/*========================== begin_copyright_notice ============================

Copyright (C) 2017-2022 Intel Corporation

SPDX-License-Identifier: MIT

============================= end_copyright_notice ===========================*/

#include "Assertions.h"
#include "G4_Verifier.hpp"
#include "HWConformity.h"
#include "InstSplit.h"
#include "Optimizer.h"

using namespace vISA;

uint8_t HWConformity::checkMinExecSize(G4_opcode op) {
  if (op == G4_dp2 || op == G4_dp3 || op == G4_dp4 || op == G4_dph) {
    return 4;
  } else if (op == G4_line || op == G4_pln) {
    return 8;
  } else if (op == G4_sad2 || op == G4_sada2) {
    return 2;
  } else
    return 1;
}

void HWConformity::fixOpndTypeAlign(G4_BB *bb) {
  INST_LIST_ITER i = bb->begin();
  INST_LIST_ITER next_iter = i;
  bool needSplit = false;

  for (auto iEnd = bb->end(); i != iEnd; i = next_iter) {
    G4_INST *inst = *i;
    G4_opcode opcode = inst->opcode();
    if (opcode == G4_nop || opcode == G4_label || inst->isSend() ||
        inst->isDpas()) {
      next_iter++;
    } else if (fixInstOpndTypeAlign(i, bb)) {
      needSplit = true;
      next_iter = i;
      next_iter++;
    } else {
      next_iter++;
    }
#ifdef _DEBUG
    verifyG4Kernel(kernel, Optimizer::PI_HWConformityChk, false);
#endif
  }

  if (needSplit) {
    // make sure updated insts and new moves don't cross 2 GRF
    InstSplitPass instSplitter(&builder);
    instSplitter.runOnBB(bb);
  }
}

// Fix instructions with vector immediate as source operands.
//    mov (8) r5.0<2>:uw 0xfdb97531:uv {Align1}
// becomes
//    mov (8) r6.0<1>:uw 0xfdb97531:uv {Align1}
//    mov (8) r5.0<2>:uw r6.0<8;8,1>:uw {Align1, Q1}
//
// When an immediate vector is used in an instruction, the destination must
// be 128-bit aligned with destination horizontal stride equivalent to a
// word for an immediate integer vector (v) and equivalent to a DWord for an
// immediate float vector (vf).
// For Xe2+, :vf datatype is not supported and :v and :uv must not be used
// when destination is any of the float datatypes.
bool HWConformity::fixDstAlignmentWithVectorImm(INST_LIST_ITER iter,
                                                G4_BB *bb) {
  bool changed = false;
  G4_INST *inst = *iter;
  G4_DstRegRegion *reg = inst->getDst();
  uint8_t execSize = inst->getExecSize();

  bool dstAligned = builder.tryToAlignOperand(reg, 16);

  unsigned hsInBytes = reg->getHorzStride() * reg->getTypeSize();
  for (int k = 0, e = inst->getNumSrc(); k < e; ++k) {
    G4_Operand *src = inst->getSrc(k);
    if (!src->isVectImm())
      continue;

    G4_Type ty = src->getType();
    if (!builder.hasPackedRestrictedFloatVector())
      vISA_ASSERT(ty != Type_VF,
                  ":vf datatype is not supported on this platform");
    G4_Type moveTy = (ty == Type_V)    ? Type_W
                     : (ty == Type_UV) ? Type_UW
                                       : Type_F;
    if (!dstAligned || (builder.getPlatform() >= Xe2 &&
                        IS_TYPE_FLOAT_ALL(reg->getType()))) {
      inst->setSrc(insertMovBefore(iter, k, moveTy, bb), k);
      changed = true;
    } else if (hsInBytes != TypeSize(moveTy)) {
      if (hsInBytes == 4 && execSize < 8) {
        // for the case where dst is dword and execution size is < 8,
        // we can interleave the vector to avoid a move
        // e.g., mov (2) r1.0<1>:d 0x21:uv  -->
        //       mov (2) r1.0<1>:d 0x0201:uv
        uint64_t bitValue = 0;
        uint32_t immBits = static_cast<uint32_t>(src->asImm()->getImm());
        for (int i = 0; i < execSize; ++i) {
          uint64_t val = (immBits >> (i * 4)) & 0xF;
          bitValue |= val << (i * 8);
        }
        inst->setSrc(builder.createImm(bitValue, ty), k);
      } else {
        inst->setSrc(insertMovBefore(iter, k, moveTy, bb), k);
        changed = true;
      }
    }
  }

  return changed;
}

// Do basic HW conformity check related to operand type and dst alignment before
// resucing execution size to avoid splitting of the MOV inserted in this stage.
// This function is called for some instructions generated in later stages.
bool HWConformity::fixInstOpndTypeAlign(INST_LIST_ITER i, G4_BB *bb) {
  G4_INST *inst = *i;
  bool insertedInst = false;

  if (inst->opcode() == G4_srnd) {
    // Operands can be packed.
    return false;
  }

  int extypesize = 0;
  G4_Type extype = inst->getOpExecType(extypesize);

  if (extypesize == kernel.numEltPerGRF<Type_UB>() / 2 &&
      inst->opcode() != G4_mov) {
    fixPackedSource(i, bb);
    extype = inst->getOpExecType(extypesize);
  }

