llvm/lib/Target/X86/X86SchedHaswell.td

   1 //=- X86SchedHaswell.td - X86 Haswell Scheduling -------------*- tablegen -*-=//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9 // This file defines the machine model for Haswell to support instruction
  10 // scheduling and other instruction cost heuristics.
  11 //
  12 // Note that we define some instructions here that are not supported by haswell,
  13 // but we still have to define them because KNL uses the HSW model.
  14 // They are currently tagged with a comment `Unsupported = 1`.
  15 // FIXME: Use Unsupported = 1 once KNL has its own model.
  16 //
  17 //===----------------------------------------------------------------------===//
  18
  19 def HaswellModel : SchedMachineModel {
  20   // All x86 instructions are modeled as a single micro-op, and HW can decode 4
  21   // instructions per cycle.
  22   let IssueWidth = 4;
  23   let MicroOpBufferSize = 192; // Based on the reorder buffer.
  24   let LoadLatency = 5;
  25   let MispredictPenalty = 16;
  26
  27   // Based on the LSD (loop-stream detector) queue size and benchmarking data.
  28   let LoopMicroOpBufferSize = 50;
  29
  30   // This flag is set to allow the scheduler to assign a default model to
  31   // unrecognized opcodes.
  32   let CompleteModel = 0;
  33 }
  34
  35 let SchedModel = HaswellModel in {
  36
  37 // Haswell can issue micro-ops to 8 different ports in one cycle.
  38
  39 // Ports 0, 1, 5, and 6 handle all computation.
  40 // Port 4 gets the data half of stores. Store data can be available later than
  41 // the store address, but since we don't model the latency of stores, we can
  42 // ignore that.
  43 // Ports 2 and 3 are identical. They handle loads and the address half of
  44 // stores. Port 7 can handle address calculations.
  45 def HWPort0 : ProcResource<1>;
  46 def HWPort1 : ProcResource<1>;
  47 def HWPort2 : ProcResource<1>;
  48 def HWPort3 : ProcResource<1>;
  49 def HWPort4 : ProcResource<1>;
  50 def HWPort5 : ProcResource<1>;
  51 def HWPort6 : ProcResource<1>;
  52 def HWPort7 : ProcResource<1>;
  53
  54 // Many micro-ops are capable of issuing on multiple ports.
  55 def HWPort01  : ProcResGroup<[HWPort0, HWPort1]>;
  56 def HWPort23  : ProcResGroup<[HWPort2, HWPort3]>;
  57 def HWPort237 : ProcResGroup<[HWPort2, HWPort3, HWPort7]>;
  58 def HWPort04  : ProcResGroup<[HWPort0, HWPort4]>;
  59 def HWPort05  : ProcResGroup<[HWPort0, HWPort5]>;
  60 def HWPort06  : ProcResGroup<[HWPort0, HWPort6]>;
  61 def HWPort15  : ProcResGroup<[HWPort1, HWPort5]>;
  62 def HWPort16  : ProcResGroup<[HWPort1, HWPort6]>;
  63 def HWPort56  : ProcResGroup<[HWPort5, HWPort6]>;
  64 def HWPort015 : ProcResGroup<[HWPort0, HWPort1, HWPort5]>;
  65 def HWPort056 : ProcResGroup<[HWPort0, HWPort5, HWPort6]>;
  66 def HWPort0156: ProcResGroup<[HWPort0, HWPort1, HWPort5, HWPort6]>;
  67
  68 // 60 Entry Unified Scheduler
  69 def HWPortAny : ProcResGroup<[HWPort0, HWPort1, HWPort2, HWPort3, HWPort4,
  70                               HWPort5, HWPort6, HWPort7]> {
  71   let BufferSize=60;
  72 }
  73
  74 // Integer division issued on port 0.
  75 def HWDivider : ProcResource<1>;
  76 // FP division and sqrt on port 0.
  77 def HWFPDivider : ProcResource<1>;
  78
  79 // Integer loads are 5 cycles, so ReadAfterLd registers needn't be available until 5
  80 // cycles after the memory operand.
  81 def : ReadAdvance<ReadAfterLd, 5>;
  82
  83 // Vector loads are 5/6/7 cycles, so ReadAfterVec*Ld registers needn't be available
  84 // until 5/6/7 cycles after the memory operand.
  85 def : ReadAdvance<ReadAfterVecLd, 5>;
  86 def : ReadAdvance<ReadAfterVecXLd, 6>;
  87 def : ReadAdvance<ReadAfterVecYLd, 7>;
  88
  89 def : ReadAdvance<ReadInt2Fpu, 0>;
  90
  91 // Many SchedWrites are defined in pairs with and without a folded load.
  92 // Instructions with folded loads are usually micro-fused, so they only appear
  93 // as two micro-ops when queued in the reservation station.
  94 // This multiclass defines the resource usage for variants with and without
  95 // folded loads.
  96 multiclass HWWriteResPair<X86FoldableSchedWrite SchedRW,
  97                           list<ProcResourceKind> ExePorts,
  98                           int Lat, list<int> Res = [1], int UOps = 1,
  99                           int LoadLat = 5> {
 100   // Register variant is using a single cycle on ExePort.
 101   def : WriteRes<SchedRW, ExePorts> {
 102     let Latency = Lat;
 103     let ResourceCycles = Res;
 104     let NumMicroOps = UOps;
 105   }
 106
 107   // Memory variant also uses a cycle on port 2/3 and adds LoadLat cycles to
 108   // the latency (default = 5).
 109   def : WriteRes<SchedRW.Folded, !listconcat([HWPort23], ExePorts)> {
 110     let Latency = !add(Lat, LoadLat);
 111     let ResourceCycles = !listconcat([1], Res);
 112     let NumMicroOps = !add(UOps, 1);
 113   }
 114 }
 115
 116 // A folded store needs a cycle on port 4 for the store data, and an extra port
 117 // 2/3/7 cycle to recompute the address.
 118 def : WriteRes<WriteRMW, [HWPort237,HWPort4]>;
 119
 120 // Store_addr on 237.
 121 // Store_data on 4.
 122 defm : X86WriteRes<WriteStore,   [HWPort237, HWPort4], 1, [1,1], 1>;
 123 defm : X86WriteRes<WriteStoreNT, [HWPort237, HWPort4], 1, [1,1], 2>;
 124 defm : X86WriteRes<WriteLoad,    [HWPort23], 5, [1], 1>;
 125 defm : X86WriteRes<WriteMove,    [HWPort0156], 1, [1], 1>;
 126 def  : WriteRes<WriteZero,       []>;
 127
 128 // Model the effect of clobbering the read-write mask operand of the GATHER operation.
 129 // Does not cost anything by itself, only has latency, matching that of the WriteLoad,
 130 defm : X86WriteRes<WriteVecMaskedGatherWriteback, [], 5, [], 0>;
 131
 132 // Arithmetic.
 133 defm : HWWriteResPair<WriteALU,    [HWPort0156], 1>;
 134 defm : HWWriteResPair<WriteADC,    [HWPort06, HWPort0156], 2, [1,1], 2>;
 135
 136 // Integer multiplication.
 137 defm : HWWriteResPair<WriteIMul8,     [HWPort1],   3>;
 138 defm : HWWriteResPair<WriteIMul16,    [HWPort1,HWPort06,HWPort0156], 4, [1,1,2], 4>;
 139 defm : X86WriteRes<WriteIMul16Imm,    [HWPort1,HWPort0156], 4, [1,1], 2>;
 140 defm : X86WriteRes<WriteIMul16ImmLd,  [HWPort1,HWPort0156,HWPort23], 8, [1,1,1], 3>;
 141 defm : HWWriteResPair<WriteIMul16Reg, [HWPort1],   3>;
 142 defm : HWWriteResPair<WriteIMul32,    [HWPort1,HWPort06,HWPort0156], 4, [1,1,1], 3>;
 143 defm : HWWriteResPair<WriteMULX32,    [HWPort1,HWPort06,HWPort0156], 4, [1,1,1], 3>;
 144 defm : HWWriteResPair<WriteIMul32Imm, [HWPort1],   3>;
 145 defm : HWWriteResPair<WriteIMul32Reg, [HWPort1],   3>;
 146 defm : HWWriteResPair<WriteIMul64,    [HWPort1,HWPort6], 4, [1,1], 2>;
 147 defm : HWWriteResPair<WriteMULX64,    [HWPort1,HWPort6], 4, [1,1], 2>;
 148 defm : HWWriteResPair<WriteIMul64Imm, [HWPort1],   3>;
 149 defm : HWWriteResPair<WriteIMul64Reg, [HWPort1],   3>;
 150 def  : WriteRes<WriteIMulH, []> { let Latency = 3; }
 151
 152 defm : X86WriteRes<WriteBSWAP32,   [HWPort15], 1, [1], 1>;
 153 defm : X86WriteRes<WriteBSWAP64,   [HWPort06, HWPort15], 2, [1,1], 2>;
 154 defm : X86WriteRes<WriteCMPXCHG,[HWPort06, HWPort0156], 5, [2,3], 5>;
 155 defm : X86WriteRes<WriteCMPXCHGRMW,[HWPort23,HWPort06,HWPort0156,HWPort237,HWPort4], 9, [1,2,1,1,1], 6>;
 156 defm : X86WriteRes<WriteXCHG, [HWPort0156], 2, [3], 3>;
 157
 158 // Integer shifts and rotates.
 159 defm : HWWriteResPair<WriteShift,    [HWPort06],  1>;
 160 defm : HWWriteResPair<WriteShiftCL,  [HWPort06, HWPort0156],  3, [2,1], 3>;
 161 defm : HWWriteResPair<WriteRotate,   [HWPort06],  1, [1], 1>;
 162 defm : HWWriteResPair<WriteRotateCL, [HWPort06, HWPort0156],  3, [2,1], 3>;
 163
 164 // SHLD/SHRD.
 165 defm : X86WriteRes<WriteSHDrri, [HWPort1], 3, [1], 1>;
 166 defm : X86WriteRes<WriteSHDrrcl,[HWPort1, HWPort06, HWPort0156], 6, [1, 1, 2], 4>;
 167 defm : X86WriteRes<WriteSHDmri, [HWPort1, HWPort23, HWPort237, HWPort0156], 10, [1, 1, 1, 1], 4>;
 168 defm : X86WriteRes<WriteSHDmrcl,[HWPort1, HWPort23, HWPort237, HWPort06, HWPort0156], 12, [1, 1, 1, 1, 2], 6>;
 169
 170 defm : HWWriteResPair<WriteJump,   [HWPort06],  1>;
 171 defm : HWWriteResPair<WriteCRC32,  [HWPort1],   3>;
 172
 173 defm : HWWriteResPair<WriteCMOV,  [HWPort06,HWPort0156], 2, [1,1], 2>; // Conditional move.
 174 defm : X86WriteRes<WriteFCMOV, [HWPort1], 3, [1], 1>; // x87 conditional move.
 175 def  : WriteRes<WriteSETCC, [HWPort06]>; // Setcc.
 176 def  : WriteRes<WriteSETCCStore, [HWPort06,HWPort4,HWPort237]> {
 177   let Latency = 2;
 178   let NumMicroOps = 3;
 179 }
 180
 181 defm : X86WriteRes<WriteLAHFSAHF,        [HWPort06], 1, [1], 1>;
 182 defm : X86WriteRes<WriteBitTest,         [HWPort06], 1, [1], 1>;
 183 defm : X86WriteRes<WriteBitTestImmLd,    [HWPort06,HWPort23], 6, [1,1], 2>;
 184 defm : X86WriteRes<WriteBitTestRegLd,    [], 1, [], 10>;
 185 defm : X86WriteRes<WriteBitTestSet,      [HWPort06], 1, [1], 1>;
 186 defm : X86WriteRes<WriteBitTestSetImmLd, [HWPort06,HWPort23], 6, [1,1], 3>;
 187 //defm : X86WriteRes<WriteBitTestSetRegLd, [], 1, [], 11>;
 188
 189 // This is for simple LEAs with one or two input operands.
 190 // The complex ones can only execute on port 1, and they require two cycles on
 191 // the port to read all inputs. We don't model that.
 192 def : WriteRes<WriteLEA, [HWPort15]>;
 193
 194 // Bit counts.
 195 defm : HWWriteResPair<WriteBSF, [HWPort1], 3>;
 196 defm : HWWriteResPair<WriteBSR, [HWPort1], 3>;
 197 defm : HWWriteResPair<WriteLZCNT,          [HWPort1], 3>;
 198 defm : HWWriteResPair<WriteTZCNT,          [HWPort1], 3>;
 199 defm : HWWriteResPair<WritePOPCNT,         [HWPort1], 3>;
 200
 201 // BMI1 BEXTR/BLS, BMI2 BZHI
 202 defm : HWWriteResPair<WriteBEXTR, [HWPort06,HWPort15], 2, [1,1], 2>;
 203 defm : HWWriteResPair<WriteBLS,   [HWPort15], 1>;
 204 defm : HWWriteResPair<WriteBZHI,  [HWPort15], 1>;
 205
 206 // TODO: Why isn't the HWDivider used?
 207 defm : X86WriteRes<WriteDiv8,     [HWPort0,HWPort1,HWPort5,HWPort6], 22, [], 9>;
 208 defm : X86WriteRes<WriteDiv16,    [HWPort0,HWPort1,HWPort5,HWPort6,HWPort01,HWPort0156], 98, [7,7,3,3,1,11], 32>;
 209 defm : X86WriteRes<WriteDiv32,    [HWPort0,HWPort1,HWPort5,HWPort6,HWPort01,HWPort0156], 98, [7,7,3,3,1,11], 32>;
 210 defm : X86WriteRes<WriteDiv64,    [HWPort0,HWPort1,HWPort5,HWPort6,HWPort01,HWPort0156], 98, [7,7,3,3,1,11], 32>;
 211 defm : X86WriteRes<WriteDiv8Ld,   [HWPort0,HWPort23,HWDivider], 29, [1,1,10], 2>;
 212 defm : X86WriteRes<WriteDiv16Ld,  [HWPort0,HWPort23,HWDivider], 29, [1,1,10], 2>;
 213 defm : X86WriteRes<WriteDiv32Ld,  [HWPort0,HWPort23,HWDivider], 29, [1,1,10], 2>;
 214 defm : X86WriteRes<WriteDiv64Ld,  [HWPort0,HWPort23,HWDivider], 29, [1,1,10], 2>;
 215
 216 defm : X86WriteRes<WriteIDiv8,    [HWPort0,HWPort1,HWPort5,HWPort6], 23, [], 9>;
 217 defm : X86WriteRes<WriteIDiv16,   [HWPort0,HWPort1,HWPort5,HWPort6,HWPort06,HWPort0156], 112, [4,2,4,8,14,34], 66>;
 218 defm : X86WriteRes<WriteIDiv32,   [HWPort0,HWPort1,HWPort5,HWPort6,HWPort06,HWPort0156], 112, [4,2,4,8,14,34], 66>;
 219 defm : X86WriteRes<WriteIDiv64,   [HWPort0,HWPort1,HWPort5,HWPort6,HWPort06,HWPort0156], 112, [4,2,4,8,14,34], 66>;
 220 defm : X86WriteRes<WriteIDiv8Ld,  [HWPort0,HWPort23,HWDivider], 29, [1,1,10], 2>;
 221 defm : X86WriteRes<WriteIDiv16Ld, [HWPort0,HWPort23,HWDivider], 29, [1,1,10], 2>;
 222 defm : X86WriteRes<WriteIDiv32Ld, [HWPort0,HWPort23,HWDivider], 29, [1,1,10], 2>;
 223 defm : X86WriteRes<WriteIDiv64Ld, [HWPort0,HWPort23,HWDivider], 29, [1,1,10], 2>;
 224
 225 // Scalar and vector floating point.
