llvm-project/llvm/lib/Target/X86/X86ScheduleBtVer2.td

//=- X86ScheduleBtVer2.td - X86 BtVer2 (Jaguar) Scheduling ---*- tablegen -*-=//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines the machine model for AMD btver2 (Jaguar) to support
// instruction scheduling and other instruction cost heuristics. Based off AMD Software
// Optimization Guide for AMD Family 16h Processors & Instruction Latency appendix.
//
//===----------------------------------------------------------------------===//

def BtVer2Model : SchedMachineModel {
  // All x86 instructions are modeled as a single micro-op, and btver2 can
  // decode 2 instructions per cycle.
  let IssueWidth = 2;
  let MicroOpBufferSize = 64; // Retire Control Unit
  let LoadLatency = 5; // FPU latency (worse case cf Integer 3 cycle latency)
  let HighLatency = 25;
  let MispredictPenalty = 14; // Minimum branch misdirection penalty
  let PostRAScheduler = 1;

  // FIXME: SSE4/AVX is unimplemented. This flag is set to allow
  // the scheduler to assign a default model to unrecognized opcodes.
  let CompleteModel = 0;
}

let SchedModel = BtVer2Model in {

// Jaguar can issue up to 6 micro-ops in one cycle
def JALU0 : ProcResource<1>; // Integer Pipe0: integer ALU0 (also handle FP->INT jam)
def JALU1 : ProcResource<1>; // Integer Pipe1: integer ALU1/MUL/DIV
def JLAGU : ProcResource<1>; // Integer Pipe2: LAGU
def JSAGU : ProcResource<1>; // Integer Pipe3: SAGU (also handles 3-operand LEA)
def JFPU0 : ProcResource<1>; // Vector/FPU Pipe0: VALU0/VIMUL/FPA
def JFPU1 : ProcResource<1>; // Vector/FPU Pipe1: VALU1/STC/FPM

// Any pipe - FIXME we need this until we can discriminate between int/fpu load/store/moves properly
def JAny : ProcResGroup<[JALU0, JALU1, JLAGU, JSAGU, JFPU0, JFPU1]>;

// Integer Pipe Scheduler
def JALU01 : ProcResGroup<[JALU0, JALU1]> {
  let BufferSize=20;
}

// AGU Pipe Scheduler
def JLSAGU : ProcResGroup<[JLAGU, JSAGU]> {
  let BufferSize=12;
}

// Fpu Pipe Scheduler
def JFPU01 : ProcResGroup<[JFPU0, JFPU1]> {
  let BufferSize=18;
}

def JDiv    : ProcResource<1>; // integer division
def JMul    : ProcResource<1>; // integer multiplication
def JVALU0  : ProcResource<1>; // vector integer
def JVALU1  : ProcResource<1>; // vector integer
def JVIMUL  : ProcResource<1>; // vector integer multiplication
def JSTC    : ProcResource<1>; // vector store/convert
def JFPM    : ProcResource<1>; // FP multiplication
def JFPA    : ProcResource<1>; // FP addition

// Integer loads are 3 cycles, so ReadAfterLd registers needn't be available until 3
// cycles after the memory operand.
def : ReadAdvance<ReadAfterLd, 3>;

// Many SchedWrites are defined in pairs with and without a folded load.
// Instructions with folded loads are usually micro-fused, so they only appear
// as two micro-ops when dispatched by the schedulers.
// This multiclass defines the resource usage for variants with and without
// folded loads.
multiclass JWriteResIntPair<X86FoldableSchedWrite SchedRW,
                          ProcResourceKind ExePort,
                          int Lat> {
  // Register variant is using a single cycle on ExePort.
  def : WriteRes<SchedRW, [ExePort]> { let Latency = Lat; }

  // Memory variant also uses a cycle on JLAGU and adds 3 cycles to the
  // latency.
  def : WriteRes<SchedRW.Folded, [JLAGU, ExePort]> {
     let Latency = !add(Lat, 3);
  }
}

multiclass JWriteResFpuPair<X86FoldableSchedWrite SchedRW,
                          ProcResourceKind ExePort,
                          int Lat> {
  // Register variant is using a single cycle on ExePort.
  def : WriteRes<SchedRW, [ExePort]> { let Latency = Lat; }

  // Memory variant also uses a cycle on JLAGU and adds 5 cycles to the
  // latency.
  def : WriteRes<SchedRW.Folded, [JLAGU, ExePort]> {
     let Latency = !add(Lat, 5);
  }
}

// A folded store needs a cycle on the SAGU for the store data.
def : WriteRes<WriteRMW, [JSAGU]>;

