2, 1]>,
1586   //
1587   // Integer to Lane Move
1588   InstrItinData<IIC_VMOVISL , [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1589                                InstrStage<1, [A9_MUX0], 0>,
1590                                InstrStage<1, [A9_DRegsN],   0, Required>,
1591                                InstrStage<4, [A9_DRegsVFP], 0, Reserved>,
1592                                InstrStage<2, [A9_NPipe]>],
1593                               [3, 1, 1]>,
1594
1595   //
1596   // Vector narrow move
1597   InstrItinData<IIC_VMOVN,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1598                                InstrStage<1, [A9_MUX0], 0>,
1599                                InstrStage<1, [A9_DRegsN],   0, Required>,
1600                                // Extra latency cycles since wbck is 6 cycles
1601                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1602                                InstrStage<1, [A9_NPipe]>],
1603                               [3, 1]>,
1604   //
1605   // Double-register FP Unary
1606   InstrItinData<IIC_VUNAD,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1607                                InstrStage<1, [A9_MUX0], 0>,
1608                                InstrStage<1, [A9_DRegsN],   0, Required>,
1609                                // Extra latency cycles since wbck is 6 cycles
1610                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1611                                InstrStage<1, [A9_NPipe]>],
1612                               [5, 2]>,
1613   //
1614   // Quad-register FP Unary
1615   // Result written in N5, but that is relative to the last cycle of multicycle,
1616   // so we use 6 for those cases
1617   InstrItinData<IIC_VUNAQ,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1618                                InstrStage<1, [A9_MUX0], 0>,
1619                                InstrStage<1, [A9_DRegsN],   0, Required>,
1620                                // Extra latency cycles since wbck is 7 cycles
1621                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1622                                InstrStage<2, [A9_NPipe]>],
1623                               [6, 2]>,
1624   //
1625   // Double-register FP Binary
1626   // FIXME: We're using this itin for many instructions and [2, 2] here is too
1627   // optimistic.
1628   InstrItinData<IIC_VBIND,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1629                                InstrStage<1, [A9_MUX0], 0>,
1630                                InstrStage<1, [A9_DRegsN],   0, Required>,
1631                                // Extra latency cycles since wbck is 6 cycles
1632                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1633                                InstrStage<1, [A9_NPipe]>],
1634                               [5, 2, 2]>,
1635
1636   //
1637   // VPADD, etc.
1638   InstrItinData<IIC_VPBIND,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1639                                InstrStage<1, [A9_MUX0], 0>,
1640                                InstrStage<1, [A9_DRegsN],   0, Required>,
1641                                // Extra latency cycles since wbck is 6 cycles
1642                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1643                                InstrStage<1, [A9_NPipe]>],
1644                               [5, 1, 1]>,
1645   //
1646   // Double-register FP VMUL
1647   InstrItinData<IIC_VFMULD,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1648                                InstrStage<1, [A9_MUX0], 0>,
1649                                InstrStage<1, [A9_DRegsN],   0, Required>,
1650                                // Extra latency cycles since wbck is 6 cycles
1651                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1652                                InstrStage<1, [A9_NPipe]>],
1653                               [5, 2, 1]>,
1654   //
1655   // Quad-register FP Binary
1656   // Result written in N5, but that is relative to the last cycle of multicycle,
1657   // so we use 6 for those cases
1658   // FIXME: We're using this itin for many instructions and [2, 2] here is too
1659   // optimistic.