 226 defm : X86WriteRes<WriteFLD0,          [HWPort01], 1, [1], 1>;
 227 defm : X86WriteRes<WriteFLD1,          [HWPort01], 1, [2], 2>;
 228 defm : X86WriteRes<WriteFLDC,          [HWPort01], 1, [2], 2>;
 229 defm : X86WriteRes<WriteFLoad,         [HWPort23], 5, [1], 1>;
 230 defm : X86WriteRes<WriteFLoadX,        [HWPort23], 6, [1], 1>;
 231 defm : X86WriteRes<WriteFLoadY,        [HWPort23], 7, [1], 1>;
 232 defm : X86WriteRes<WriteFMaskedLoad,   [HWPort23,HWPort5], 8, [1,2], 3>;
 233 defm : X86WriteRes<WriteFMaskedLoadY,  [HWPort23,HWPort5], 9, [1,2], 3>;
 234 defm : X86WriteRes<WriteFStore,        [HWPort237,HWPort4], 1, [1,1], 2>;
 235 defm : X86WriteRes<WriteFStoreX,       [HWPort237,HWPort4], 1, [1,1], 2>;
 236 defm : X86WriteRes<WriteFStoreY,       [HWPort237,HWPort4], 1, [1,1], 2>;
 237 defm : X86WriteRes<WriteFStoreNT,      [HWPort237,HWPort4], 1, [1,1], 2>;
 238 defm : X86WriteRes<WriteFStoreNTX,     [HWPort237,HWPort4], 1, [1,1], 2>;
 239 defm : X86WriteRes<WriteFStoreNTY,     [HWPort237,HWPort4], 1, [1,1], 2>;
 240
 241 defm : X86WriteRes<WriteFMaskedStore32,  [HWPort0,HWPort4,HWPort237,HWPort15], 5, [1,1,1,1], 4>;
 242 defm : X86WriteRes<WriteFMaskedStore32Y, [HWPort0,HWPort4,HWPort237,HWPort15], 5, [1,1,1,1], 4>;
 243 defm : X86WriteRes<WriteFMaskedStore64,  [HWPort0,HWPort4,HWPort237,HWPort15], 5, [1,1,1,1], 4>;
 244 defm : X86WriteRes<WriteFMaskedStore64Y, [HWPort0,HWPort4,HWPort237,HWPort15], 5, [1,1,1,1], 4>;
 245
 246 defm : X86WriteRes<WriteFMove,         [HWPort5], 1, [1], 1>;
 247 defm : X86WriteRes<WriteFMoveX,        [HWPort5], 1, [1], 1>;
 248 defm : X86WriteRes<WriteFMoveY,        [HWPort5], 1, [1], 1>;
 249 defm : X86WriteRes<WriteEMMS,          [HWPort01,HWPort15,HWPort015,HWPort0156], 31, [8,1,21,1], 31>;
 250
 251 defm : HWWriteResPair<WriteFAdd,    [HWPort1],  3, [1], 1, 5>;
 252 defm : HWWriteResPair<WriteFAddX,   [HWPort1],  3, [1], 1, 6>;
 253 defm : HWWriteResPair<WriteFAddY,   [HWPort1],  3, [1], 1, 7>;
 254 defm : HWWriteResPair<WriteFAddZ,   [HWPort1],  3, [1], 1, 7>; // Unsupported = 1
 255 defm : HWWriteResPair<WriteFAdd64,  [HWPort1],  3, [1], 1, 5>;
 256 defm : HWWriteResPair<WriteFAdd64X, [HWPort1],  3, [1], 1, 6>;
 257 defm : HWWriteResPair<WriteFAdd64Y, [HWPort1],  3, [1], 1, 7>;
 258 defm : HWWriteResPair<WriteFAdd64Z, [HWPort1],  3, [1], 1, 7>; // Unsupported = 1
 259
 260 defm : HWWriteResPair<WriteFCmp,    [HWPort1],  3, [1], 1, 5>;
 261 defm : HWWriteResPair<WriteFCmpX,   [HWPort1],  3, [1], 1, 6>;
 262 defm : HWWriteResPair<WriteFCmpY,   [HWPort1],  3, [1], 1, 7>;
 263 defm : HWWriteResPair<WriteFCmpZ,   [HWPort1],  3, [1], 1, 7>; // Unsupported = 1
 264 defm : HWWriteResPair<WriteFCmp64,  [HWPort1],  3, [1], 1, 5>;
 265 defm : HWWriteResPair<WriteFCmp64X, [HWPort1],  3, [1], 1, 6>;
 266 defm : HWWriteResPair<WriteFCmp64Y, [HWPort1],  3, [1], 1, 7>;
 267 defm : HWWriteResPair<WriteFCmp64Z, [HWPort1],  3, [1], 1, 7>; // Unsupported = 1
 268
 269 defm : HWWriteResPair<WriteFCom,    [HWPort1],  3>;
 270 defm : HWWriteResPair<WriteFComX,   [HWPort1],  3>;
 271
 272 defm : HWWriteResPair<WriteFMul,    [HWPort01],  5, [1], 1, 5>;
 273 defm : HWWriteResPair<WriteFMulX,   [HWPort01],  5, [1], 1, 6>;
 274 defm : HWWriteResPair<WriteFMulY,   [HWPort01],  5, [1], 1, 7>;
 275 defm : HWWriteResPair<WriteFMulZ,   [HWPort01],  5, [1], 1, 7>; // Unsupported = 1
 276 defm : HWWriteResPair<WriteFMul64,  [HWPort01],  5, [1], 1, 5>;
 277 defm : HWWriteResPair<WriteFMul64X, [HWPort01],  5, [1], 1, 6>;
 278 defm : HWWriteResPair<WriteFMul64Y, [HWPort01],  5, [1], 1, 7>;
 279 defm : HWWriteResPair<WriteFMul64Z, [HWPort01],  5, [1], 1, 7>; // Unsupported = 1
 280
 281 defm : HWWriteResPair<WriteFDiv,    [HWPort0,HWFPDivider], 13, [1,7], 1, 5>;
 282 defm : HWWriteResPair<WriteFDivX,   [HWPort0,HWFPDivider], 13, [1,7], 1, 6>;
 283 defm : HWWriteResPair<WriteFDivY,   [HWPort0,HWPort15,HWFPDivider], 21, [2,1,14], 3, 7>;
 284 defm : HWWriteResPair<WriteFDivZ,   [HWPort0,HWPort15,HWFPDivider], 21, [2,1,14], 3, 7>; // Unsupported = 1
 285 defm : HWWriteResPair<WriteFDiv64,  [HWPort0,HWFPDivider], 20, [1,14], 1, 5>;
 286 defm : HWWriteResPair<WriteFDiv64X, [HWPort0,HWFPDivider], 20, [1,14], 1, 6>;
 287 defm : HWWriteResPair<WriteFDiv64Y, [HWPort0,HWPort15,HWFPDivider], 35, [2,1,28], 3, 7>;
 288 defm : HWWriteResPair<WriteFDiv64Z, [HWPort0,HWPort15,HWFPDivider], 35, [2,1,28], 3, 7>; // Unsupported = 1
 289
 290 defm : HWWriteResPair<WriteFRcp,   [HWPort0],  5, [1], 1, 5>;
 291 defm : HWWriteResPair<WriteFRcpX,  [HWPort0],  5, [1], 1, 6>;
 292 defm : HWWriteResPair<WriteFRcpY,  [HWPort0,HWPort015], 11, [2,1], 3, 7>;
 293 defm : HWWriteResPair<WriteFRcpZ,  [HWPort0,HWPort015], 11, [2,1], 3, 7>; // Unsupported = 1
 294
 295 defm : HWWriteResPair<WriteFRsqrt, [HWPort0],  5, [1], 1, 5>;
 296 defm : HWWriteResPair<WriteFRsqrtX,[HWPort0],  5, [1], 1, 6>;
 297 defm : HWWriteResPair<WriteFRsqrtY,[HWPort0,HWPort015], 11, [2,1], 3, 7>;
 298 defm : HWWriteResPair<WriteFRsqrtZ,[HWPort0,HWPort015], 11, [2,1], 3, 7>; // Unsupported = 1
 299
 300 defm : HWWriteResPair<WriteFSqrt,    [HWPort0,HWFPDivider], 11, [1,7], 1, 5>;
 301 defm : HWWriteResPair<WriteFSqrtX,   [HWPort0,HWFPDivider], 11, [1,7], 1, 6>;
 302 defm : HWWriteResPair<WriteFSqrtY,   [HWPort0,HWPort15,HWFPDivider], 21, [2,1,14], 3, 7>;
 303 defm : HWWriteResPair<WriteFSqrtZ,   [HWPort0,HWPort15,HWFPDivider], 21, [2,1,14], 3, 7>; // Unsupported = 1
 304 defm : HWWriteResPair<WriteFSqrt64,  [HWPort0,HWFPDivider], 16, [1,14], 1, 5>;
 305 defm : HWWriteResPair<WriteFSqrt64X, [HWPort0,HWFPDivider], 16, [1,14], 1, 6>;
 306 defm : HWWriteResPair<WriteFSqrt64Y, [HWPort0,HWPort15,HWFPDivider], 35, [2,1,28], 3, 7>;
 307 defm : HWWriteResPair<WriteFSqrt64Z, [HWPort0,HWPort15,HWFPDivider], 35, [2,1,28], 3, 7>; // Unsupported = 1
 308 defm : HWWriteResPair<WriteFSqrt80,  [HWPort0,HWFPDivider], 23, [1,17]>;
 309
 310 defm : HWWriteResPair<WriteFMA,   [HWPort01], 5, [1], 1, 5>;
 311 defm : HWWriteResPair<WriteFMAX,  [HWPort01], 5, [1], 1, 6>;
 312 defm : HWWriteResPair<WriteFMAY,  [HWPort01], 5, [1], 1, 7>;
 313 defm : HWWriteResPair<WriteFMAZ,  [HWPort01], 5, [1], 1, 7>; // Unsupported = 1
 314 defm : HWWriteResPair<WriteDPPD,  [HWPort0,HWPort1,HWPort5],  9, [1,1,1], 3, 6>;
 315 defm : HWWriteResPair<WriteDPPS,  [HWPort0,HWPort1,HWPort5], 14, [2,1,1], 4, 6>;
 316 defm : HWWriteResPair<WriteDPPSY, [HWPort0,HWPort1,HWPort5], 14, [2,1,1], 4, 7>;
 317 defm : HWWriteResPair<WriteDPPSZ, [HWPort0,HWPort1,HWPort5], 14, [2,1,1], 4, 7>; // Unsupported = 1
 318 defm : HWWriteResPair<WriteFSign,  [HWPort0], 1>;
 319 defm : X86WriteRes<WriteFRnd,            [HWPort23],  6, [1],   1>;
 320 defm : X86WriteRes<WriteFRndY,           [HWPort23],  6, [1],   1>;
 321 defm : X86WriteRes<WriteFRndZ,           [HWPort23],  6, [1],   1>; // Unsupported = 1
 322 defm : X86WriteRes<WriteFRndLd,  [HWPort1,HWPort23], 12, [2,1], 3>;
 323 defm : X86WriteRes<WriteFRndYLd, [HWPort1,HWPort23], 13, [2,1], 3>;
 324 defm : X86WriteRes<WriteFRndZLd, [HWPort1,HWPort23], 13, [2,1], 3>; // Unsupported = 1
 325 defm : HWWriteResPair<WriteFLogic,  [HWPort5], 1, [1], 1, 6>;
 326 defm : HWWriteResPair<WriteFLogicY, [HWPort5], 1, [1], 1, 7>;
 327 defm : HWWriteResPair<WriteFLogicZ, [HWPort5], 1, [1], 1, 7>; // Unsupported = 1
 328 defm : HWWriteResPair<WriteFTest,   [HWPort0], 1, [1], 1, 6>;
 329 defm : HWWriteResPair<WriteFTestY,  [HWPort0], 1, [1], 1, 7>;
 330 defm : HWWriteResPair<WriteFTestZ,  [HWPort0], 1, [1], 1, 7>; // Unsupported = 1
 331 defm : HWWriteResPair<WriteFShuffle,  [HWPort5], 1, [1], 1, 6>;
 332 defm : HWWriteResPair<WriteFShuffleY, [HWPort5], 1, [1], 1, 7>;
 333 defm : HWWriteResPair<WriteFShuffleZ, [HWPort5], 1, [1], 1, 7>; // Unsupported = 1
 334 defm : HWWriteResPair<WriteFVarShuffle,  [HWPort5], 1, [1], 1, 6>;
 335 defm : HWWriteResPair<WriteFVarShuffleY, [HWPort5], 1, [1], 1, 7>;
 336 defm : HWWriteResPair<WriteFVarShuffleZ, [HWPort5], 1, [1], 1, 7>; // Unsupported = 1
 337 defm : HWWriteResPair<WriteFBlend,  [HWPort015], 1, [1], 1, 6>;
 338 defm : HWWriteResPair<WriteFBlendY, [HWPort015], 1, [1], 1, 7>;
 339 defm : HWWriteResPair<WriteFBlendZ, [HWPort015], 1, [1], 1, 7>; // Unsupported = 1
 340 defm : HWWriteResPair<WriteFShuffle256, [HWPort5], 3, [1], 1, 7>;
 341 defm : HWWriteResPair<WriteFVarShuffle256, [HWPort5], 3, [1], 1, 7>;
 342 defm : HWWriteResPair<WriteFVarBlend,  [HWPort5], 2, [2], 2, 6>;
 343 defm : HWWriteResPair<WriteFVarBlendY, [HWPort5], 2, [2], 2, 7>;
 344 defm : HWWriteResPair<WriteFVarBlendZ, [HWPort5], 2, [2], 2, 7>; // Unsupported = 1
 345
 346 // Conversion between integer and float.