////////////////////////////////////////////////////////////////////////////////
// Arithmetic.
////////////////////////////////////////////////////////////////////////////////

defm : JWriteResIntPair<WriteALU,   JALU01, 1>;
defm : JWriteResIntPair<WriteIMul,  JALU1,  3>;

def  : WriteRes<WriteIMulH, [JALU1]> {
  let Latency = 6;
  let ResourceCycles = [4];
}

// FIXME 8/16 bit divisions
def : WriteRes<WriteIDiv, [JALU1, JDiv]> {
  let Latency = 25;
  let ResourceCycles = [1, 25];
}
def : WriteRes<WriteIDivLd, [JALU1, JLAGU, JDiv]> {
  let Latency = 41;
  let ResourceCycles = [1, 1, 25];
}

// This is for simple LEAs with one or two input operands.
// FIXME: SAGU 3-operand LEA
def : WriteRes<WriteLEA, [JALU01]>;

////////////////////////////////////////////////////////////////////////////////
// Integer shifts and rotates.
////////////////////////////////////////////////////////////////////////////////

defm : JWriteResIntPair<WriteShift, JALU01, 1>;

def WriteSHLDrri : SchedWriteRes<[JALU01]> {
  let Latency = 3;
  let ResourceCycles = [6];
  let NumMicroOps = 6;
}
def: InstRW<[WriteSHLDrri], (instrs SHLD16rri8, SHLD32rri8, SHLD64rri8,
                                    SHRD16rri8, SHRD32rri8, SHRD64rri8)>;

def WriteSHLDrrCL : SchedWriteRes<[JALU01]> {
  let Latency = 4;
  let ResourceCycles = [8];
  let NumMicroOps = 7;
}
def: InstRW<[WriteSHLDrrCL], (instrs SHLD16rrCL, SHLD32rrCL, SHLD64rrCL,
                                     SHRD16rrCL, SHRD32rrCL, SHRD64rrCL)>;

def WriteSHLDm : SchedWriteRes<[JLAGU, JALU01]> {
  let Latency = 9;
  let ResourceCycles = [1, 22];
  let NumMicroOps = 8;
}
def: InstRW<[WriteSHLDm],(instrs SHLD16mri8, SHLD32mri8, SHLD64mri8,
                                 SHLD16mrCL, SHLD32mrCL, SHLD64mrCL,
                                 SHRD16mri8, SHRD32mri8, SHRD64mri8,
                                 SHRD16mrCL, SHRD32mrCL, SHRD64mrCL)>;

////////////////////////////////////////////////////////////////////////////////
// Loads, stores, and moves, not folded with other operations.
// FIXME: Split x86 and SSE load/store/moves
////////////////////////////////////////////////////////////////////////////////

def : WriteRes<WriteLoad,  [JLAGU]> { let Latency = 5; }
def : WriteRes<WriteStore, [JSAGU]>;
def : WriteRes<WriteMove,  [JALU01]>;

// Treat misc copies as a move.
def : InstRW<[WriteMove], (instrs COPY)>;

////////////////////////////////////////////////////////////////////////////////
// Idioms that clear a register, like xorps %xmm0, %xmm0.
// These can often bypass execution ports completely.
////////////////////////////////////////////////////////////////////////////////

def : WriteRes<WriteZero,  []>;

////////////////////////////////////////////////////////////////////////////////
// Branches don't produce values, so they have no latency, but they still
// consume resources. Indirect branches can fold loads.
////////////////////////////////////////////////////////////////////////////////

defm : JWriteResIntPair<WriteJump,  JALU01, 1>;

////////////////////////////////////////////////////////////////////////////////
// Floating point. This covers both scalar and vector operations.
// FIXME: should we bother splitting JFPU pipe + unit stages for fast instructions?
// FIXME: Double precision latencies
// FIXME: SS vs PS latencies
// FIXME: ymm latencies
////////////////////////////////////////////////////////////////////////////////

defm : JWriteResFpuPair<WriteFAdd,        JFPU0,  3>;
defm : JWriteResFpuPair<WriteFMul,        JFPU1,  2>;
defm : JWriteResFpuPair<WriteFMA,         JFPU1,  2>; // NOTE: Doesn't exist on Jaguar.
defm : JWriteResFpuPair<WriteFRcp,        JFPU1,  2>;
defm : JWriteResFpuPair<WriteFRsqrt,      JFPU1,  2>;
defm : JWriteResFpuPair<WriteFShuffle,   JFPU01,  1>;
defm : JWriteResFpuPair<WriteFBlend,     JFPU01,  1>;
defm : JWriteResFpuPair<WriteFShuffle256, JFPU01, 1>;