1660   InstrItinData<IIC_VBINQ,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1661                                InstrStage<1, [A9_MUX0], 0>,
1662                                InstrStage<1, [A9_DRegsN],   0, Required>,
1663                                // Extra latency cycles since wbck is 7 cycles
1664                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1665                                InstrStage<2, [A9_NPipe]>],
1666                               [6, 2, 2]>,
1667   //
1668   // Quad-register FP VMUL
1669   InstrItinData<IIC_VFMULQ,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1670                                InstrStage<1, [A9_MUX0], 0>,
1671                                InstrStage<1, [A9_DRegsN],   0, Required>,
1672                                // Extra latency cycles since wbck is 7 cycles
1673                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1674                                InstrStage<1, [A9_NPipe]>],
1675                               [6, 2, 1]>,
1676   //
1677   // Double-register FP Multiple-Accumulate
1678   InstrItinData<IIC_VMACD,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1679                                InstrStage<1, [A9_MUX0], 0>,
1680                                InstrStage<1, [A9_DRegsN],   0, Required>,
1681                                // Extra latency cycles since wbck is 7 cycles
1682                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1683                                InstrStage<2, [A9_NPipe]>],
1684                               [6, 3, 2, 1]>,
1685   //
1686   // Quad-register FP Multiple-Accumulate
1687   // Result written in N9, but that is relative to the last cycle of multicycle,
1688   // so we use 10 for those cases
1689   InstrItinData<IIC_VMACQ,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1690                                InstrStage<1, [A9_MUX0], 0>,
1691                                InstrStage<1, [A9_DRegsN],   0, Required>,
1692                                // Extra latency cycles since wbck is 9 cycles
1693                                InstrStage<10, [A9_DRegsVFP], 0, Reserved>,
1694                                InstrStage<4, [A9_NPipe]>],
1695                               [8, 4, 2, 1]>,
1696   //
1697   // Double-register Reciprical Step
1698   InstrItinData<IIC_VRECSD,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1699                                InstrStage<1, [A9_MUX0], 0>,
1700                                InstrStage<1, [A9_DRegsN],   0, Required>,
1701                                // Extra latency cycles since wbck is 10 cycles
1702                                InstrStage<11, [A9_DRegsVFP], 0, Reserved>,
1703                                InstrStage<1, [A9_NPipe]>],
1704                               [9, 2, 2]>,
1705   //
1706   // Quad-register Reciprical Step
1707   InstrItinData<IIC_VRECSQ,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1708                                InstrStage<1, [A9_MUX0], 0>,
1709                                InstrStage<1, [A9_DRegsN],   0, Required>,
1710                                // Extra latency cycles since wbck is 11 cycles
1711                                InstrStage<12, [A9_DRegsVFP], 0, Reserved>,
1712                                InstrStage<2, [A9_NPipe]>],
1713                               [10, 2, 2]>,
1714   //
1715   // Double-register Permute
1716   InstrItinData<IIC_VPERMD,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1717                                InstrStage<1, [A9_MUX0], 0>,
1718                                InstrStage<1, [A9_DRegsN],   0, Required>,
1719                                // Extra latency cycles since wbck is 6 cycles
1720                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1721                                InstrStage<1, [A9_NPipe]>],
1722                               [2, 2, 1, 1]>,
1723   //
1724   // Quad-register Permute
1725   // Result written in N2, but that is relative to the last cycle of multicycle,
1726   // so we use 3 for those cases
1727   InstrItinData<IIC_VPERMQ,   [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1728                                InstrStage<1, [A9_MUX0], 0>,
1729                                InstrStage<1, [A9_DRegsN],   0, Required>,
1730                                // Extra latency cycles since wbck is 7 cycles
1731                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1732                                InstrStage<2, [A9_NPipe]>],
1733                               [3, 3, 1, 1]>,
1734   //
1735   // Quad-register Permute (3 cycle issue)
1736   // Result written in N2, but that is relative to the last cycle of multicycle,
1737   // so we use 4 for those cases
1738   InstrItinData<IIC_VPERMQ3,  [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1739                                InstrStage<1, [A9_MUX0], 0>,
1740                                InstrStage<1, [A9_DRegsN],   0, Required>,
1741                                // Extra latency cycles since wbck is 8 cycles
1742                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
1743                                InstrStage<3, [A9_NPipe]>],