 347 defm : HWWriteResPair<WriteCvtSD2I,   [HWPort1], 3>;
 348 defm : HWWriteResPair<WriteCvtPD2I,   [HWPort1], 3>;
 349 defm : HWWriteResPair<WriteCvtPD2IY,  [HWPort1], 3>;
 350 defm : HWWriteResPair<WriteCvtPD2IZ,  [HWPort1], 3>; // Unsupported = 1
 351 defm : HWWriteResPair<WriteCvtSS2I,   [HWPort1], 3>;
 352 defm : HWWriteResPair<WriteCvtPS2I,   [HWPort1], 3>;
 353 defm : HWWriteResPair<WriteCvtPS2IY,  [HWPort1], 3>;
 354 defm : HWWriteResPair<WriteCvtPS2IZ,  [HWPort1], 3>; // Unsupported = 1
 355
 356 defm : HWWriteResPair<WriteCvtI2SD,   [HWPort1], 4>;
 357 defm : HWWriteResPair<WriteCvtI2PD,   [HWPort1], 4>;
 358 defm : HWWriteResPair<WriteCvtI2PDY,  [HWPort1], 4>;
 359 defm : HWWriteResPair<WriteCvtI2PDZ,  [HWPort1], 4>; // Unsupported = 1
 360 defm : HWWriteResPair<WriteCvtI2SS,   [HWPort1], 4>;
 361 defm : HWWriteResPair<WriteCvtI2PS,   [HWPort1], 4>;
 362 defm : HWWriteResPair<WriteCvtI2PSY,  [HWPort1], 4>;
 363 defm : HWWriteResPair<WriteCvtI2PSZ,  [HWPort1], 4>; // Unsupported = 1
 364
 365 defm : HWWriteResPair<WriteCvtSS2SD,  [HWPort1], 3>;
 366 defm : HWWriteResPair<WriteCvtPS2PD,  [HWPort1], 3>;
 367 defm : HWWriteResPair<WriteCvtPS2PDY, [HWPort1], 3>;
 368 defm : HWWriteResPair<WriteCvtPS2PDZ, [HWPort1], 3>; // Unsupported = 1
 369 defm : HWWriteResPair<WriteCvtSD2SS,  [HWPort1], 3>;
 370 defm : HWWriteResPair<WriteCvtPD2PS,  [HWPort1], 3>;
 371 defm : HWWriteResPair<WriteCvtPD2PSY, [HWPort1], 3>;
 372 defm : HWWriteResPair<WriteCvtPD2PSZ, [HWPort1], 3>; // Unsupported = 1
 373
 374 defm : X86WriteRes<WriteCvtPH2PS,     [HWPort0,HWPort5], 2, [1,1], 2>;
 375 defm : X86WriteRes<WriteCvtPH2PSY,    [HWPort0,HWPort5], 2, [1,1], 2>;
 376 defm : X86WriteRes<WriteCvtPH2PSZ,    [HWPort0,HWPort5], 2, [1,1], 2>; // Unsupported = 1
 377 defm : X86WriteRes<WriteCvtPH2PSLd,  [HWPort0,HWPort23], 6, [1,1], 2>;
 378 defm : X86WriteRes<WriteCvtPH2PSYLd, [HWPort0,HWPort23], 7, [1,1], 2>;
 379 defm : X86WriteRes<WriteCvtPH2PSZLd, [HWPort0,HWPort23], 7, [1,1], 2>; // Unsupported = 1
 380
 381 defm : X86WriteRes<WriteCvtPS2PH,    [HWPort1,HWPort5], 4, [1,1], 2>;
 382 defm : X86WriteRes<WriteCvtPS2PHY,   [HWPort1,HWPort5], 6, [1,1], 2>;
 383 defm : X86WriteRes<WriteCvtPS2PHZ,   [HWPort1,HWPort5], 6, [1,1], 2>; // Unsupported = 1
 384 defm : X86WriteRes<WriteCvtPS2PHSt,  [HWPort1,HWPort4,HWPort5,HWPort237], 5, [1,1,1,1], 4>;
 385 defm : X86WriteRes<WriteCvtPS2PHYSt, [HWPort1,HWPort4,HWPort5,HWPort237], 7, [1,1,1,1], 4>;
 386 defm : X86WriteRes<WriteCvtPS2PHZSt, [HWPort1,HWPort4,HWPort5,HWPort237], 7, [1,1,1,1], 4>; // Unsupported = 1
 387
 388 // Vector integer operations.
 389 defm : X86WriteRes<WriteVecLoad,         [HWPort23], 5, [1], 1>;
 390 defm : X86WriteRes<WriteVecLoadX,        [HWPort23], 6, [1], 1>;
 391 defm : X86WriteRes<WriteVecLoadY,        [HWPort23], 7, [1], 1>;
 392 defm : X86WriteRes<WriteVecLoadNT,       [HWPort23], 6, [1], 1>;
 393 defm : X86WriteRes<WriteVecLoadNTY,      [HWPort23], 7, [1], 1>;
 394 defm : X86WriteRes<WriteVecMaskedLoad,   [HWPort23,HWPort5], 8, [1,2], 3>;
 395 defm : X86WriteRes<WriteVecMaskedLoadY,  [HWPort23,HWPort5], 9, [1,2], 3>;
 396 defm : X86WriteRes<WriteVecStore,        [HWPort237,HWPort4], 1, [1,1], 2>;
 397 defm : X86WriteRes<WriteVecStoreX,       [HWPort237,HWPort4], 1, [1,1], 2>;
 398 defm : X86WriteRes<WriteVecStoreY,       [HWPort237,HWPort4], 1, [1,1], 2>;
 399 defm : X86WriteRes<WriteVecStoreNT,      [HWPort237,HWPort4], 1, [1,1], 2>;
 400 defm : X86WriteRes<WriteVecStoreNTY,     [HWPort237,HWPort4], 1, [1,1], 2>;
 401 defm : X86WriteRes<WriteVecMaskedStore32,  [HWPort0,HWPort4,HWPort237,HWPort15], 5, [1,1,1,1], 4>;
 402 defm : X86WriteRes<WriteVecMaskedStore32Y, [HWPort0,HWPort4,HWPort237,HWPort15], 5, [1,1,1,1], 4>;
 403 defm : X86WriteRes<WriteVecMaskedStore64,  [HWPort0,HWPort4,HWPort237,HWPort15], 5, [1,1,1,1], 4>;
 404 defm : X86WriteRes<WriteVecMaskedStore64Y, [HWPort0,HWPort4,HWPort237,HWPort15], 5, [1,1,1,1], 4>;
 405 defm : X86WriteRes<WriteVecMove,         [HWPort015], 1, [1], 1>;
 406 defm : X86WriteRes<WriteVecMoveX,        [HWPort015], 1, [1], 1>;
 407 defm : X86WriteRes<WriteVecMoveY,        [HWPort015], 1, [1], 1>;
 408 defm : X86WriteRes<WriteVecMoveToGpr,    [HWPort0], 1, [1], 1>;
 409 defm : X86WriteRes<WriteVecMoveFromGpr,  [HWPort5], 1, [1], 1>;
 410
 411 defm : HWWriteResPair<WriteVecLogic, [HWPort015], 1, [1], 1, 5>;
 412 defm : HWWriteResPair<WriteVecLogicX,[HWPort015], 1, [1], 1, 6>;
 413 defm : HWWriteResPair<WriteVecLogicY,[HWPort015], 1, [1], 1, 7>;
 414 defm : HWWriteResPair<WriteVecLogicZ,[HWPort015], 1, [1], 1, 7>; // Unsupported = 1
 415 defm : HWWriteResPair<WriteVecTest,  [HWPort0,HWPort5], 2, [1,1], 2, 6>;
 416 defm : HWWriteResPair<WriteVecTestY, [HWPort0,HWPort5], 4, [1,1], 2, 7>;
 417 defm : HWWriteResPair<WriteVecTestZ, [HWPort0,HWPort5], 4, [1,1], 2, 7>; // Unsupported = 1
 418 defm : HWWriteResPair<WriteVecALU,   [HWPort15],  1, [1], 1, 5>;
 419 defm : HWWriteResPair<WriteVecALUX,  [HWPort15],  1, [1], 1, 6>;
 420 defm : HWWriteResPair<WriteVecALUY,  [HWPort15],  1, [1], 1, 7>;
 421 defm : HWWriteResPair<WriteVecALUZ,  [HWPort15],  1, [1], 1, 7>; // Unsupported = 1
 422 defm : HWWriteResPair<WriteVecIMul,  [HWPort0],  5, [1], 1, 5>;
 423 defm : HWWriteResPair<WriteVecIMulX, [HWPort0],  5, [1], 1, 6>;
 424 defm : HWWriteResPair<WriteVecIMulY, [HWPort0],  5, [1], 1, 7>;
 425 defm : HWWriteResPair<WriteVecIMulZ, [HWPort0],  5, [1], 1, 7>; // Unsupported = 1
 426 defm : HWWriteResPair<WritePMULLD,   [HWPort0], 10, [2], 2, 6>;
 427 defm : HWWriteResPair<WritePMULLDY,  [HWPort0], 10, [2], 2, 7>;
 428 defm : HWWriteResPair<WritePMULLDZ,  [HWPort0], 10, [2], 2, 7>; // Unsupported = 1
 429 defm : HWWriteResPair<WriteShuffle,  [HWPort5],  1, [1], 1, 5>;
 430 defm : HWWriteResPair<WriteShuffleX, [HWPort5],  1, [1], 1, 6>;
 431 defm : HWWriteResPair<WriteShuffleY, [HWPort5],  1, [1], 1, 7>;
 432 defm : HWWriteResPair<WriteShuffleZ, [HWPort5],  1, [1], 1, 7>; // Unsupported = 1
 433 defm : HWWriteResPair<WriteVarShuffle, [HWPort5], 1, [1], 1, 5>;
 434 defm : HWWriteResPair<WriteVarShuffleX,[HWPort5], 1, [1], 1, 6>;
 435 defm : HWWriteResPair<WriteVarShuffleY,[HWPort5], 1, [1], 1, 7>;
 436 defm : HWWriteResPair<WriteVarShuffleZ,[HWPort5], 1, [1], 1, 7>; // Unsupported = 1
 437 defm : HWWriteResPair<WriteBlend,  [HWPort5], 1, [1], 1, 6>;
 438 defm : HWWriteResPair<WriteBlendY, [HWPort5], 1, [1], 1, 7>;
 439 defm : HWWriteResPair<WriteBlendZ, [HWPort5], 1, [1], 1, 7>; // Unsupported = 1
 440 defm : HWWriteResPair<WriteShuffle256, [HWPort5], 3, [1], 1, 7>;
 441 defm : HWWriteResPair<WriteVPMOV256, [HWPort5], 3, [1], 1, 7>;
 442 defm : HWWriteResPair<WriteVarShuffle256, [HWPort5], 3, [1], 1, 7>;
 443 defm : HWWriteResPair<WriteVarBlend,  [HWPort5], 2, [2], 2, 6>;
 444 defm : HWWriteResPair<WriteVarBlendY, [HWPort5], 2, [2], 2, 7>;
 445 defm : HWWriteResPair<WriteVarBlendZ, [HWPort5], 2, [2], 2, 7>; // Unsupported = 1
 446 defm : HWWriteResPair<WriteMPSAD,  [HWPort0, HWPort5], 7, [1, 2], 3, 6>;
 447 defm : HWWriteResPair<WriteMPSADY, [HWPort0, HWPort5], 7, [1, 2], 3, 7>;
 448 defm : HWWriteResPair<WriteMPSADZ, [HWPort0, HWPort5], 7, [1, 2], 3, 7>; // Unsupported = 1
 449 defm : HWWriteResPair<WritePSADBW,  [HWPort0], 5, [1], 1, 5>;
 450 defm : HWWriteResPair<WritePSADBWX, [HWPort0], 5, [1], 1, 6>;
 451 defm : HWWriteResPair<WritePSADBWY, [HWPort0], 5, [1], 1, 7>;
 452 defm : HWWriteResPair<WritePSADBWZ, [HWPort0], 5, [1], 1, 7>; // Unsupported = 1
 453 defm : HWWriteResPair<WritePHMINPOS, [HWPort0],  5, [1], 1, 6>;
 454
 455 // Vector integer shifts.
 456 defm : HWWriteResPair<WriteVecShift,     [HWPort0], 1, [1], 1, 5>;
 457 defm : HWWriteResPair<WriteVecShiftX,    [HWPort0,HWPort5],  2, [1,1], 2, 6>;
 458 defm : X86WriteRes<WriteVecShiftY,       [HWPort0,HWPort5],  4, [1,1], 2>;
 459 defm : X86WriteRes<WriteVecShiftZ,       [HWPort0,HWPort5],  4, [1,1], 2>; // Unsupported = 1
 460 defm : X86WriteRes<WriteVecShiftYLd,     [HWPort0,HWPort23], 8, [1,1], 2>;
 461 defm : X86WriteRes<WriteVecShiftZLd,     [HWPort0,HWPort23], 8, [1,1], 2>; // Unsupported = 1
 462
 463 defm : HWWriteResPair<WriteVecShiftImm,  [HWPort0], 1, [1], 1, 5>;
 464 defm : HWWriteResPair<WriteVecShiftImmX, [HWPort0], 1, [1], 1, 6>;
 465 defm : HWWriteResPair<WriteVecShiftImmY, [HWPort0], 1, [1], 1, 7>;
 466 defm : HWWriteResPair<WriteVecShiftImmZ, [HWPort0], 1, [1], 1, 7>; // Unsupported = 1
 467 defm : HWWriteResPair<WriteVarVecShift,  [HWPort0, HWPort5], 3, [2,1], 3, 6>;
 468 defm : HWWriteResPair<WriteVarVecShiftY, [HWPort0, HWPort5], 3, [2,1], 3, 7>;
 469 defm : HWWriteResPair<WriteVarVecShiftZ, [HWPort0, HWPort5], 3, [2,1], 3, 7>; // Unsupported = 1
 470
 471 // Vector insert/extract operations.
 472 def : WriteRes<WriteVecInsert, [HWPort5]> {
 473   let Latency = 2;
 474   let NumMicroOps = 2;
 475   let ResourceCycles = [2];
 476 }
 477 def : WriteRes<WriteVecInsertLd, [HWPort5,HWPort23]> {
 478   let Latency = 6;
 479   let NumMicroOps = 2;
 480 }
 481 def: InstRW<[WriteVecInsertLd], (instregex "(V?)MOV(H|L)(PD|PS)rm")>;
 482
 483 def : WriteRes<WriteVecExtract, [HWPort0,HWPort5]> {
 484   let Latency = 2;
 485   let NumMicroOps = 2;
 486 }
 487 def : WriteRes<WriteVecExtractSt, [HWPort4,HWPort5,HWPort237]> {
 488   let Latency = 2;
 489   let NumMicroOps = 3;
 490 }
 491
 492 // String instructions.
 493
 494 // Packed Compare Implicit Length Strings, Return Mask
 495 def : WriteRes<WritePCmpIStrM, [HWPort0]> {
 496   let Latency = 11;
 497   let NumMicroOps = 3;
 498   let ResourceCycles = [3];
 499 }
 500 def : WriteRes<WritePCmpIStrMLd, [HWPort0, HWPort23]> {
 501   let Latency = 17;
 502   let NumMicroOps = 4;
 503   let ResourceCycles = [3,1];
 504 }
 505
 506 // Packed Compare Explicit Length Strings, Return Mask
 507 def : WriteRes<WritePCmpEStrM, [HWPort0, HWPort5, HWPort015, HWPort0156]> {
 508   let Latency = 19;
 509   let NumMicroOps = 9;
 510   let ResourceCycles = [4,3,1,1];
 511 }
 512 def : WriteRes<WritePCmpEStrMLd, [HWPort0, HWPort5, HWPort23, HWPort015, HWPort0156]> {
 513   let Latency = 25;
 514   let NumMicroOps = 10;
 515   let ResourceCycles = [4,3,1,1,1];
 516 }
 517
 518 // Packed Compare Implicit Length Strings, Return Index
 519 def : WriteRes<WritePCmpIStrI, [HWPort0]> {
 520   let Latency = 11;
 521   let NumMicroOps = 3;
 522   let ResourceCycles = [3];
 523 }
 524 def : WriteRes<WritePCmpIStrILd, [HWPort0, HWPort23]> {
 525   let Latency = 17;
 526   let NumMicroOps = 4;
 527   let ResourceCycles = [3,1];
 528 }
 529
 530 // Packed Compare Explicit Length Strings, Return Index
 531 def : WriteRes<WritePCmpEStrI, [HWPort0, HWPort5, HWPort0156]> {
 532   let Latency = 18;
 533   let NumMicroOps = 8;
 534   let ResourceCycles = [4,3,1];
 535 }
 536 def : WriteRes<WritePCmpEStrILd, [HWPort0, HWPort5, HWPort23, HWPort0156]> {
 537   let Latency = 24;
 538   let NumMicroOps = 9;
 539   let ResourceCycles = [4,3,1,1];
 540 }
 541
 542 // MOVMSK Instructions.