def : WriteRes<WriteFSqrt, [JFPU1, JLAGU, JFPM]> {
  let Latency = 21;
  let ResourceCycles = [1, 1, 21];
}
def : WriteRes<WriteFSqrtLd, [JFPU1, JLAGU, JFPM]> {
  let Latency = 26;
  let ResourceCycles = [1, 1, 21];
}

def : WriteRes<WriteFDiv, [JFPU1, JLAGU, JFPM]> {
  let Latency = 19;
  let ResourceCycles = [1, 1, 19];
}
def : WriteRes<WriteFDivLd, [JFPU1, JLAGU, JFPM]> {
  let Latency = 24;
  let ResourceCycles = [1, 1, 19];
}

// FIXME: integer pipes
defm : JWriteResFpuPair<WriteCvtF2I,    JFPU1,  3>; // Float -> Integer.
defm : JWriteResFpuPair<WriteCvtI2F,    JFPU1,  3>; // Integer -> Float.
defm : JWriteResFpuPair<WriteCvtF2F,    JFPU1,  3>; // Float -> Float size conversion.

def : WriteRes<WriteFVarBlend, [JFPU01]> {
  let Latency = 2;
  let ResourceCycles = [4];
  let NumMicroOps = 3;
}
def : WriteRes<WriteFVarBlendLd, [JLAGU, JFPU01]> {
  let Latency = 7;
  let ResourceCycles = [1, 4];
  let NumMicroOps = 3;
}

// Vector integer operations.
defm : JWriteResFpuPair<WriteVecALU,   JFPU01,  1>;
defm : JWriteResFpuPair<WriteVecShift, JFPU01,  1>;
defm : JWriteResFpuPair<WriteVecIMul,  JFPU0,   2>;
defm : JWriteResFpuPair<WriteShuffle,  JFPU01,  1>;
defm : JWriteResFpuPair<WriteBlend,    JFPU01,  1>;
defm : JWriteResFpuPair<WriteVecLogic, JFPU01,  1>;
defm : JWriteResFpuPair<WriteShuffle256, JFPU01, 1>;

def : WriteRes<WriteVarBlend, [JFPU01]> {
  let Latency = 2;
  let ResourceCycles = [4];
  let NumMicroOps = 3;
}
def : WriteRes<WriteVarBlendLd, [JLAGU, JFPU01]> {
  let Latency = 7;
  let ResourceCycles = [1, 4];
  let NumMicroOps = 3;
}

// FIXME: why do we need to define AVX2 resource on CPU that doesn't have AVX2?
def : WriteRes<WriteVarVecShift, [JFPU01]> {}
def : WriteRes<WriteVarVecShiftLd, [JLAGU, JFPU01]> {
  let Latency = 6;
  let ResourceCycles = [1, 2];
}

def : WriteRes<WriteMPSAD, [JFPU0]> {
  let Latency = 3;
  let ResourceCycles = [2];
}
def : WriteRes<WriteMPSADLd, [JLAGU, JFPU0]> {
  let Latency = 8;
  let ResourceCycles = [1, 2];
}

////////////////////////////////////////////////////////////////////////////////
// String instructions.
// Packed Compare Implicit Length Strings, Return Mask
// FIXME: approximate latencies + pipe dependencies
////////////////////////////////////////////////////////////////////////////////

def : WriteRes<WritePCmpIStrM, [JFPU1,JFPU0]> {
  let Latency = 8;
  let ResourceCycles = [2, 2];
  let NumMicroOps = 3;
}
def : WriteRes<WritePCmpIStrMLd, [JLAGU, JFPU1, JFPU0]> {
  let Latency = 13;
  let ResourceCycles = [1, 2, 2];
  let NumMicroOps = 3;
}

// Packed Compare Explicit Length Strings, Return Mask
def : WriteRes<WritePCmpEStrM, [JFPU1, JLAGU, JFPU01,JFPU1, JFPU0]> {
  let Latency = 14;
  let ResourceCycles = [5, 5, 5, 5, 5];
  let NumMicroOps = 9;
}
def : WriteRes<WritePCmpEStrMLd, [JLAGU, JFPU1, JLAGU, JFPU01,JFPU1, JFPU0]> {
  let Latency = 19;
  let ResourceCycles = [1, 5, 5, 5, 5, 5];
  let NumMicroOps = 9;
}