1744                               [4, 4, 1, 1]>,
1745
1746   //
1747   // Double-register VEXT
1748   InstrItinData<IIC_VEXTD,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1749                                InstrStage<1, [A9_MUX0], 0>,
1750                                InstrStage<1, [A9_DRegsN],   0, Required>,
1751                                // Extra latency cycles since wbck is 6 cycles
1752                                InstrStage<7, [A9_DRegsVFP], 0, Reserved>,
1753                                InstrStage<1, [A9_NPipe]>],
1754                               [2, 1, 1]>,
1755   //
1756   // Quad-register VEXT
1757   InstrItinData<IIC_VEXTQ,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1758                                InstrStage<1, [A9_MUX0], 0>,
1759                                InstrStage<1, [A9_DRegsN],   0, Required>,
1760                                // Extra latency cycles since wbck is 7 cycles
1761                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1762                                InstrStage<2, [A9_NPipe]>],
1763                               [3, 1, 2]>,
1764   //
1765   // VTB
1766   InstrItinData<IIC_VTB1,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1767                                InstrStage<1, [A9_MUX0], 0>,
1768                                InstrStage<1, [A9_DRegsN],   0, Required>,
1769                                // Extra latency cycles since wbck is 7 cycles
1770                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1771                                InstrStage<2, [A9_NPipe]>],
1772                               [3, 2, 1]>,
1773   InstrItinData<IIC_VTB2,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1774                                InstrStage<1, [A9_MUX0], 0>,
1775                                InstrStage<2, [A9_DRegsN],   0, Required>,
1776                                // Extra latency cycles since wbck is 7 cycles
1777                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1778                                InstrStage<2, [A9_NPipe]>],
1779                               [3, 2, 2, 1]>,
1780   InstrItinData<IIC_VTB3,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1781                                InstrStage<1, [A9_MUX0], 0>,
1782                                InstrStage<2, [A9_DRegsN],   0, Required>,
1783                                // Extra latency cycles since wbck is 8 cycles
1784                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
1785                                InstrStage<3, [A9_NPipe]>],
1786                               [4, 2, 2, 3, 1]>,
1787   InstrItinData<IIC_VTB4,     [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1788                                InstrStage<1, [A9_MUX0], 0>,
1789                                InstrStage<1, [A9_DRegsN],   0, Required>,
1790                                // Extra latency cycles since wbck is 8 cycles
1791                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
1792                                InstrStage<3, [A9_NPipe]>],
1793                               [4, 2, 2, 3, 3, 1]>,
1794   //
1795   // VTBX
1796   InstrItinData<IIC_VTBX1,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1797                                InstrStage<1, [A9_MUX0], 0>,
1798                                InstrStage<1, [A9_DRegsN],   0, Required>,
1799                                // Extra latency cycles since wbck is 7 cycles
1800                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1801                                InstrStage<2, [A9_NPipe]>],
1802                               [3, 1, 2, 1]>,
1803   InstrItinData<IIC_VTBX2,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1804                                InstrStage<1, [A9_MUX0], 0>,
1805                                InstrStage<1, [A9_DRegsN],   0, Required>,
1806                                // Extra latency cycles since wbck is 7 cycles
1807                                InstrStage<8, [A9_DRegsVFP], 0, Reserved>,
1808                                InstrStage<2, [A9_NPipe]>],
1809                               [3, 1, 2, 2, 1]>,
1810   InstrItinData<IIC_VTBX3,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1811                                InstrStage<1, [A9_MUX0], 0>,
1812                                InstrStage<1, [A9_DRegsN],   0, Required>,
1813                                // Extra latency cycles since wbck is 8 cycles
1814                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
1815                                InstrStage<3, [A9_NPipe]>],
1816                               [4, 1, 2, 2, 3, 1]>,
1817   InstrItinData<IIC_VTBX4,    [InstrStage<1, [A9_Issue0, A9_Issue1], 0>,
1818                                InstrStage<1, [A9_MUX0], 0>,
1819                                InstrStage<1, [A9_DRegsN],   0, Required>,
1820                                // Extra latency cycles since wbck is 8 cycles
1821                                InstrStage<9, [A9_DRegsVFP], 0, Reserved>,
1822                                InstrStage<2, [A9_NPipe]>],
1823                               [4, 1, 2, 2, 3, 3, 1]>
1824 ]>;