 543 def : WriteRes<WriteFMOVMSK,    [HWPort0]> { let Latency = 3; }
 544 def : WriteRes<WriteVecMOVMSK,  [HWPort0]> { let Latency = 3; }
 545 def : WriteRes<WriteVecMOVMSKY, [HWPort0]> { let Latency = 3; }
 546 def : WriteRes<WriteMMXMOVMSK,  [HWPort0]> { let Latency = 1; }
 547
 548 // AES Instructions.
 549 def : WriteRes<WriteAESDecEnc, [HWPort5]> {
 550   let Latency = 7;
 551   let NumMicroOps = 1;
 552   let ResourceCycles = [1];
 553 }
 554 def : WriteRes<WriteAESDecEncLd, [HWPort5, HWPort23]> {
 555   let Latency = 13;
 556   let NumMicroOps = 2;
 557   let ResourceCycles = [1,1];
 558 }
 559
 560 def : WriteRes<WriteAESIMC, [HWPort5]> {
 561   let Latency = 14;
 562   let NumMicroOps = 2;
 563   let ResourceCycles = [2];
 564 }
 565 def : WriteRes<WriteAESIMCLd, [HWPort5, HWPort23]> {
 566   let Latency = 20;
 567   let NumMicroOps = 3;
 568   let ResourceCycles = [2,1];
 569 }
 570
 571 def : WriteRes<WriteAESKeyGen, [HWPort0,HWPort5,HWPort015]> {
 572   let Latency = 29;
 573   let NumMicroOps = 11;
 574   let ResourceCycles = [2,7,2];
 575 }
 576 def : WriteRes<WriteAESKeyGenLd, [HWPort0,HWPort5,HWPort23,HWPort015]> {
 577   let Latency = 34;
 578   let NumMicroOps = 11;
 579   let ResourceCycles = [2,7,1,1];
 580 }
 581
 582 // Carry-less multiplication instructions.
 583 def : WriteRes<WriteCLMul, [HWPort0, HWPort5]> {
 584   let Latency = 11;
 585   let NumMicroOps = 3;
 586   let ResourceCycles = [2,1];
 587 }
 588 def : WriteRes<WriteCLMulLd, [HWPort0, HWPort5, HWPort23]> {
 589   let Latency = 17;
 590   let NumMicroOps = 4;
 591   let ResourceCycles = [2,1,1];
 592 }
 593
 594 // Load/store MXCSR.
 595 def : WriteRes<WriteLDMXCSR, [HWPort0,HWPort23,HWPort0156]> { let Latency = 7; let NumMicroOps = 3; let ResourceCycles = [1,1,1]; }
 596 def : WriteRes<WriteSTMXCSR, [HWPort4,HWPort5,HWPort237]> { let Latency = 2; let NumMicroOps = 3; let ResourceCycles = [1,1,1]; }
 597
 598 def : WriteRes<WriteSystem,     [HWPort0156]> { let Latency = 100; }
 599 def : WriteRes<WriteMicrocoded, [HWPort0156]> { let Latency = 100; }
 600 def : WriteRes<WriteFence,  [HWPort23, HWPort4]>;
 601 def : WriteRes<WriteNop, []>;
 602
 603 //================ Exceptions ================//
 604
 605 //-- Specific Scheduling Models --//
 606
 607 // Starting with P0.
 608 def HWWriteP0 : SchedWriteRes<[HWPort0]>;
 609
 610 def HWWriteP01 : SchedWriteRes<[HWPort01]>;
 611
 612 def HWWrite2P01 : SchedWriteRes<[HWPort01]> {
 613   let NumMicroOps = 2;
 614 }
 615 def HWWrite3P01 : SchedWriteRes<[HWPort01]> {
 616   let NumMicroOps = 3;
 617 }
 618
 619 def HWWriteP0156_P23 : SchedWriteRes<[HWPort0156, HWPort23]> {
 620   let NumMicroOps = 2;
 621 }
 622
 623 def HWWrite2P0156_P23 : SchedWriteRes<[HWPort0156, HWPort23]> {
 624   let NumMicroOps = 3;
 625   let ResourceCycles = [2, 1];
 626 }
 627
 628 // Starting with P1.
 629 def HWWriteP1 : SchedWriteRes<[HWPort1]>;
 630
 631
 632 def HWWrite2P1 : SchedWriteRes<[HWPort1]> {
 633   let NumMicroOps = 2;
 634   let ResourceCycles = [2];
 635 }
 636
 637 // Notation:
 638 // - r: register.
 639 // - mm: 64 bit mmx register.
 640 // - x = 128 bit xmm register.
 641 // - (x)mm = mmx or xmm register.
 642 // - y = 256 bit ymm register.
 643 // - v = any vector register.
 644 // - m = memory.
 645
 646 //=== Integer Instructions ===//
 647 //-- Move instructions --//
 648
 649 // XLAT.
 650 def HWWriteXLAT : SchedWriteRes<[]> {
 651   let Latency = 7;
 652   let NumMicroOps = 3;
 653 }
 654 def : InstRW<[HWWriteXLAT], (instrs XLAT)>;
 655
 656 // PUSHA.
 657 def HWWritePushA : SchedWriteRes<[]> {
 658   let NumMicroOps = 19;
 659 }
 660 def : InstRW<[HWWritePushA], (instregex "PUSHA(16|32)")>;
 661
 662 // POPA.
 663 def HWWritePopA : SchedWriteRes<[]> {
 664   let NumMicroOps = 18;
 665 }
 666 def : InstRW<[HWWritePopA], (instregex "POPA(16|32)")>;
 667
 668 //-- Arithmetic instructions --//
 669
 670 // BTR BTS BTC.
 671 // m,r.
 672 def HWWriteBTRSCmr : SchedWriteRes<[]> {
 673   let NumMicroOps = 11;
 674 }
 675 def : SchedAlias<WriteBitTestSetRegRMW, HWWriteBTRSCmr>;
 676
 677 //-- Control transfer instructions --//
 678
 679 // CALL.
 680 // i.
 681 def HWWriteRETI : SchedWriteRes<[HWPort23, HWPort6, HWPort015]> {
 682   let NumMicroOps = 4;
 683   let ResourceCycles = [1, 2, 1];
 684 }
 685 def : InstRW<[HWWriteRETI], (instregex "RETI(L|Q|W)", "LRETI(L|Q|W)")>;
 686
 687 // BOUND.
 688 // r,m.
 689 def HWWriteBOUND : SchedWriteRes<[]> {
 690   let NumMicroOps = 15;
 691 }
 692 def : InstRW<[HWWriteBOUND], (instregex "BOUNDS(16|32)rm")>;
 693
 694 // INTO.
 695 def HWWriteINTO : SchedWriteRes<[]> {
 696   let NumMicroOps = 4;
 697 }
 698 def : InstRW<[HWWriteINTO], (instrs INTO)>;
 699
 700 //-- String instructions --//
 701
 702 // LODSB/W.
 703 def : InstRW<[HWWrite2P0156_P23], (instregex "LODS(B|W)")>;
 704
 705 // LODSD/Q.
 706 def : InstRW<[HWWriteP0156_P23], (instregex "LODS(L|Q)")>;
 707
 708 // MOVS.
 709 def HWWriteMOVS : SchedWriteRes<[HWPort23, HWPort4, HWPort0156]> {
 710   let Latency = 4;
 711   let NumMicroOps = 5;
 712   let ResourceCycles = [2, 1, 2];
 713 }
 714 def : InstRW<[HWWriteMOVS], (instrs MOVSB, MOVSL, MOVSQ, MOVSW)>;
 715
 716 // CMPS.
 717 def HWWriteCMPS : SchedWriteRes<[HWPort23, HWPort0156]> {
 718   let Latency = 4;
 719   let NumMicroOps = 5;
 720   let ResourceCycles = [2, 3];
 721 }
 722 def : InstRW<[HWWriteCMPS], (instregex "CMPS(B|L|Q|W)")>;
 723
 724 //-- Other --//
 725
 726 // RDPMC.f
 727 def HWWriteRDPMC : SchedWriteRes<[]> {
 728   let NumMicroOps = 34;
 729 }
 730 def : InstRW<[HWWriteRDPMC], (instrs RDPMC)>;
 731
 732 // RDRAND.
 733 def HWWriteRDRAND : SchedWriteRes<[HWPort23, HWPort015]> {
 734   let NumMicroOps = 17;
 735   let ResourceCycles = [1, 16];
 736 }
 737 def : InstRW<[HWWriteRDRAND], (instrs RDRAND16r, RDRAND32r, RDRAND64r)>;
 738
 739 //=== Floating Point x87 Instructions ===//
 740 //-- Move instructions --//
 741
 742 // FLD.
 743 // m80.
 744 def : InstRW<[HWWriteP01], (instrs LD_Frr)>;
 745
 746 // FBLD.
 747 // m80.
 748 def HWWriteFBLD : SchedWriteRes<[]> {
 749   let Latency = 47;
 750   let NumMicroOps = 43;
 751 }
 752 def : InstRW<[HWWriteFBLD], (instrs FBLDm)>;
 753
 754 // FST(P).
 755 // r.
 756 def : InstRW<[HWWriteP01], (instregex "ST_(F|FP)rr")>;
 757
 758 // FFREE.
 759 def : InstRW<[HWWriteP01], (instregex "FFREE")>;
 760
 761 // FNSAVE.
 762 def HWWriteFNSAVE : SchedWriteRes<[]> {
 763   let NumMicroOps = 147;
 764 }
 765 def : InstRW<[HWWriteFNSAVE], (instrs FSAVEm)>;
 766
 767 // FRSTOR.
 768 def HWWriteFRSTOR : SchedWriteRes<[]> {
 769   let NumMicroOps = 90;
 770 }
 771 def : InstRW<[HWWriteFRSTOR], (instrs FRSTORm)>;
 772
 773 //-- Arithmetic instructions --//
 774
 775 // FCOMPP FUCOMPP.
 776 // r.
 777 def : InstRW<[HWWrite2P01], (instrs FCOMPP, UCOM_FPPr)>;
 778
 779 // FCOMI(P) FUCOMI(P).
 780 // m.
 781 def : InstRW<[HWWrite3P01], (instrs COM_FIPr, COM_FIr, UCOM_FIPr, UCOM_FIr)>;
 782
 783 // FTST.
 784 def : InstRW<[HWWriteP1], (instregex "TST_F")>;
 785
 786 // FXAM.
 787 def : InstRW<[HWWrite2P1], (instrs XAM_F)>;
 788
 789 // FPREM.
 790 def HWWriteFPREM : SchedWriteRes<[]> {
 791   let Latency = 19;
 792   let NumMicroOps = 28;
 793 }
 794 def : InstRW<[HWWriteFPREM], (instrs FPREM)>;
 795
 796 // FPREM1.
 797 def HWWriteFPREM1 : SchedWriteRes<[]> {
 798   let Latency = 27;
 799   let NumMicroOps = 41;
 800 }
 801 def : InstRW<[HWWriteFPREM1], (instrs FPREM1)>;
 802
 803 // FRNDINT.
 804 def HWWriteFRNDINT : SchedWriteRes<[]> {
 805   let Latency = 11;
 806   let NumMicroOps = 17;
 807 }
 808 def : InstRW<[HWWriteFRNDINT], (instrs FRNDINT)>;
 809
 810 //-- Math instructions --//
 811
 812 // FSCALE.
 813 def HWWriteFSCALE : SchedWriteRes<[]> {
 814   let Latency = 75; // 49-125
 815   let NumMicroOps = 50; // 25-75
 816 }
 817 def : InstRW<[HWWriteFSCALE], (instrs FSCALE)>;
 818
 819 // FXTRACT.
 820 def HWWriteFXTRACT : SchedWriteRes<[]> {
 821   let Latency = 15;
 822   let NumMicroOps = 17;
 823 }
 824 def : InstRW<[HWWriteFXTRACT], (instrs FXTRACT)>;
 825
 826 ////////////////////////////////////////////////////////////////////////////////
 827 // Horizontal add/sub  instructions.
 828 ////////////////////////////////////////////////////////////////////////////////
 829
 830 defm : HWWriteResPair<WriteFHAdd,  [HWPort1, HWPort5], 5, [1,2], 3, 6>;
 831 defm : HWWriteResPair<WriteFHAddY, [HWPort1, HWPort5], 5, [1,2], 3, 7>;
 832 defm : HWWriteResPair<WritePHAdd,  [HWPort5, HWPort15], 3, [2,1], 3, 5>;
 833 defm : HWWriteResPair<WritePHAddX, [HWPort5, HWPort15], 3, [2,1], 3, 6>;
 834 defm : HWWriteResPair<WritePHAddY, [HWPort5, HWPort15], 3, [2,1], 3, 7>;
 835
 836 //=== Floating Point XMM and YMM Instructions ===//
 837
 838 // Remaining instrs.