// Packed Compare Implicit Length Strings, Return Index
def : WriteRes<WritePCmpIStrI, [JFPU1, JFPU0]> {
  let Latency = 7;
  let ResourceCycles = [2, 2];
}
def : WriteRes<WritePCmpIStrILd, [JLAGU, JFPU1, JFPU0]> {
  let Latency = 12;
  let ResourceCycles = [1, 2, 2];
}

// Packed Compare Explicit Length Strings, Return Index
def : WriteRes<WritePCmpEStrI, [JFPU1, JLAGU, JFPU01,JFPU1, JFPU0]> {
  let Latency = 14;
  let ResourceCycles = [5, 5, 5, 5, 5];
  let NumMicroOps = 9;
}
def : WriteRes<WritePCmpEStrILd, [JLAGU, JFPU1, JLAGU, JFPU01,JFPU1, JFPU0]> {
  let Latency = 19;
  let ResourceCycles = [1, 5, 5, 5, 5, 5];
  let NumMicroOps = 9;
}

////////////////////////////////////////////////////////////////////////////////
// AES Instructions.
////////////////////////////////////////////////////////////////////////////////

def : WriteRes<WriteAESDecEnc, [JFPU01, JVIMUL]> {
  let Latency = 3;
  let ResourceCycles = [1, 1];
}
def : WriteRes<WriteAESDecEncLd, [JFPU01, JLAGU, JVIMUL]> {
  let Latency = 8;
  let ResourceCycles = [1, 1, 1];
}

def : WriteRes<WriteAESIMC, [JVIMUL]> {
  let Latency = 2;
  let ResourceCycles = [1];
}
def : WriteRes<WriteAESIMCLd, [JLAGU, JVIMUL]> {
  let Latency = 7;
  let ResourceCycles = [1, 1];
}

def : WriteRes<WriteAESKeyGen, [JVIMUL]> {
  let Latency = 2;
  let ResourceCycles = [1];
}
def : WriteRes<WriteAESKeyGenLd, [JLAGU, JVIMUL]> {
  let Latency = 7;
  let ResourceCycles = [1, 1];
}

////////////////////////////////////////////////////////////////////////////////
// Horizontal add/sub  instructions.
////////////////////////////////////////////////////////////////////////////////

def : WriteRes<WriteFHAdd, [JFPU0]> {
  let Latency = 3;
}

def : WriteRes<WriteFHAddLd, [JLAGU, JFPU0]> {
  let Latency = 8;
}

def : WriteRes<WritePHAdd, [JFPU01]> {
  let ResourceCycles = [1];
}
def : WriteRes<WritePHAddLd, [JLAGU, JFPU01 ]> {
  let Latency = 6;
  let ResourceCycles = [1, 1];
}

def WriteFHAddY: SchedWriteRes<[JFPU0]> {
  let Latency = 3;
  let ResourceCycles = [2];
}
def : InstRW<[WriteFHAddY], (instrs VHADDPDYrr, VHADDPSYrr, VHSUBPDYrr, VHSUBPSYrr)>;

def WriteFHAddYLd: SchedWriteRes<[JLAGU, JFPU0]> {
  let Latency = 8;
  let ResourceCycles = [1, 2];
}
def : InstRW<[WriteFHAddYLd], (instrs VHADDPDYrm, VHADDPSYrm, VHSUBPDYrm, VHSUBPSYrm)>;

////////////////////////////////////////////////////////////////////////////////
// Carry-less multiplication instructions.
////////////////////////////////////////////////////////////////////////////////

def : WriteRes<WriteCLMul, [JVIMUL]> {
  let Latency = 2;
  let ResourceCycles = [1];
}
def : WriteRes<WriteCLMulLd, [JLAGU, JVIMUL]> {
  let Latency = 7;
  let ResourceCycles = [1, 1];
}

// FIXME: pipe for system/microcode?
def : WriteRes<WriteSystem,     [JAny]> { let Latency = 100; }
def : WriteRes<WriteMicrocoded, [JAny]> { let Latency = 100; }
def : WriteRes<WriteFence,  [JSAGU]>;
def : WriteRes<WriteNop, []>;