  // fixes opernds including
  // swapping sel,
  fixOpnds(i, bb, extype);

  extype = inst->getOpExecType(extypesize);
  if (inst->getDst() && !(inst->isSend()) && !(inst->isRawMov())) {
    if (extypesize < (int)kernel.numEltPerGRF<Type_UB>() / 2) {
      uint32_t dst_elsize = inst->getDst()->getTypeSize();
      if (dst_elsize < (unsigned int)extypesize ||
          // indirect float type needs to be handled as well.
          // See fixDstAlignment for detail
          ((extype == Type_F || extype == Type_HF) &&
           inst->getDst()->getRegAccess() != Direct)) {
        if (fixDstAlignment(i, bb, extype, dst_elsize)) {
          insertedInst = true;
        }
      }
    }

    // There are vector immediate source operands.
    if ((*i)->hasVectImm()) {
      if ((insertedInst = fixDstAlignmentWithVectorImm(i, bb))) {
        // Recompute the execution type size if there is some change.
        // This allows fixDstAlignment to fix possible conformity issues.
        extype = inst->getOpExecType(extypesize);
        uint32_t dst_elsize = inst->getDst()->getTypeSize();
        if (dst_elsize < unsigned(extypesize)) {
          if (fixDstAlignment(i, bb, extype, dst_elsize)) {
            insertedInst = true;
          }
        }
      }
    }
  }

  return insertedInst;
}

// check Rule 2H
// VertStride must be used to cross GRF register boundaries. This rule implies
// that elements within a 'Width' cannot cross GRF boundaries. This is a
// separate function from fixSrcRegion because we may need to split the
// instruction to satisfy this rule
bool HWConformity::checkSrcCrossGRF(INST_LIST_ITER &iter, G4_BB *bb) {
  G4_INST *inst = *iter;
  for (int i = 0, numSrc = inst->getNumSrc(); i < numSrc; i++) {
    if (inst->getSrc(i) && inst->getSrc(i)->isSrcRegRegion()) {
      G4_SrcRegRegion *src = inst->getSrc(i)->asSrcRegRegion();
      bool widthCrossingGRF = false;
      const RegionDesc *srcRegion = src->getRegion();
      uint16_t vs = srcRegion->vertStride, wd = srcRegion->width,
               hs = srcRegion->horzStride;
      uint8_t exSize = inst->getExecSize();
      if (src->getRegAccess() == Direct && src->crossGRF(builder)) {
        int elementSize = src->getTypeSize();
        int startOffset = src->getLeftBound() % kernel.numEltPerGRF<Type_UB>();
        for (int row = 0; row < exSize / wd; row++) {
          int rowOffset = (startOffset + row * vs * elementSize) %
                          kernel.numEltPerGRF<Type_UB>();
          if (rowOffset + (wd - 1) * hs * elementSize >=
              (int)kernel.numEltPerGRF<Type_UB>()) {
            widthCrossingGRF = true;
            break;
          }
        }
      } else if (src->getRegAccess() == IndirGRF) {
        widthCrossingGRF = wd > 1 && hs != 0;
      }

      auto doSplit = [&](bool canCrossGRF) -> void {
        if (inst->usesFlag() ||
            (!bb->isAllLaneActive() && !inst->isWriteEnableInst())) {
          // splitting may be unsafe, insert a move then split the move
          G4_Operand *newSrc =
              insertMovBefore(iter, i, inst->getSrc(i)->getType(), bb);
          inst->setSrc(newSrc, i);
          auto movIter = iter;
          --movIter;
          splitInstruction(movIter, bb, false, 0, false, canCrossGRF);
        } else {
          splitInstruction(iter, bb, false, 0, false, canCrossGRF);
        }
      };