 839
 840 def HWWriteResGroup0 : SchedWriteRes<[HWPort23]> {
 841   let Latency = 6;
 842   let NumMicroOps = 1;
 843   let ResourceCycles = [1];
 844 }
 845 def: InstRW<[HWWriteResGroup0], (instrs VBROADCASTSSrm)>;
 846 def: InstRW<[HWWriteResGroup0], (instregex "(V?)MOVSHDUPrm",
 847                                            "(V?)MOVSLDUPrm",
 848                                            "VPBROADCAST(D|Q)rm")>;
 849
 850 def HWWriteResGroup0_1 : SchedWriteRes<[HWPort23]> {
 851   let Latency = 7;
 852   let NumMicroOps = 1;
 853   let ResourceCycles = [1];
 854 }
 855 def: InstRW<[HWWriteResGroup0_1], (instrs VBROADCASTF128,
 856                                           VBROADCASTI128,
 857                                           VBROADCASTSDYrm,
 858                                           VBROADCASTSSYrm,
 859                                           VMOVDDUPYrm,
 860                                           VMOVSHDUPYrm,
 861                                           VMOVSLDUPYrm)>;
 862 def: InstRW<[HWWriteResGroup0_1], (instregex "LD_F(32|64|80)m",
 863                                              "VPBROADCAST(D|Q)Yrm")>;
 864
 865 def HWWriteResGroup0_2 : SchedWriteRes<[HWPort23]> {
 866   let Latency = 5;
 867   let NumMicroOps = 1;
 868   let ResourceCycles = [1];
 869 }
 870 def: InstRW<[HWWriteResGroup0_2], (instregex "MOVSX(16|32|64)rm(8|16|32)",
 871                                              "MOVZX(16|32|64)rm(8|16)",
 872                                              "(V?)MOVDDUPrm")>;
 873
 874 def HWWriteResGroup1 : SchedWriteRes<[HWPort4,HWPort237]> {
 875   let Latency = 1;
 876   let NumMicroOps = 2;
 877   let ResourceCycles = [1,1];
 878 }
 879 def: InstRW<[HWWriteResGroup1], (instrs FBSTPm, VMPTRSTm)>;
 880 def: InstRW<[HWWriteResGroup1], (instregex "ST_FP(32|64|80)m")>;
 881
 882 def HWWriteResGroup2 : SchedWriteRes<[HWPort0]> {
 883   let Latency = 1;
 884   let NumMicroOps = 1;
 885   let ResourceCycles = [1];
 886 }
 887 def: InstRW<[HWWriteResGroup2], (instregex "VPSLLVQ(Y?)rr",
 888                                            "VPSRLVQ(Y?)rr")>;
 889
 890 def HWWriteResGroup3 : SchedWriteRes<[HWPort1]> {
 891   let Latency = 1;
 892   let NumMicroOps = 1;
 893   let ResourceCycles = [1];
 894 }
 895 def: InstRW<[HWWriteResGroup3], (instregex "COM(P?)_FST0r",
 896                                            "UCOM_F(P?)r")>;
 897
 898 def HWWriteResGroup4 : SchedWriteRes<[HWPort5]> {
 899   let Latency = 1;
 900   let NumMicroOps = 1;
 901   let ResourceCycles = [1];
 902 }
 903 def: InstRW<[HWWriteResGroup4], (instrs MMX_MOVQ2DQrr)>;
 904
 905 def HWWriteResGroup5 : SchedWriteRes<[HWPort6]> {
 906   let Latency = 1;
 907   let NumMicroOps = 1;
 908   let ResourceCycles = [1];
 909 }
 910 def: InstRW<[HWWriteResGroup5], (instregex "JMP(16|32|64)r")>;
 911
 912 def HWWriteResGroup6 : SchedWriteRes<[HWPort01]> {
 913   let Latency = 1;
 914   let NumMicroOps = 1;
 915   let ResourceCycles = [1];
 916 }
 917 def: InstRW<[HWWriteResGroup6], (instrs FINCSTP, FNOP)>;
 918
 919 def HWWriteResGroup7 : SchedWriteRes<[HWPort06]> {
 920   let Latency = 1;
 921   let NumMicroOps = 1;
 922   let ResourceCycles = [1];
 923 }
 924 def: InstRW<[HWWriteResGroup7], (instrs CDQ, CQO)>;
 925
 926 def HWWriteResGroup8 : SchedWriteRes<[HWPort15]> {
 927   let Latency = 1;
 928   let NumMicroOps = 1;
 929   let ResourceCycles = [1];
 930 }
 931 def: InstRW<[HWWriteResGroup8], (instregex "ANDN(32|64)rr")>;
 932
 933 def HWWriteResGroup9 : SchedWriteRes<[HWPort015]> {
 934   let Latency = 1;
 935   let NumMicroOps = 1;
 936   let ResourceCycles = [1];
 937 }
 938 def: InstRW<[HWWriteResGroup9], (instregex "VPBLENDD(Y?)rri")>;
 939
 940 def HWWriteResGroup10 : SchedWriteRes<[HWPort0156]> {
 941   let Latency = 1;
 942   let NumMicroOps = 1;
 943   let ResourceCycles = [1];
 944 }
 945 def: InstRW<[HWWriteResGroup10], (instrs CBW, CWDE, CDQE,
 946                                          CMC, STC,
 947                                          SGDT64m,
 948                                          SIDT64m,
 949                                          SMSW16m,
 950                                          STRm,
 951                                          SYSCALL)>;
 952
 953 def HWWriteResGroup11 : SchedWriteRes<[HWPort0,HWPort23]> {
 954   let Latency = 6;
 955   let NumMicroOps = 2;
 956   let ResourceCycles = [1,1];
 957 }
 958 def: InstRW<[HWWriteResGroup11], (instregex "(V?)CVTPS2PDrm")>;
 959
 960 def HWWriteResGroup11_1 : SchedWriteRes<[HWPort0,HWPort23]> {
 961   let Latency = 7;
 962   let NumMicroOps = 2;
 963   let ResourceCycles = [1,1];
 964 }
 965 def: InstRW<[HWWriteResGroup11_1], (instrs VPSLLVQrm, VPSRLVQrm)>;
 966 def: InstRW<[HWWriteResGroup11_1], (instregex "(V?)CVTSS2SDrm")>;
 967
 968 def HWWriteResGroup11_2 : SchedWriteRes<[HWPort0,HWPort23]> {
 969   let Latency = 8;
 970   let NumMicroOps = 2;
 971   let ResourceCycles = [1,1];
 972 }
 973 def: InstRW<[HWWriteResGroup11_2], (instrs VPSLLVQYrm, VPSRLVQYrm)>;
 974
 975 def HWWriteResGroup12 : SchedWriteRes<[HWPort1,HWPort23]> {
 976   let Latency = 8;
 977   let NumMicroOps = 2;
 978   let ResourceCycles = [1,1];
 979 }
 980 def: InstRW<[HWWriteResGroup12], (instrs MMX_CVTPI2PSirm)>;
 981 def: InstRW<[HWWriteResGroup12], (instregex "P(DEP|EXT)(32|64)rm")>;
 982
 983 def HWWriteResGroup13 : SchedWriteRes<[HWPort5,HWPort23]> {
 984   let Latency = 6;
 985   let NumMicroOps = 2;
 986   let ResourceCycles = [1,1];
 987 }
 988 def: InstRW<[HWWriteResGroup13], (instregex "(V?)PMOV(SX|ZX)BDrm",
 989                                             "(V?)PMOV(SX|ZX)BQrm",
 990                                             "(V?)PMOV(SX|ZX)BWrm",
 991                                             "(V?)PMOV(SX|ZX)DQrm",
 992                                             "(V?)PMOV(SX|ZX)WDrm",
 993                                             "(V?)PMOV(SX|ZX)WQrm")>;
 994
 995 def HWWriteResGroup13_1 : SchedWriteRes<[HWPort5,HWPort23]> {
 996   let Latency = 8;
 997   let NumMicroOps = 2;
 998   let ResourceCycles = [1,1];
 999 }
1000 def: InstRW<[HWWriteResGroup13_1], (instrs VPMOVSXBDYrm,
1001                                            VPMOVSXBQYrm,
1002                                            VPMOVSXWQYrm)>;
1003
1004 def HWWriteResGroup14 : SchedWriteRes<[HWPort6,HWPort23]> {
1005   let Latency = 6;
1006   let NumMicroOps = 2;
1007   let ResourceCycles = [1,1];
1008 }
1009 def: InstRW<[HWWriteResGroup14], (instrs FARJMP64m)>;
1010 def: InstRW<[HWWriteResGroup14], (instregex "JMP(16|32|64)m")>;
1011
1012 def HWWriteResGroup16 : SchedWriteRes<[HWPort23,HWPort15]> {
1013   let Latency = 6;
1014   let NumMicroOps = 2;
1015   let ResourceCycles = [1,1];
1016 }
1017 def: InstRW<[HWWriteResGroup16], (instregex "ANDN(32|64)rm",
1018                                             "MOVBE(16|32|64)rm")>;
1019
1020 def HWWriteResGroup17 : SchedWriteRes<[HWPort23,HWPort015]> {
1021   let Latency = 7;
1022   let NumMicroOps = 2;
1023   let ResourceCycles = [1,1];
1024 }
1025 def: InstRW<[HWWriteResGroup17], (instrs VINSERTF128rm,
1026                                          VINSERTI128rm,
1027                                          VPBLENDDrmi)>;
1028
1029 def HWWriteResGroup17_2 : SchedWriteRes<[HWPort23,HWPort015]> {
1030   let Latency = 8;
1031   let NumMicroOps = 2;
1032   let ResourceCycles = [1,1];
1033 }
1034 def: InstRW<[HWWriteResGroup17_2], (instrs VPBLENDDYrmi)>;
1035
1036 def HWWriteResGroup18 : SchedWriteRes<[HWPort23,HWPort0156]> {
1037   let Latency = 6;
1038   let NumMicroOps = 2;
1039   let ResourceCycles = [1,1];
1040 }
1041 def: InstRW<[HWWriteResGroup18], (instrs POP16r, POP32r, POP64r)>;
1042 def: InstRW<[HWWriteResGroup18], (instregex "POP(16|32|64)rmr")>;
1043
1044 def HWWriteResGroup19 : SchedWriteRes<[HWPort237,HWPort0156]> {
1045   let Latency = 2;
1046   let NumMicroOps = 2;
1047   let ResourceCycles = [1,1];
1048 }
1049 def: InstRW<[HWWriteResGroup19], (instrs SFENCE)>;
1050
1051 def HWWriteResGroup21 : SchedWriteRes<[HWPort4,HWPort6,HWPort237]> {
1052   let Latency = 2;
1053   let NumMicroOps = 3;
1054   let ResourceCycles = [1,1,1];
1055 }
1056 def: InstRW<[HWWriteResGroup21], (instrs FNSTCW16m)>;
1057
1058 def HWWriteResGroup23 : SchedWriteRes<[HWPort4,HWPort237,HWPort15]> {
1059   let Latency = 2;
1060   let NumMicroOps = 3;
1061   let ResourceCycles = [1,1,1];
1062 }
1063 def: InstRW<[HWWriteResGroup23], (instregex "MOVBE(32|64)mr")>;
1064
1065 def HWWriteResGroup23_16 : SchedWriteRes<[HWPort06, HWPort237, HWPort4]> {
1066   let Latency = 2;
1067   let NumMicroOps = 3;
1068   let ResourceCycles = [1,1,1];
1069 }
1070 def: InstRW<[HWWriteResGroup23_16], (instrs MOVBE16mr)>;
1071
1072 def HWWriteResGroup24 : SchedWriteRes<[HWPort4,HWPort237,HWPort0156]> {
1073   let Latency = 2;
1074   let NumMicroOps = 3;
1075   let ResourceCycles = [1,1,1];
1076 }
1077 def: InstRW<[HWWriteResGroup24], (instrs PUSH16r, PUSH32r, PUSH64r, PUSH64i8,
1078                                          STOSB, STOSL, STOSQ, STOSW)>;
1079 def: InstRW<[HWWriteResGroup24], (instregex "PUSH(16|32|64)rmr")>;
1080
1081 def HWWriteResGroup25 : SchedWriteRes<[HWPort4,HWPort23,HWPort237,HWPort06]> {
1082   let Latency = 7;
1083   let NumMicroOps = 4;
1084   let ResourceCycles = [1,1,1,1];
1085 }
1086 def: InstRW<[HWWriteResGroup25], (instregex "SAR(8|16|32|64)m(1|i)",
1087                                             "SHL(8|16|32|64)m(1|i)",
1088                                             "SHR(8|16|32|64)m(1|i)")>;
1089
1090 def HWWriteResGroup26 : SchedWriteRes<[HWPort4,HWPort23,HWPort237,HWPort0156]> {
1091   let Latency = 7;
1092   let NumMicroOps = 4;
1093   let ResourceCycles = [1,1,1,1];
1094 }
1095 def: InstRW<[HWWriteResGroup26], (instregex "POP(16|32|64)rmm",
1096                                             "PUSH(16|32|64)rmm")>;
1097
1098 def HWWriteResGroup28 : SchedWriteRes<[HWPort01]> {
1099   let Latency = 2;
1100   let NumMicroOps = 2;
1101   let ResourceCycles = [2];
1102 }
1103 def: InstRW<[HWWriteResGroup28], (instrs FDECSTP)>;
1104
1105 def HWWriteResGroup30 : SchedWriteRes<[HWPort0156]> {
1106   let Latency = 2;
1107   let NumMicroOps = 2;
1108   let ResourceCycles = [2];
1109 }
1110 def: InstRW<[HWWriteResGroup30], (instrs LFENCE,
1111                                          MFENCE,
1112                                          WAIT,
1113                                          XGETBV)>;
1114
1115 def HWWriteResGroup31 : SchedWriteRes<[HWPort0,HWPort5]> {
1116   let Latency = 2;
1117   let NumMicroOps = 2;
1118   let ResourceCycles = [1,1];
1119 }
1120 def: InstRW<[HWWriteResGroup31], (instregex "(V?)CVTPS2PDrr",
1121                                             "(V?)CVTSS2SDrr")>;
1122
1123 def HWWriteResGroup32 : SchedWriteRes<[HWPort6,HWPort0156]> {
1124   let Latency = 2;
1125   let NumMicroOps = 2;
1126   let ResourceCycles = [1,1];
1127 }
1128 def: InstRW<[HWWriteResGroup32], (instregex "CLFLUSH")>;
1129
1130 def HWWriteResGroup33 : SchedWriteRes<[HWPort01,HWPort015]> {
1131   let Latency = 2;
1132   let NumMicroOps = 2;
1133   let ResourceCycles = [1,1];
1134 }
1135 def: InstRW<[HWWriteResGroup33], (instrs MMX_MOVDQ2Qrr)>;
1136
1137 def HWWriteResGroup35 : SchedWriteRes<[HWPort06,HWPort0156]> {
1138   let Latency = 2;
1139   let NumMicroOps = 2;
1140   let ResourceCycles = [1,1];
1141 }
1142 def: InstRW<[HWWriteResGroup35], (instrs CWD, JCXZ, JECXZ, JRCXZ)>;
1143
1144 def HWWriteResGroup36_2 : SchedWriteRes<[HWPort5,HWPort23]> {