////////////////////////////////////////////////////////////////////////////////
// SSE4.1 instructions.
////////////////////////////////////////////////////////////////////////////////

def WriteDPPS: SchedWriteRes<[JFPU0, JFPU1]> {
  let Latency = 11;
  let ResourceCycles = [3,3];
  let NumMicroOps = 5;
}
def : InstRW<[WriteDPPS], (instrs DPPSrri, VDPPSrri)>;

def WriteDPPSLd: SchedWriteRes<[JLAGU, JFPU0, JFPU1]> {
  let Latency = 16;
  let ResourceCycles = [1,3,3];
  let NumMicroOps = 6;
}
def : InstRW<[WriteDPPSLd], (instrs DPPSrmi, VDPPSrmi)>;

def WriteDPPD: SchedWriteRes<[JFPU0, JFPU1]> {
  let Latency = 9;
  let ResourceCycles = [3,3];
  let NumMicroOps = 3;
}
def : InstRW<[WriteDPPD], (instrs DPPDrri, VDPPDrri)>;

def WriteDPPDLd: SchedWriteRes<[JLAGU, JFPU0, JFPU1]> {
  let Latency = 14;
  let ResourceCycles = [1,3,3];
  let NumMicroOps = 3;
}
def : InstRW<[WriteDPPDLd], (instrs DPPDrmi, VDPPDrmi)>;

////////////////////////////////////////////////////////////////////////////////
// SSE4A instructions.
////////////////////////////////////////////////////////////////////////////////

def WriteEXTRQ: SchedWriteRes<[JFPU01]> {
  let Latency = 1;
  let ResourceCycles = [1];
}
def : InstRW<[WriteEXTRQ], (instrs EXTRQ, EXTRQI)>;

def WriteINSERTQ: SchedWriteRes<[JFPU01]> {
  let Latency = 2;
  let ResourceCycles = [4];
}
def : InstRW<[WriteINSERTQ], (instrs INSERTQ, INSERTQI)>;

////////////////////////////////////////////////////////////////////////////////
// F16C instructions.
////////////////////////////////////////////////////////////////////////////////

def WriteCVT3: SchedWriteRes<[JFPU1]> {
  let Latency = 3;
}
def : InstRW<[WriteCVT3], (instrs VCVTPS2PHrr, VCVTPH2PSrr)>;

def WriteCVT3St: SchedWriteRes<[JFPU1, JSAGU]> {
  let Latency = 3;
  let ResourceCycles = [1, 1];
}
def : InstRW<[WriteCVT3St], (instrs VCVTPS2PHmr)>;

def WriteCVT3Ld: SchedWriteRes<[JLAGU, JFPU1]> {
  let Latency = 8;
  let ResourceCycles = [1, 1];
}
def : InstRW<[WriteCVT3Ld], (instrs VCVTPH2PSrm)>;

def WriteCVTPS2PHY: SchedWriteRes<[JFPU1, JFPU01]> {
  let Latency = 6;
  let ResourceCycles = [2,2];
  let NumMicroOps = 3;
}
def : InstRW<[WriteCVTPS2PHY], (instrs VCVTPS2PHYrr)>;

def WriteCVTPS2PHYSt: SchedWriteRes<[JFPU1, JFPU01, JSAGU]> {
  let Latency = 11;
  let ResourceCycles = [2,2,1];
  let NumMicroOps = 3;
}
def : InstRW<[WriteCVTPS2PHYSt], (instrs VCVTPS2PHYmr)>;

def WriteCVTPH2PSY: SchedWriteRes<[JFPU1]> {
  let Latency = 3;
  let ResourceCycles = [2];
  let NumMicroOps = 2;
}
def : InstRW<[WriteCVTPH2PSY], (instrs VCVTPH2PSYrr)>;

def WriteCVTPH2PSYLd: SchedWriteRes<[JLAGU, JFPU1]> {
  let Latency = 8;
  let ResourceCycles = [1,2];
  let NumMicroOps = 2;
}
def : InstRW<[WriteCVTPH2PSYLd], (instrs VCVTPH2PSYrm)>;

////////////////////////////////////////////////////////////////////////////////
// AVX instructions.
////////////////////////////////////////////////////////////////////////////////

def WriteLogicY: SchedWriteRes<[JFPU01]> {
  let Latency = 1;
  let ResourceCycles = [2];
  let NumMicroOps = 2;
}
def : InstRW<[WriteLogicY], (instrs VORPDYrr, VORPSYrr,
                                    VXORPDYrr, VXORPSYrr,
                                    VANDPDYrr, VANDPSYrr,
                                    VANDNPDYrr, VANDNPSYrr)>;

def WriteLogicYLd: SchedWriteRes<[JLAGU, JFPU01]> {
  let Latency = 6;
  let ResourceCycles = [1, 2];
  let NumMicroOps = 3;
}
def : InstRW<[WriteLogicYLd], (instrs VORPDYrm, VORPSYrm,
                                      VXORPDYrm, VXORPSYrm,
                                      VANDPDYrm, VANDPSYrm,
                                      VANDNPDYrm, VANDNPSYrm)>;

def WriteVDPPSY: SchedWriteRes<[JFPU1, JFPU0]> {
  let Latency = 12;
  let ResourceCycles = [6, 6];
  let NumMicroOps = 10;
}
def : InstRW<[WriteVDPPSY], (instrs VDPPSYrri)>;

def WriteVDPPSYLd: SchedWriteRes<[JLAGU, JFPU1, JFPU0]> {
  let Latency = 17;
  let ResourceCycles = [1, 6, 6];
  let NumMicroOps = 11;
}
def : InstRW<[WriteVDPPSYLd, ReadAfterLd], (instrs VDPPSYrmi)>;

def WriteFAddY: SchedWriteRes<[JFPU0]> {
  let Latency = 3;
  let ResourceCycles = [2];
}
def : InstRW<[WriteFAddY], (instrs VADDPDYrr, VADDPSYrr,
                                   VSUBPDYrr, VSUBPSYrr,
                                   VADDSUBPDYrr, VADDSUBPSYrr)>;

def WriteFAddYLd: SchedWriteRes<[JLAGU, JFPU0]> {
  let Latency = 8;
  let ResourceCycles = [1, 2];
}
def : InstRW<[WriteFAddYLd, ReadAfterLd], (instrs VADDPDYrm, VADDPSYrm,
                                                  VSUBPDYrm, VSUBPSYrm,
                                                  VADDSUBPDYrm, VADDSUBPSYrm)>;

def WriteFDivY: SchedWriteRes<[JFPU1]> {
  let Latency = 38;
  let ResourceCycles = [38];
}
def : InstRW<[WriteFDivY], (instrs VDIVPDYrr, VDIVPSYrr)>;

def WriteFDivYLd: SchedWriteRes<[JLAGU, JFPU1]> {
  let Latency = 43;
  let ResourceCycles = [1, 38];
}
def : InstRW<[WriteFDivYLd, ReadAfterLd], (instrs VDIVPDYrm, VDIVPSYrm)>;

def WriteVMULYPD: SchedWriteRes<[JFPU1]> {
  let Latency = 4;
  let ResourceCycles = [4];
}
def : InstRW<[WriteVMULYPD], (instrs VMULPDYrr)>;

def WriteVMULYPDLd: SchedWriteRes<[JLAGU, JFPU1]> {
  let Latency = 9;
  let ResourceCycles = [1, 4];
}
def : InstRW<[WriteVMULYPDLd, ReadAfterLd], (instrs VMULPDYrm)>;

def WriteVMULYPS: SchedWriteRes<[JFPU1]> {
  let Latency = 2;
  let ResourceCycles = [2];
}
def : InstRW<[WriteVMULYPS], (instrs VMULPSYrr, VRCPPSYr, VRSQRTPSYr)>;

def WriteVMULYPSLd: SchedWriteRes<[JLAGU, JFPU1]> {
  let Latency = 7;
  let ResourceCycles = [1, 2];
}
def : InstRW<[WriteVMULYPSLd, ReadAfterLd], (instrs VMULPSYrm, VRCPPSYm, VRSQRTPSYm)>;

def WriteVMULPD: SchedWriteRes<[JFPU1]> {
  let Latency = 4;
  let ResourceCycles = [2];
}
def : InstRW<[WriteVMULPD], (instrs MULPDrr, MULSDrr, VMULPDrr, VMULSDrr)>;

def WriteVMULPDLd: SchedWriteRes<[JLAGU, JFPU1]> {
  let Latency = 9;
  let ResourceCycles = [1, 2];
}
def : InstRW<[WriteVMULPDLd], (instrs MULPDrm, MULSDrm, VMULPDrm, VMULSDrm)>;

def WriteVCVTY: SchedWriteRes<[JSTC]> {
  let Latency = 3;
  let ResourceCycles = [2];
}
def : InstRW<[WriteVCVTY], (instrs VCVTDQ2PDYrr, VCVTDQ2PSYrr,
                                   VCVTPS2DQYrr, VCVTTPS2DQYrr,
                                   VROUNDYPDr,   VROUNDYPSr)>;

def WriteVCVTYLd: SchedWriteRes<[JLAGU, JSTC]> {
  let Latency = 8;
  let ResourceCycles = [1, 2];
}
def : InstRW<[WriteVCVTYLd, ReadAfterLd], (instrs VCVTDQ2PDYrm, VCVTDQ2PSYrm,
                                                  VCVTPS2DQYrm, VCVTTPS2DQYrm,
                                                  VROUNDYPDm,   VROUNDYPSm)>;