1145   let Latency = 7;
1146   let NumMicroOps = 3;
1147   let ResourceCycles = [2,1];
1148 }
1149 def: InstRW<[HWWriteResGroup36_2], (instrs MMX_PACKSSDWirm,
1150                                            MMX_PACKSSWBirm,
1151                                            MMX_PACKUSWBirm)>;
1152
1153 def HWWriteResGroup37 : SchedWriteRes<[HWPort23,HWPort0156]> {
1154   let Latency = 7;
1155   let NumMicroOps = 3;
1156   let ResourceCycles = [1,2];
1157 }
1158 def: InstRW<[HWWriteResGroup37], (instrs LEAVE, LEAVE64,
1159                                          SCASB, SCASL, SCASQ, SCASW)>;
1160
1161 def HWWriteResGroup39 : SchedWriteRes<[HWPort0,HWPort01,HWPort23]> {
1162   let Latency = 7;
1163   let NumMicroOps = 3;
1164   let ResourceCycles = [1,1,1];
1165 }
1166 def: InstRW<[HWWriteResGroup39], (instrs FLDCW16m)>;
1167
1168 def HWWriteResGroup41 : SchedWriteRes<[HWPort6,HWPort23,HWPort0156]> {
1169   let Latency = 7;
1170   let NumMicroOps = 3;
1171   let ResourceCycles = [1,1,1];
1172 }
1173 def: InstRW<[HWWriteResGroup41], (instrs LRETQ, RETL, RETQ)>;
1174
1175 def HWWriteResGroup44 : SchedWriteRes<[HWPort4,HWPort6,HWPort237,HWPort0156]> {
1176   let Latency = 3;
1177   let NumMicroOps = 4;
1178   let ResourceCycles = [1,1,1,1];
1179 }
1180 def: InstRW<[HWWriteResGroup44], (instregex "CALL(16|32|64)r")>;
1181
1182 def HWWriteResGroup45 : SchedWriteRes<[HWPort4,HWPort237,HWPort06,HWPort0156]> {
1183   let Latency = 3;
1184   let NumMicroOps = 4;
1185   let ResourceCycles = [1,1,1,1];
1186 }
1187 def: InstRW<[HWWriteResGroup45], (instrs CALL64pcrel32)>;
1188
1189 def HWWriteResGroup46 : SchedWriteRes<[HWPort4,HWPort23,HWPort237,HWPort06]> {
1190   let Latency = 8;
1191   let NumMicroOps = 5;
1192   let ResourceCycles = [1,1,1,2];
1193 }
1194 def: InstRW<[HWWriteResGroup46], (instregex "ROL(8|16|32|64)m(1|i)",
1195                                             "ROR(8|16|32|64)m(1|i)")>;
1196
1197 def HWWriteResGroup46_1 : SchedWriteRes<[HWPort06]> {
1198   let Latency = 2;
1199   let NumMicroOps = 2;
1200   let ResourceCycles = [2];
1201 }
1202 def: InstRW<[HWWriteResGroup46_1], (instrs ROL8r1, ROL16r1, ROL32r1, ROL64r1,
1203                                            ROR8r1, ROR16r1, ROR32r1, ROR64r1)>;
1204
1205 def HWWriteResGroup47 : SchedWriteRes<[HWPort4,HWPort23,HWPort237,HWPort0156]> {
1206   let Latency = 8;
1207   let NumMicroOps = 5;
1208   let ResourceCycles = [1,1,1,2];
1209 }
1210 def: InstRW<[HWWriteResGroup47], (instregex "XADD(8|16|32|64)rm")>;
1211
1212 def HWWriteResGroup48 : SchedWriteRes<[HWPort4,HWPort6,HWPort23,HWPort237,HWPort0156]> {
1213   let Latency = 8;
1214   let NumMicroOps = 5;
1215   let ResourceCycles = [1,1,1,1,1];
1216 }
1217 def: InstRW<[HWWriteResGroup48], (instregex "CALL(16|32|64)m")>;
1218 def: InstRW<[HWWriteResGroup48], (instrs FARCALL64m)>;
1219
1220 def HWWriteResGroup50 : SchedWriteRes<[HWPort1]> {
1221   let Latency = 3;
1222   let NumMicroOps = 1;
1223   let ResourceCycles = [1];
1224 }
1225 def: InstRW<[HWWriteResGroup50], (instrs MMX_CVTPI2PSirr)>;
1226 def: InstRW<[HWWriteResGroup50], (instregex "P(DEP|EXT)(32|64)rr",
1227                                             "(V?)CVTDQ2PS(Y?)rr")>;
1228
1229 def HWWriteResGroup51 : SchedWriteRes<[HWPort5]> {
1230   let Latency = 3;
1231   let NumMicroOps = 1;
1232   let ResourceCycles = [1];
1233 }
1234 def: InstRW<[HWWriteResGroup51], (instregex "VPBROADCAST(B|W)rr")>;
1235
1236 def HWWriteResGroup52 : SchedWriteRes<[HWPort1,HWPort23]> {
1237   let Latency = 9;
1238   let NumMicroOps = 2;
1239   let ResourceCycles = [1,1];
1240 }
1241 def: InstRW<[HWWriteResGroup52], (instregex "(V?)CVTPS2DQrm",
1242                                             "(V?)CVTTPS2DQrm")>;
1243
1244 def HWWriteResGroup52_1 : SchedWriteRes<[HWPort1,HWPort23]> {
1245   let Latency = 10;
1246   let NumMicroOps = 2;
1247   let ResourceCycles = [1,1];
1248 }
1249 def: InstRW<[HWWriteResGroup52_1], (instregex "(ADD|SUB|SUBR)_F(32|64)m",
1250                                               "ILD_F(16|32|64)m")>;
1251 def: InstRW<[HWWriteResGroup52_1], (instrs VCVTDQ2PSYrm,
1252                                            VCVTPS2DQYrm,
1253                                            VCVTTPS2DQYrm)>;
1254
1255 def HWWriteResGroup53_1 : SchedWriteRes<[HWPort5,HWPort23]> {
1256   let Latency = 9;
1257   let NumMicroOps = 2;
1258   let ResourceCycles = [1,1];
1259 }
1260 def: InstRW<[HWWriteResGroup53_1], (instrs VPMOVSXBWYrm,
1261                                            VPMOVSXDQYrm,
1262                                            VPMOVSXWDYrm,
1263                                            VPMOVZXWDYrm)>;
1264
1265 def HWWriteResGroup57 : SchedWriteRes<[HWPort5,HWPort0156]> {
1266   let Latency = 3;
1267   let NumMicroOps = 3;
1268   let ResourceCycles = [2,1];
1269 }
1270 def: InstRW<[HWWriteResGroup57], (instrs MMX_PACKSSDWirr,
1271                                          MMX_PACKSSWBirr,
1272                                          MMX_PACKUSWBirr)>;
1273
1274 def HWWriteResGroup58 : SchedWriteRes<[HWPort6,HWPort0156]> {
1275   let Latency = 3;
1276   let NumMicroOps = 3;
1277   let ResourceCycles = [1,2];
1278 }
1279 def: InstRW<[HWWriteResGroup58], (instregex "CLD")>;
1280
1281 def HWWriteResGroup59 : SchedWriteRes<[HWPort06,HWPort0156]> {
1282   let Latency = 3;
1283   let NumMicroOps = 3;
1284   let ResourceCycles = [1,2];
1285 }
1286 def: InstRW<[HWWriteResGroup59], (instregex "RCL(8|16|32|64)r(1|i)",
1287                                             "RCR(8|16|32|64)r(1|i)")>;
1288
1289 def HWWriteResGroup61 : SchedWriteRes<[HWPort0,HWPort4,HWPort237]> {
1290   let Latency = 4;
1291   let NumMicroOps = 3;
1292   let ResourceCycles = [1,1,1];
1293 }
1294 def: InstRW<[HWWriteResGroup61], (instrs FNSTSWm)>;
1295
1296 def HWWriteResGroup62 : SchedWriteRes<[HWPort1,HWPort4,HWPort237]> {
1297   let Latency = 4;
1298   let NumMicroOps = 3;
1299   let ResourceCycles = [1,1,1];
1300 }
1301 def: InstRW<[HWWriteResGroup62], (instregex "IST(T?)_FP(16|32|64)m",
1302                                             "IST_F(16|32)m")>;
1303
1304 def HWWriteResGroup66 : SchedWriteRes<[HWPort23,HWPort237,HWPort06,HWPort0156]> {
1305   let Latency = 9;
1306   let NumMicroOps = 5;
1307   let ResourceCycles = [1,1,1,2];
1308 }
1309 def: InstRW<[HWWriteResGroup66], (instregex "RCL(8|16|32|64)m(1|i)",
1310                                             "RCR(8|16|32|64)m(1|i)")>;
1311
1312 def HWWriteResGroup68 : SchedWriteRes<[HWPort4,HWPort23,HWPort237,HWPort0156]> {
1313   let Latency = 9;
1314   let NumMicroOps = 6;
1315   let ResourceCycles = [1,1,1,3];
1316 }
1317 def: InstRW<[HWWriteResGroup68], (instregex "XCHG(8|16|32|64)rm")>;
1318
1319 def HWWriteResGroup69 : SchedWriteRes<[HWPort4,HWPort23,HWPort237,HWPort06,HWPort0156]> {
1320   let Latency = 9;
1321   let NumMicroOps = 6;
1322   let ResourceCycles = [1,1,1,2,1];
1323 }
1324 def: InstRW<[HWWriteResGroup69], (instregex "ROL(8|16|32|64)mCL",
1325                                             "ROR(8|16|32|64)mCL",
1326                                             "SAR(8|16|32|64)mCL",
1327                                             "SHL(8|16|32|64)mCL",
1328                                             "SHR(8|16|32|64)mCL")>;
1329 def: SchedAlias<WriteADCRMW, HWWriteResGroup69>;
1330
1331 def HWWriteResGroup70 : SchedWriteRes<[HWPort0,HWPort1]> {
1332   let Latency = 4;
1333   let NumMicroOps = 2;
1334   let ResourceCycles = [1,1];
1335 }
1336 def: InstRW<[HWWriteResGroup70], (instregex "(V?)CVT(T?)SD2SI(64)?rr",
1337                                             "(V?)CVT(T?)SS2SI(64)?rr")>;
1338
1339 def HWWriteResGroup71 : SchedWriteRes<[HWPort0,HWPort5]> {
1340   let Latency = 4;
1341   let NumMicroOps = 2;
1342   let ResourceCycles = [1,1];
1343 }
1344 def: InstRW<[HWWriteResGroup71], (instrs VCVTPS2PDYrr)>;
1345
1346 def HWWriteResGroup72 : SchedWriteRes<[HWPort0,HWPort0156]> {
1347   let Latency = 4;
1348   let NumMicroOps = 2;
1349   let ResourceCycles = [1,1];
1350 }
1351 def: InstRW<[HWWriteResGroup72], (instrs FNSTSW16r)>;
1352
1353 def HWWriteResGroup73 : SchedWriteRes<[HWPort1,HWPort5]> {
1354   let Latency = 4;
1355   let NumMicroOps = 2;
1356   let ResourceCycles = [1,1];
1357 }
1358 def: InstRW<[HWWriteResGroup73], (instrs MMX_CVTPI2PDirr,
1359                                          MMX_CVTPD2PIirr,
1360                                          MMX_CVTPS2PIirr,
1361                                          MMX_CVTTPD2PIirr,
1362                                          MMX_CVTTPS2PIirr)>;
1363 def: InstRW<[HWWriteResGroup73], (instregex "(V?)CVTDQ2PDrr",
1364                                             "(V?)CVTPD2PSrr",
1365                                             "(V?)CVTSD2SSrr",
1366                                             "(V?)CVTSI(64)?2SDrr",
1367                                             "(V?)CVTSI2SSrr",
1368                                             "(V?)CVT(T?)PD2DQrr")>;
1369
1370 def HWWriteResGroup75 : SchedWriteRes<[HWPort1,HWPort23]> {
1371   let Latency = 11;
1372   let NumMicroOps = 3;
1373   let ResourceCycles = [2,1];
1374 }
1375 def: InstRW<[HWWriteResGroup75], (instregex "FICOM(P?)(16|32)m")>;
1376
1377 def HWWriteResGroup76 : SchedWriteRes<[HWPort0,HWPort1,HWPort23]> {
1378   let Latency = 9;
1379   let NumMicroOps = 3;
1380   let ResourceCycles = [1,1,1];
1381 }
1382 def: InstRW<[HWWriteResGroup76], (instregex "(V?)CVTSD2SI(64)?rm",
1383                                             "(V?)CVTSS2SI(64)?rm",
1384                                             "(V?)CVTTSD2SI(64)?rm",
1385                                             "VCVTTSS2SI64rm",
1386                                             "(V?)CVTTSS2SIrm")>;
1387
1388 def HWWriteResGroup77 : SchedWriteRes<[HWPort0,HWPort5,HWPort23]> {
1389   let Latency = 10;
1390   let NumMicroOps = 3;
1391   let ResourceCycles = [1,1,1];
1392 }
1393 def: InstRW<[HWWriteResGroup77], (instrs VCVTPS2PDYrm)>;
1394
1395 def HWWriteResGroup78 : SchedWriteRes<[HWPort1,HWPort5,HWPort23]> {
1396   let Latency = 10;
1397   let NumMicroOps = 3;
1398   let ResourceCycles = [1,1,1];
1399 }
1400 def: InstRW<[HWWriteResGroup78], (instrs CVTPD2PSrm,
1401                                          CVTPD2DQrm,
1402                                          CVTTPD2DQrm,
1403                                          MMX_CVTPD2PIirm,
1404                                          MMX_CVTTPD2PIirm,
1405                                          CVTDQ2PDrm,
1406                                          VCVTDQ2PDrm)>;
1407
1408 def HWWriteResGroup78_1 : SchedWriteRes<[HWPort1,HWPort5,HWPort23]> {
1409   let Latency = 9;
1410   let NumMicroOps = 3;
1411   let ResourceCycles = [1,1,1];
1412 }
1413 def: InstRW<[HWWriteResGroup78_1], (instrs MMX_CVTPI2PDirm,
1414                                            CVTSD2SSrm, CVTSD2SSrm_Int,
1415                                            VCVTSD2SSrm, VCVTSD2SSrm_Int)>;
1416
1417 def HWWriteResGroup80 : SchedWriteRes<[HWPort5,HWPort23,HWPort015]> {
1418   let Latency = 9;
1419   let NumMicroOps = 3;
1420   let ResourceCycles = [1,1,1];
1421 }
1422 def: InstRW<[HWWriteResGroup80], (instregex "VPBROADCAST(B|W)(Y?)rm")>;
1423
1424 def HWWriteResGroup81 : SchedWriteRes<[HWPort0156]> {
1425   let Latency = 4;
1426   let NumMicroOps = 4;
1427   let ResourceCycles = [4];
1428 }
1429 def: InstRW<[HWWriteResGroup81], (instrs FNCLEX)>;
1430
1431 def HWWriteResGroup82 : SchedWriteRes<[]> {