def WriteVMOVNTDQSt: SchedWriteRes<[JSTC, JSAGU]> {
  let Latency = 2;
}
def : InstRW<[WriteVMOVNTDQSt], (instrs MOVNTDQmr, VMOVNTDQmr)>;

def WriteMOVNTSt: SchedWriteRes<[JSTC, JSAGU]> {
  let Latency = 3;
}
def : InstRW<[WriteMOVNTSt], (instrs MOVNTPDmr, MOVNTPSmr, MOVNTSD, MOVNTSS, VMOVNTPDmr, VMOVNTPSmr)>;

def WriteVMOVNTPYSt: SchedWriteRes<[JSTC, JSAGU]> {
  let Latency = 3;
  let ResourceCycles = [2,1];
}
def : InstRW<[WriteVMOVNTPYSt], (instrs VMOVNTDQYmr, VMOVNTPDYmr, VMOVNTPSYmr)>;

def WriteFCmp: SchedWriteRes<[JFPU0]> {
  let Latency = 2;
}

def : InstRW<[WriteFCmp], (instregex "(V)?M(AX|IN)(P|S)(D|S)rr",
                                     "(V)?CMPP(S|D)rri", "(V)?CMPS(S|D)rr")>;

def WriteFCmpLd: SchedWriteRes<[JLAGU, JFPU0]> {
  let Latency = 7;
}

def : InstRW<[WriteFCmpLd], (instregex "(V)?M(AX|IN)(P|S)(D|S)rm",
                                       "(V)?CMPP(S|D)rmi", "(V)?CMPS(S|D)rm")>;

def WriteVCVTPDY: SchedWriteRes<[JSTC, JFPU01]> {
  let Latency = 6;
  let ResourceCycles = [2, 4];
}
def : InstRW<[WriteVCVTPDY], (instrs VCVTPD2DQYrr, VCVTTPD2DQYrr, VCVTPD2PSYrr)>;

def WriteVCVTPDYLd: SchedWriteRes<[JLAGU, JSTC, JFPU01]> {
  let Latency = 11;
  let ResourceCycles = [1, 2, 4];
}
def : InstRW<[WriteVCVTPDYLd, ReadAfterLd], (instrs VCVTPD2DQYrm, VCVTTPD2DQYrm, VCVTPD2PSYrm)>;

def WriteVBlendVPY: SchedWriteRes<[JFPU01]> {
  let Latency = 3;
  let ResourceCycles = [6];
}
def : InstRW<[WriteVBlendVPY], (instrs VBLENDVPDYrr, VBLENDVPSYrr, VPERMILPDYrr, VPERMILPSYrr)>;

def WriteVBlendVPYLd: SchedWriteRes<[JLAGU, JFPU01]> {
  let Latency = 8;
  let ResourceCycles = [1, 6];
}
def : InstRW<[WriteVBlendVPYLd, ReadAfterLd], (instrs VBLENDVPDYrm, VBLENDVPSYrm)>;

def WriteVBROADCASTYLd: SchedWriteRes<[JLAGU, JFPU01]> {
  let Latency = 6;
  let ResourceCycles = [1, 4];
}
def : InstRW<[WriteVBROADCASTYLd, ReadAfterLd], (instrs VBROADCASTSDYrm, VBROADCASTSSYrm)>;

def WriteFPAY22: SchedWriteRes<[JFPU0]> {
  let Latency = 2;
  let ResourceCycles = [2];
}
def : InstRW<[WriteFPAY22], (instregex "VCMPP(S|D)Yrri", "VM(AX|IN)P(D|S)Yrr")>;

def WriteFPAY22Ld: SchedWriteRes<[JLAGU, JFPU0]> {
  let Latency = 7;
  let ResourceCycles = [1, 2];
}
def : InstRW<[WriteFPAY22Ld, ReadAfterLd], (instregex "VCMPP(S|D)Yrmi", "VM(AX|IN)P(D|S)Yrm")>;

def WriteVHAddSubY: SchedWriteRes<[JFPU0]> {
  let Latency = 3;
  let ResourceCycles = [2];
}
def : InstRW<[WriteVHAddSubY], (instrs VHADDPDYrr, VHADDPSYrr, VHSUBPDYrr, VHSUBPSYrr)>;

def WriteVHAddSubYLd: SchedWriteRes<[JLAGU, JFPU0]> {
  let Latency = 8;
  let ResourceCycles = [1, 2];
}
def : InstRW<[WriteVHAddSubYLd], (instrs VHADDPDYrm, VHADDPSYrm, VHSUBPDYrm, VHSUBPSYrm)>;