1432   let Latency = 0;
1433   let NumMicroOps = 4;
1434   let ResourceCycles = [];
1435 }
1436 def: InstRW<[HWWriteResGroup82], (instrs VZEROUPPER)>;
1437
1438 def HWWriteResGroup83 : SchedWriteRes<[HWPort1,HWPort6,HWPort0156]> {
1439   let Latency = 4;
1440   let NumMicroOps = 4;
1441   let ResourceCycles = [1,1,2];
1442 }
1443 def: InstRW<[HWWriteResGroup83], (instregex "LAR(16|32|64)rr")>;
1444
1445 def HWWriteResGroup87 : SchedWriteRes<[HWPort1,HWPort6,HWPort23,HWPort0156]> {
1446   let Latency = 9;
1447   let NumMicroOps = 5;
1448   let ResourceCycles = [1,2,1,1];
1449 }
1450 def: InstRW<[HWWriteResGroup87], (instregex "LAR(16|32|64)rm",
1451                                             "LSL(16|32|64)rm")>;
1452
1453 def HWWriteResGroup88 : SchedWriteRes<[HWPort4,HWPort237,HWPort0156]> {
1454   let Latency = 5;
1455   let NumMicroOps = 6;
1456   let ResourceCycles = [1,1,4];
1457 }
1458 def: InstRW<[HWWriteResGroup88], (instregex "PUSHF(16|64)")>;
1459
1460 def HWWriteResGroup89 : SchedWriteRes<[HWPort0]> {
1461   let Latency = 5;
1462   let NumMicroOps = 1;
1463   let ResourceCycles = [1];
1464 }
1465 def: InstRW<[HWWriteResGroup89], (instregex "MUL_(FPrST0|FST0r|FrST0)")>;
1466
1467 def HWWriteResGroup91_2 : SchedWriteRes<[HWPort0,HWPort23]> {
1468   let Latency = 11;
1469   let NumMicroOps = 2;
1470   let ResourceCycles = [1,1];
1471 }
1472 def: InstRW<[HWWriteResGroup91_2], (instregex "(V?)PCMPGTQrm")>;
1473
1474 def HWWriteResGroup91_3 : SchedWriteRes<[HWPort0,HWPort23]> {
1475   let Latency = 12;
1476   let NumMicroOps = 2;
1477   let ResourceCycles = [1,1];
1478 }
1479 def: InstRW<[HWWriteResGroup91_3], (instregex "MUL_F(32|64)m")>;
1480 def: InstRW<[HWWriteResGroup91_3], (instrs VPCMPGTQYrm)>;
1481
1482 def HWWriteResGroup93 : SchedWriteRes<[HWPort1,HWPort5]> {
1483   let Latency = 5;
1484   let NumMicroOps = 3;
1485   let ResourceCycles = [1,2];
1486 }
1487 def: InstRW<[HWWriteResGroup93], (instregex "(V?)CVTSI642SSrr")>;
1488
1489 def HWWriteResGroup94 : SchedWriteRes<[HWPort1,HWPort6,HWPort06]> {
1490   let Latency = 5;
1491   let NumMicroOps = 3;
1492   let ResourceCycles = [1,1,1];
1493 }
1494 def: InstRW<[HWWriteResGroup94], (instregex "STR(16|32|64)r")>;
1495
1496 def HWWriteResGroup97 : SchedWriteRes<[HWPort0,HWPort1,HWPort5,HWPort23]> {
1497   let Latency = 10;
1498   let NumMicroOps = 4;
1499   let ResourceCycles = [1,1,1,1];
1500 }
1501 def: InstRW<[HWWriteResGroup97], (instregex "CVTTSS2SI64rm")>;
1502
1503 def HWWriteResGroup99 : SchedWriteRes<[HWPort6,HWPort0156]> {
1504   let Latency = 5;
1505   let NumMicroOps = 5;
1506   let ResourceCycles = [1,4];
1507 }
1508 def: InstRW<[HWWriteResGroup99], (instrs PAUSE)>;
1509
1510 def HWWriteResGroup100 : SchedWriteRes<[HWPort06,HWPort0156]> {
1511   let Latency = 5;
1512   let NumMicroOps = 5;
1513   let ResourceCycles = [1,4];
1514 }
1515 def: InstRW<[HWWriteResGroup100], (instrs XSETBV)>;
1516
1517 def HWWriteResGroup102 : SchedWriteRes<[HWPort1,HWPort5]> {
1518   let Latency = 6;
1519   let NumMicroOps = 2;
1520   let ResourceCycles = [1,1];
1521 }
1522 def: InstRW<[HWWriteResGroup102], (instrs VCVTDQ2PDYrr,
1523                                           VCVTPD2PSYrr,
1524                                           VCVTPD2DQYrr,
1525                                           VCVTTPD2DQYrr)>;
1526
1527 def HWWriteResGroup103 : SchedWriteRes<[HWPort1,HWPort23]> {
1528   let Latency = 13;
1529   let NumMicroOps = 3;
1530   let ResourceCycles = [2,1];
1531 }
1532 def: InstRW<[HWWriteResGroup103], (instregex "(ADD|SUB|SUBR)_FI(16|32)m")>;
1533
1534 def HWWriteResGroup104 : SchedWriteRes<[HWPort1,HWPort5,HWPort23]> {
1535   let Latency = 12;
1536   let NumMicroOps = 3;
1537   let ResourceCycles = [1,1,1];
1538 }
1539 def: InstRW<[HWWriteResGroup104], (instrs VCVTDQ2PDYrm)>;
1540
1541 def HWWriteResGroup107 : SchedWriteRes<[HWPort1,HWPort6,HWPort06,HWPort0156]> {
1542   let Latency = 6;
1543   let NumMicroOps = 4;
1544   let ResourceCycles = [1,1,1,1];
1545 }
1546 def: InstRW<[HWWriteResGroup107], (instregex "SLDT(16|32|64)r")>;
1547
1548 def HWWriteResGroup108 : SchedWriteRes<[HWPort6,HWPort0156]> {
1549   let Latency = 6;
1550   let NumMicroOps = 6;
1551   let ResourceCycles = [1,5];
1552 }
1553 def: InstRW<[HWWriteResGroup108], (instrs STD)>;
1554
1555 def HWWriteResGroup114 : SchedWriteRes<[HWPort6,HWPort06,HWPort15,HWPort0156]> {
1556   let Latency = 7;
1557   let NumMicroOps = 7;
1558   let ResourceCycles = [2,2,1,2];
1559 }
1560 def: InstRW<[HWWriteResGroup114], (instrs LOOP)>;
1561
1562 def HWWriteResGroup115 : SchedWriteRes<[HWPort0,HWPort1,HWPort23]> {
1563   let Latency = 15;
1564   let NumMicroOps = 3;
1565   let ResourceCycles = [1,1,1];
1566 }
1567 def: InstRW<[HWWriteResGroup115], (instregex "MUL_FI(16|32)m")>;
1568
1569 def HWWriteResGroup120 : SchedWriteRes<[HWPort1,HWPort23,HWPort237,HWPort06,HWPort15,HWPort0156]> {
1570   let Latency = 16;
1571   let NumMicroOps = 10;
1572   let ResourceCycles = [1,1,1,4,1,2];
1573 }
1574 def: InstRW<[HWWriteResGroup120], (instregex "RCL(8|16|32|64)mCL")>;
1575
1576 def HWWriteResGroup129 : SchedWriteRes<[HWPort1,HWPort06,HWPort0156]> {
1577   let Latency = 11;
1578   let NumMicroOps = 7;
1579   let ResourceCycles = [2,2,3];
1580 }
1581 def: InstRW<[HWWriteResGroup129], (instregex "RCL(16|32|64)rCL",
1582                                              "RCR(16|32|64)rCL")>;
1583
1584 def HWWriteResGroup130 : SchedWriteRes<[HWPort1,HWPort06,HWPort15,HWPort0156]> {
1585   let Latency = 11;
1586   let NumMicroOps = 9;
1587   let ResourceCycles = [1,4,1,3];
1588 }
1589 def: InstRW<[HWWriteResGroup130], (instrs RCL8rCL)>;
1590
1591 def HWWriteResGroup131 : SchedWriteRes<[HWPort06,HWPort0156]> {
1592   let Latency = 11;
1593   let NumMicroOps = 11;
1594   let ResourceCycles = [2,9];
1595 }
1596 def: InstRW<[HWWriteResGroup131], (instrs LOOPE, LOOPNE)>;
1597
1598 def HWWriteResGroup132 : SchedWriteRes<[HWPort4,HWPort23,HWPort237,HWPort06,HWPort15,HWPort0156]> {
1599   let Latency = 17;
1600   let NumMicroOps = 14;
1601   let ResourceCycles = [1,1,1,4,2,5];
1602 }
1603 def: InstRW<[HWWriteResGroup132], (instrs CMPXCHG8B)>;
1604
1605 def HWWriteResGroup135 : SchedWriteRes<[HWPort1,HWPort23,HWPort237,HWPort06,HWPort15,HWPort0156]> {
1606   let Latency = 19;
1607   let NumMicroOps = 11;
1608   let ResourceCycles = [2,1,1,3,1,3];
1609 }
1610 def: InstRW<[HWWriteResGroup135], (instregex "RCR(8|16|32|64)mCL")>;
1611
1612 def HWWriteResGroup142 : SchedWriteRes<[HWPort1,HWPort06,HWPort15,HWPort0156]> {
1613   let Latency = 14;
1614   let NumMicroOps = 10;
1615   let ResourceCycles = [2,3,1,4];
1616 }
1617 def: InstRW<[HWWriteResGroup142], (instrs RCR8rCL)>;
1618
1619 def HWWriteResGroup143 : SchedWriteRes<[HWPort23,HWPort0156]> {
1620   let Latency = 19;
1621   let NumMicroOps = 15;
1622   let ResourceCycles = [1,14];
1623 }
1624 def: InstRW<[HWWriteResGroup143], (instrs POPF16)>;
1625
1626 def HWWriteResGroup144 : SchedWriteRes<[HWPort4,HWPort5,HWPort6,HWPort23,HWPort237,HWPort06,HWPort0156]> {
1627   let Latency = 21;
1628   let NumMicroOps = 8;
1629   let ResourceCycles = [1,1,1,1,1,1,2];
1630 }
1631 def: InstRW<[HWWriteResGroup144], (instrs INSB, INSL, INSW)>;
1632
1633 def HWWriteResGroup145 : SchedWriteRes<[HWPort5, HWPort6]> {
1634   let Latency = 8;
1635   let NumMicroOps = 20;
1636   let ResourceCycles = [1,1];
1637 }
1638 def: InstRW<[HWWriteResGroup145], (instrs VZEROALL)>;
1639
1640 def HWWriteResGroup146 : SchedWriteRes<[HWPort0,HWPort4,HWPort5,HWPort23,HWPort237,HWPort06,HWPort0156]> {
1641   let Latency = 22;
1642   let NumMicroOps = 19;
1643   let ResourceCycles = [2,1,4,1,1,4,6];
1644 }
1645 def: InstRW<[HWWriteResGroup146], (instrs CMPXCHG16B)>;
1646
1647 def HWWriteResGroup147 : SchedWriteRes<[HWPort0,HWPort1,HWPort5,HWPort6,HWPort01,HWPort0156]> {
1648   let Latency = 17;
1649   let NumMicroOps = 15;
1650   let ResourceCycles = [2,1,2,4,2,4];
1651 }
1652 def: InstRW<[HWWriteResGroup147], (instrs XCH_F)>;
1653
1654 def HWWriteResGroup149 : SchedWriteRes<[HWPort5,HWPort6,HWPort06,HWPort0156]> {
1655   let Latency = 18;
1656   let NumMicroOps = 8;
1657   let ResourceCycles = [1,1,1,5];
1658 }
1659 def: InstRW<[HWWriteResGroup149], (instrs CPUID, RDTSC)>;
1660
1661 def HWWriteResGroup151 : SchedWriteRes<[HWPort6,HWPort23,HWPort0156]> {
1662   let Latency = 23;
1663   let NumMicroOps = 19;
1664   let ResourceCycles = [3,1,15];
1665 }
1666 def: InstRW<[HWWriteResGroup151], (instregex "XRSTOR(64)?")>;
1667
1668 def HWWriteResGroup154 : SchedWriteRes<[HWPort0]> {
1669   let Latency = 20;
1670   let NumMicroOps = 1;
1671   let ResourceCycles = [1];
1672 }
1673 def: InstRW<[HWWriteResGroup154], (instregex "DIV_(FPrST0|FST0r|FrST0)")>;
1674
1675 def HWWriteResGroup155 : SchedWriteRes<[HWPort0,HWPort23]> {
1676   let Latency = 27;
1677   let NumMicroOps = 2;
1678   let ResourceCycles = [1,1];
1679 }
1680 def: InstRW<[HWWriteResGroup155], (instregex "DIVR_F(32|64)m")>;
1681
1682 def HWWriteResGroup156 : SchedWriteRes<[HWPort5,HWPort6,HWPort0156]> {
1683   let Latency = 20;
1684   let NumMicroOps = 10;
1685   let ResourceCycles = [1,2,7];
1686 }
1687 def: InstRW<[HWWriteResGroup156], (instrs MWAITrr)>;
1688
1689 def HWWriteResGroup161 : SchedWriteRes<[HWPort0,HWPort1,HWPort23]> {
1690   let Latency = 30;
1691   let NumMicroOps = 3;
1692   let ResourceCycles = [1,1,1];
1693 }
1694 def: InstRW<[HWWriteResGroup161], (instregex "DIVR_FI(16|32)m")>;
1695
1696 def HWWriteResGroup162 : SchedWriteRes<[HWPort0]> {
1697   let Latency = 24;
1698   let NumMicroOps = 1;
1699   let ResourceCycles = [1];
1700 }
1701 def: InstRW<[HWWriteResGroup162], (instregex "DIVR_(FPrST0|FST0r|FrST0)")>;
1702
1703 def HWWriteResGroup163 : SchedWriteRes<[HWPort0,HWPort23]> {
1704   let Latency = 31;
1705   let NumMicroOps = 2;
1706   let ResourceCycles = [1,1];
1707 }
1708 def: InstRW<[HWWriteResGroup163], (instregex "DIV_F(32|64)m")>;
1709
1710 def HWWriteResGroup164 : SchedWriteRes<[HWPort4,HWPort6,HWPort23,HWPort237,HWPort0156]> {
1711   let Latency = 30;
1712   let NumMicroOps = 27;
1713   let ResourceCycles = [1,5,1,1,19];
1714 }
1715 def: InstRW<[HWWriteResGroup164], (instrs XSAVE64)>;
1716
1717 def HWWriteResGroup165 : SchedWriteRes<[HWPort4,HWPort6,HWPort23,HWPort237,HWPort0156]> {
1718   let Latency = 31;
1719   let NumMicroOps = 28;
1720   let ResourceCycles = [1,6,1,1,19];
1721 }
1722 def: InstRW<[HWWriteResGroup165], (instrs XSAVE)>;
1723 def: InstRW<[HWWriteResGroup165], (instregex "XSAVEC", "XSAVES", "XSAVEOPT")>;
1724
1725 def HWWriteResGroup166 : SchedWriteRes<[HWPort0,HWPort1,HWPort23]> {
1726   let Latency = 34;
1727   let NumMicroOps = 3;
1728   let ResourceCycles = [1,1,1];
1729 }
1730 def: InstRW<[HWWriteResGroup166], (instregex "DIV_FI(16|32)m")>;
1731
1732 def HWWriteResGroup170 : SchedWriteRes<[HWPort5,HWPort6,HWPort23,HWPort06,HWPort0156]> {
1733   let Latency = 35;
1734   let NumMicroOps = 23;
1735   let ResourceCycles = [1,5,3,4,10];
1736 }
1737 def: InstRW<[HWWriteResGroup170], (instregex "IN(8|16|32)ri",
1738                                              "IN(8|16|32)rr")>;
1739
1740 def HWWriteResGroup171 : SchedWriteRes<[HWPort5,HWPort6,HWPort23,HWPort237,HWPort06,HWPort0156]> {
1741   let Latency = 36;
1742   let NumMicroOps = 23;
1743   let ResourceCycles = [1,5,2,1,4,10];
1744 }
1745 def: InstRW<[HWWriteResGroup171], (instregex "OUT(8|16|32)ir",
1746                                              "OUT(8|16|32)rr")>;
1747