def WriteVMaskMovLd: SchedWriteRes<[JLAGU,JFPU01]> {
  let Latency = 6;
  let ResourceCycles = [1, 2];
}
def : InstRW<[WriteVMaskMovLd], (instrs VMASKMOVPDrm, VMASKMOVPSrm)>;

def WriteVMaskMovYLd: SchedWriteRes<[JLAGU,JFPU01]> {
  let Latency = 6;
  let ResourceCycles = [1, 4];
}
def : InstRW<[WriteVMaskMovYLd], (instrs VMASKMOVPDYrm, VMASKMOVPSYrm)>;

def WriteVMaskMovSt: SchedWriteRes<[JFPU01,JSAGU]> {
  let Latency = 6;
  let ResourceCycles = [4, 1];
}
def : InstRW<[WriteVMaskMovSt], (instrs VMASKMOVPDmr, VMASKMOVPSmr)>;

def WriteVMaskMovYSt: SchedWriteRes<[JFPU01,JSAGU]> {
  let Latency = 6;
  let ResourceCycles = [4, 1];
}
def : InstRW<[WriteVMaskMovYSt], (instrs VMASKMOVPDYmr, VMASKMOVPSYmr)>;

// TODO: In fact we have latency '2+i'. The +i represents an additional 1 cycle transfer
// operation which moves the floating point result to the integer unit. During this
// additional cycle the floating point unit execution resources are not occupied
// and ALU0 in the integer unit is occupied instead.
def WriteVMOVMSK: SchedWriteRes<[JFPU0]> {
  let Latency = 3;
}
def : InstRW<[WriteVMOVMSK], (instrs VMOVMSKPDrr, VMOVMSKPDYrr, VMOVMSKPSrr, VMOVMSKPSYrr)>;

// TODO: In fact we have latency '3+i'. The +i represents an additional 1 cycle transfer
// operation which moves the floating point result to the integer unit. During this
// additional cycle the floating point unit execution resources are not occupied
// and ALU0 in the integer unit is occupied instead.
def WriteVTESTY: SchedWriteRes<[JFPU01, JFPU0]> {
  let Latency = 4;
  let ResourceCycles = [2, 2];
  let NumMicroOps = 3;
}
def : InstRW<[WriteVTESTY], (instrs VPTESTYrr, VTESTPDYrr, VTESTPSYrr)>;

def WriteVTESTYLd: SchedWriteRes<[JLAGU, JFPU01, JFPU0]> {
  let Latency = 9;
  let ResourceCycles = [1, 2, 2];
  let NumMicroOps = 3;
}
def : InstRW<[WriteVTESTYLd], (instrs VPTESTYrm, VTESTPDYrm, VTESTPSYrm)>;

def WriteVTEST: SchedWriteRes<[JFPU0]> {
  let Latency = 3;
}
def : InstRW<[WriteVTEST], (instrs PTESTrr, VPTESTrr, VTESTPDrr, VTESTPSrr)>;

def WriteVTESTLd: SchedWriteRes<[JLAGU, JFPU0]> {
  let Latency = 8;
}
def : InstRW<[WriteVTESTLd], (instrs PTESTrm, VPTESTrm, VTESTPDrm, VTESTPSrm)>;

def WriteVSQRTYPD: SchedWriteRes<[JFPU1]> {
  let Latency = 54;
  let ResourceCycles = [54];
}
def : InstRW<[WriteVSQRTYPD], (instrs VSQRTPDYr)>;

def WriteVSQRTYPDLd: SchedWriteRes<[JLAGU, JFPU1]> {
  let Latency = 59;
  let ResourceCycles = [1, 54];
}
def : InstRW<[WriteVSQRTYPDLd], (instrs VSQRTPDYm)>;

def WriteVSQRTYPS: SchedWriteRes<[JFPU1]> {
  let Latency = 42;
  let ResourceCycles = [42];
}
def : InstRW<[WriteVSQRTYPS], (instrs VSQRTPSYr)>;

def WriteVSQRTYPSLd: SchedWriteRes<[JLAGU, JFPU1]> {
  let Latency = 47;
  let ResourceCycles = [1, 42];
}
def : InstRW<[WriteVSQRTYPSLd], (instrs VSQRTPSYm)>;

def WriteJVZEROALL: SchedWriteRes<[]> {
  let Latency = 90;
  let NumMicroOps = 73;
}
def : InstRW<[WriteJVZEROALL], (instrs VZEROALL)>;

def WriteJVZEROUPPER: SchedWriteRes<[]> {
  let Latency = 46;
  let NumMicroOps = 37;
}
def : InstRW<[WriteJVZEROUPPER], (instrs VZEROUPPER)>;
} // SchedModel