1748 def HWWriteResGroup175 : SchedWriteRes<[HWPort1,HWPort4,HWPort5,HWPort6,HWPort23,HWPort237,HWPort15,HWPort0156]> {
1749   let Latency = 41;
1750   let NumMicroOps = 18;
1751   let ResourceCycles = [1,1,2,3,1,1,1,8];
1752 }
1753 def: InstRW<[HWWriteResGroup175], (instrs VMCLEARm)>;
1754
1755 def HWWriteResGroup176 : SchedWriteRes<[HWPort5,HWPort0156]> {
1756   let Latency = 42;
1757   let NumMicroOps = 22;
1758   let ResourceCycles = [2,20];
1759 }
1760 def: InstRW<[HWWriteResGroup176], (instrs RDTSCP)>;
1761
1762 def HWWriteResGroup177 : SchedWriteRes<[HWPort0,HWPort01,HWPort23,HWPort05,HWPort06,HWPort015,HWPort0156]> {
1763   let Latency = 61;
1764   let NumMicroOps = 64;
1765   let ResourceCycles = [2,2,8,1,10,2,39];
1766 }
1767 def: InstRW<[HWWriteResGroup177], (instrs FLDENVm)>;
1768
1769 def HWWriteResGroup178 : SchedWriteRes<[HWPort0,HWPort6,HWPort23,HWPort05,HWPort06,HWPort15,HWPort0156]> {
1770   let Latency = 64;
1771   let NumMicroOps = 88;
1772   let ResourceCycles = [4,4,31,1,2,1,45];
1773 }
1774 def: InstRW<[HWWriteResGroup178], (instrs FXRSTOR64)>;
1775
1776 def HWWriteResGroup179 : SchedWriteRes<[HWPort0,HWPort6,HWPort23,HWPort05,HWPort06,HWPort15,HWPort0156]> {
1777   let Latency = 64;
1778   let NumMicroOps = 90;
1779   let ResourceCycles = [4,2,33,1,2,1,47];
1780 }
1781 def: InstRW<[HWWriteResGroup179], (instrs FXRSTOR)>;
1782
1783 def HWWriteResGroup180 : SchedWriteRes<[HWPort5,HWPort01,HWPort0156]> {
1784   let Latency = 75;
1785   let NumMicroOps = 15;
1786   let ResourceCycles = [6,3,6];
1787 }
1788 def: InstRW<[HWWriteResGroup180], (instrs FNINIT)>;
1789
1790 def HWWriteResGroup183 : SchedWriteRes<[HWPort0,HWPort1,HWPort4,HWPort5,HWPort6,HWPort237,HWPort06,HWPort0156]> {
1791   let Latency = 115;
1792   let NumMicroOps = 100;
1793   let ResourceCycles = [9,9,11,8,1,11,21,30];
1794 }
1795 def: InstRW<[HWWriteResGroup183], (instrs FSTENVm)>;
1796
1797 def HWWriteResGroup184 : SchedWriteRes<[HWPort0,HWPort5,HWPort06,HWPort15,HWPort015,HWPort23]> {
1798   let Latency = 14;
1799   let NumMicroOps = 12;
1800   let ResourceCycles = [2,2,2,1,3,2];
1801 }
1802 def: InstRW<[HWWriteResGroup184], (instrs VGATHERDPDrm, VPGATHERDQrm)>;
1803
1804 def HWWriteResGroup185 : SchedWriteRes<[HWPort0,HWPort5,HWPort06,HWPort15,HWPort015,HWPort23]> {
1805   let Latency = 17;
1806   let NumMicroOps = 20;
1807   let ResourceCycles = [3,3,4,1,5,4];
1808 }
1809 def: InstRW<[HWWriteResGroup185], (instrs VGATHERDPDYrm, VPGATHERDQYrm)>;
1810
1811 def HWWriteResGroup186 : SchedWriteRes<[HWPort0,HWPort5,HWPort06,HWPort15,HWPort015,HWPort23]> {
1812   let Latency = 16;
1813   let NumMicroOps = 20;
1814   let ResourceCycles = [3,3,4,1,5,4];
1815 }
1816 def: InstRW<[HWWriteResGroup186], (instrs VGATHERDPSrm, VPGATHERDDrm)>;
1817
1818 def HWWriteResGroup187 : SchedWriteRes<[HWPort0,HWPort5,HWPort06,HWPort15,HWPort015,HWPort23]> {
1819   let Latency = 22;
1820   let NumMicroOps = 34;
1821   let ResourceCycles = [5,3,8,1,9,8];
1822 }
1823 def: InstRW<[HWWriteResGroup187], (instrs VGATHERDPSYrm, VPGATHERDDYrm)>;
1824
1825 def HWWriteResGroup188 : SchedWriteRes<[HWPort0,HWPort5,HWPort06,HWPort15,HWPort015,HWPort23]> {
1826   let Latency = 15;
1827   let NumMicroOps = 14;
1828   let ResourceCycles = [3,3,2,1,3,2];
1829 }
1830 def: InstRW<[HWWriteResGroup188], (instrs VGATHERQPDrm, VPGATHERQQrm)>;
1831
1832 def HWWriteResGroup189 : SchedWriteRes<[HWPort0,HWPort5,HWPort06,HWPort15,HWPort015,HWPort23]> {
1833   let Latency = 17;
1834   let NumMicroOps = 22;
1835   let ResourceCycles = [5,3,4,1,5,4];
1836 }
1837 def: InstRW<[HWWriteResGroup189], (instrs VGATHERQPDYrm, VPGATHERQQYrm,
1838                                           VGATHERQPSYrm, VPGATHERQDYrm)>;
1839
1840 def HWWriteResGroup190 : SchedWriteRes<[HWPort0,HWPort5,HWPort06,HWPort15,HWPort015,HWPort23]> {
1841   let Latency = 16;
1842   let NumMicroOps = 15;
1843   let ResourceCycles = [3,3,2,1,4,2];
1844 }
1845 def: InstRW<[HWWriteResGroup190], (instrs VGATHERQPSrm, VPGATHERQDrm)>;
1846
1847 def: InstRW<[WriteZero], (instrs CLC)>;
1848
1849
1850 // Instruction variants handled by the renamer. These might not need execution
1851 // ports in certain conditions.
1852 // See Agner's Fog "The microarchitecture of Intel, AMD and VIA CPUs",
1853 // section "Haswell and Broadwell Pipeline" > "Register allocation and
1854 // renaming".
1855 // These can be investigated with llvm-exegesis, e.g.
1856 // echo 'pxor %mm0, %mm0' | /tmp/llvm-exegesis -mode=uops -snippets-file=-
1857 // echo 'vxorpd %xmm0, %xmm0, %xmm1' | /tmp/llvm-exegesis -mode=uops -snippets-file=-
1858
1859 def HWWriteZeroLatency : SchedWriteRes<[]> {
1860   let Latency = 0;
1861 }
1862
1863 def HWWriteZeroIdiom : SchedWriteVariant<[
1864     SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [HWWriteZeroLatency]>,
1865     SchedVar<NoSchedPred,                          [WriteALU]>
1866 ]>;
1867 def : InstRW<[HWWriteZeroIdiom], (instrs SUB32rr, SUB64rr,
1868                                          XOR32rr, XOR64rr)>;
1869
1870 def HWWriteFZeroIdiom : SchedWriteVariant<[
1871     SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [HWWriteZeroLatency]>,
1872     SchedVar<NoSchedPred,                          [WriteFLogic]>
1873 ]>;
1874 def : InstRW<[HWWriteFZeroIdiom], (instrs XORPSrr, VXORPSrr, XORPDrr,
1875                                           VXORPDrr)>;
1876
1877 def HWWriteFZeroIdiomY : SchedWriteVariant<[
1878     SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [HWWriteZeroLatency]>,
1879     SchedVar<NoSchedPred,                          [WriteFLogicY]>
1880 ]>;
1881 def : InstRW<[HWWriteFZeroIdiomY], (instrs VXORPSYrr, VXORPDYrr)>;
1882
1883 def HWWriteVZeroIdiomLogicX : SchedWriteVariant<[
1884     SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [HWWriteZeroLatency]>,
1885     SchedVar<NoSchedPred,                          [WriteVecLogicX]>
1886 ]>;
1887 def : InstRW<[HWWriteVZeroIdiomLogicX], (instrs PXORrr, VPXORrr)>;
1888
1889 def HWWriteVZeroIdiomLogicY : SchedWriteVariant<[
1890     SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [HWWriteZeroLatency]>,
1891     SchedVar<NoSchedPred,                          [WriteVecLogicY]>
1892 ]>;
1893 def : InstRW<[HWWriteVZeroIdiomLogicY], (instrs VPXORYrr)>;
1894
1895 def HWWriteVZeroIdiomALUX : SchedWriteVariant<[
1896     SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [HWWriteZeroLatency]>,
1897     SchedVar<NoSchedPred,                          [WriteVecALUX]>
1898 ]>;
1899 def : InstRW<[HWWriteVZeroIdiomALUX], (instrs PSUBBrr, VPSUBBrr,
1900                                               PSUBDrr, VPSUBDrr,
1901                                               PSUBQrr, VPSUBQrr,
1902                                               PSUBWrr, VPSUBWrr,
1903                                               PCMPGTBrr, VPCMPGTBrr,
1904                                               PCMPGTDrr, VPCMPGTDrr,
1905                                               PCMPGTWrr, VPCMPGTWrr)>;
1906
1907 def HWWriteVZeroIdiomALUY : SchedWriteVariant<[
1908     SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [HWWriteZeroLatency]>,
1909     SchedVar<NoSchedPred,                          [WriteVecALUY]>
1910 ]>;
1911 def : InstRW<[HWWriteVZeroIdiomALUY], (instrs VPSUBBYrr,
1912                                               VPSUBDYrr,
1913                                               VPSUBQYrr,
1914                                               VPSUBWYrr,
1915                                               VPCMPGTBYrr,
1916                                               VPCMPGTDYrr,
1917                                               VPCMPGTWYrr)>;
1918
1919 def HWWritePCMPGTQ : SchedWriteRes<[HWPort0]> {
1920   let Latency = 5;
1921   let NumMicroOps = 1;
1922   let ResourceCycles = [1];
1923 }
1924
1925 def HWWriteVZeroIdiomPCMPGTQ : SchedWriteVariant<[
1926     SchedVar<MCSchedPredicate<ZeroIdiomPredicate>, [HWWriteZeroLatency]>,
1927     SchedVar<NoSchedPred,                          [HWWritePCMPGTQ]>
1928 ]>;
1929 def : InstRW<[HWWriteVZeroIdiomPCMPGTQ], (instrs PCMPGTQrr, VPCMPGTQrr,
1930                                                  VPCMPGTQYrr)>;
1931
1932
1933 // The 0x83 ADC/SBB opcodes have special support for immediate 0 to only require
1934 // a single uop. It does not apply to the GR8 encoding. And only applies to the
1935 // 8-bit immediate since using larger immediate for 0 would be silly.
1936 // Unfortunately, this optimization does not apply to the AX/EAX/RAX short
1937 // encodings we convert to in MCInstLowering so we exclude AX/EAX/RAX here since
1938 // we schedule before that point.
1939 // TODO: Should we disable using the short encodings on these CPUs?
1940 def HWFastADC0 : MCSchedPredicate<
1941   CheckAll<[
1942     CheckImmOperand<2, 0>,              // Second MCOperand is Imm and has value 0.
1943     CheckNot<CheckRegOperand<1, AX>>,   // First MCOperand is not register AX
1944     CheckNot<CheckRegOperand<1, EAX>>,  // First MCOperand is not register EAX
1945     CheckNot<CheckRegOperand<1, RAX>>   // First MCOperand is not register RAX
1946   ]>
1947 >;
1948
1949 def HWWriteADC0 : SchedWriteRes<[HWPort06]> {
1950   let Latency = 1;
1951   let NumMicroOps = 1;
1952   let ResourceCycles = [1];
1953 }
1954
1955 def HWWriteADC : SchedWriteVariant<[
1956   SchedVar<HWFastADC0, [HWWriteADC0]>,
1957   SchedVar<NoSchedPred, [WriteADC]>
1958 ]>;
1959
1960 def : InstRW<[HWWriteADC], (instrs ADC16ri8, ADC32ri8, ADC64ri8,
1961                                       SBB16ri8, SBB32ri8, SBB64ri8)>;
1962
1963 // CMOVs that use both Z and C flag require an extra uop.
1964 def HWWriteCMOVA_CMOVBErr : SchedWriteRes<[HWPort06,HWPort0156]> {
1965   let Latency = 3;
1966   let ResourceCycles = [1,2];
1967   let NumMicroOps = 3;
1968 }
1969
1970 def HWWriteCMOVA_CMOVBErm : SchedWriteRes<[HWPort23,HWPort06,HWPort0156]> {
1971   let Latency = 8;
1972   let ResourceCycles = [1,1,2];
1973   let NumMicroOps = 4;
1974 }
1975
1976 def HWCMOVA_CMOVBErr :  SchedWriteVariant<[
1977   SchedVar<MCSchedPredicate<IsCMOVArr_Or_CMOVBErr>, [HWWriteCMOVA_CMOVBErr]>,
1978   SchedVar<NoSchedPred,                             [WriteCMOV]>
1979 ]>;
1980
1981 def HWCMOVA_CMOVBErm :  SchedWriteVariant<[
1982   SchedVar<MCSchedPredicate<IsCMOVArm_Or_CMOVBErm>, [HWWriteCMOVA_CMOVBErm]>,
1983   SchedVar<NoSchedPred,                             [WriteCMOV.Folded]>
1984 ]>;
1985
1986 def : InstRW<[HWCMOVA_CMOVBErr], (instrs CMOV16rr, CMOV32rr, CMOV64rr)>;
1987 def : InstRW<[HWCMOVA_CMOVBErm], (instrs CMOV16rm, CMOV32rm, CMOV64rm)>;
1988
1989 // SETCCs that use both Z and C flag require an extra uop.
1990 def HWWriteSETA_SETBEr : SchedWriteRes<[HWPort06,HWPort0156]> {
1991   let Latency = 2;
1992   let ResourceCycles = [1,1];
1993   let NumMicroOps = 2;
1994 }
1995
1996 def HWWriteSETA_SETBEm : SchedWriteRes<[HWPort4,HWPort237,HWPort06,HWPort0156]> {
1997   let Latency = 3;
1998   let ResourceCycles = [1,1,1,1];
1999   let NumMicroOps = 4;
2000 }
2001
2002 def HWSETA_SETBErr :  SchedWriteVariant<[
2003   SchedVar<MCSchedPredicate<IsSETAr_Or_SETBEr>, [HWWriteSETA_SETBEr]>,
2004   SchedVar<NoSchedPred,                         [WriteSETCC]>
2005 ]>;
2006
2007 def HWSETA_SETBErm :  SchedWriteVariant<[
2008   SchedVar<MCSchedPredicate<IsSETAm_Or_SETBEm>, [HWWriteSETA_SETBEm]>,
2009   SchedVar<NoSchedPred,                         [WriteSETCCStore]>
2010 ]>;
2011
2012 def : InstRW<[HWSETA_SETBErr], (instrs SETCCr)>;
2013 def : InstRW<[HWSETA_SETBErm], (instrs SETCCm)>;
2014
2015 } // SchedModel