      if (widthCrossingGRF) {
        uint16_t stride = 0;
        if (srcRegion->isSingleStride(exSize, stride)) {
          // replace <v;w,h> with <h;1,0>
          src->setRegion(builder, builder.createRegionDesc(stride, 1, 0), true);
        } else {
          doSplit(true);
          return true;
        }
      } else if (kernel.getKernelType() == VISA_CM &&
                 builder.no64bitRegioning() && src->getTypeSize() == 8) {
        // for CM, split non-scalar, non-contiguous source that cross GRF as HW
        // conformity may be not equipped to deal with them later
        const RegionDesc *region = src->getRegion();
        if (!region->isScalar() && !region->isContiguous(inst->getExecSize()) &&
            src->crossGRF(builder)) {
          doSplit(false);
          return true;
        }
      }
    }
  }

  return false;
}

void HWConformity::fixInstExecSize(G4_BB *bb) {
#ifdef _DEBUG
  verifyG4Kernel(kernel, Optimizer::PI_HWConformityChk, false);
#endif

  INST_LIST_ITER i = bb->begin();
  INST_LIST_ITER next_iter = i;

  for (; i != bb->end(); i = next_iter) {
    next_iter++;
    G4_INST *inst = *i;
    G4_opcode opcode = inst->opcode();
    if (opcode == G4_nop || opcode == G4_label || inst->isSend() ||
        inst->isDpas()) {
      continue;
    }

    if (reduceExecSize(i, bb)) {
      next_iter = i;
      next_iter++;
    }
  }
}
// split CISA instructions to follow Gen register region restriction
// splitOp returns true if inst is split into more than instructions
bool HWConformity::reduceExecSize(INST_LIST_ITER iter, G4_BB *bb) {
  G4_INST *inst = *iter;
  // Madw can't be split in any pass except for fixMadwInst as it will cause the
  // dst(SOA layout) unexpected. For example:
  //    madw (M1, 16) dst(0,0)<1> src0(0,0)<1;1,0> 0x38:ud 0x0:ud
  // If split here, then low result is in dst(0,0) and dst(2,0), and high result
  // is in dst(1,0) and dst(3,0)
  //    madw (M1, 8) dst(0,0)<1> src0(0,0)<1;1,0> 0x38:ud 0x0:ud
  //    madw (M8, 8) dst(2,0)<1> src0(1,0)<1;1,0> 0x38:ud 0x0:ud
  // But expected dst is low result is in dst(0,0) and dst(1,0) and high result
  // is in dst(2,0) and dst(3,0)
  if (!inst || inst->isSend() || inst->getExecSize() == 1 ||
      inst->opcode() == G4_madw) {
    return false;
  }

  bool insertMOV = false;

  G4_DstRegRegion *dst = inst->getDst();
  uint8_t minExSize = checkMinExecSize(inst->opcode());

  bool useAcc =
      (inst->hasACCSrc() || (dst && dst->isAccReg()) || inst->getImplAccDst());

  // TODO pre-processing of replicate region, VxH or all indirect sources?
  bool nullDst = inst->hasNULLDst();
  bool packedByteDst = false;
  if (!nullDst && dst) {
    packedByteDst = IS_BTYPE(dst->getType()) && (dst->getHorzStride() == 1);
  }

  unsigned char execSize = inst->getExecSize();
  bool splitOp = false, goodOneGRFDst = false;
  bool crossGRFDst = dst && dst->isCrossGRFDst(builder);
  bool goodTwoGRFDst = false;
  // for all platforms, if execution size is 8 or less and the destination
  // register is 2, flag updates are not supported.
  bool specialCondForComprInst =
      (execSize < 8 && dst && dst->getHorzStride() != 1 && inst->getCondMod() &&
       inst->opcode() != G4_sel);

  TARGET_PLATFORM genX = builder.getPlatform();

  // rules specific to math instructions
  // INT DIV function does not support SIMD16
  if (inst->isMath() && inst->asMathInst()->isMathIntDiv() && execSize == 16) {
    return reduceExecSizeForMath(iter, bb);
  }

  if (genX >= GENX_SKL) {
    // SKL removes rules for GRF alignments, so we don't have to check whether
    // the dst or src is evenly split anymore This means that any subreg of
    // source can move to any subreg of dst
    // FIXME: From the comments it seems we still have to handle
    // "specialCondForComprInst"
    return checkSrcCrossGRF(iter, bb);
  }

  // various variables needed for instruction splitting, for some reason
  G4_opcode op = inst->opcode();
  G4_Type instExecType = inst->getExecType();
  bool oneGRFSrc[3] = {false, false, false};
  bool twoGRFSrc[3] = {false, false, false};
  bool badTwoGRFSrc[3] = {false, false, false};
  bool evenTwoGRFSrc[3] = {false, false, false};
  bool fullTwoGRFSrc[3] = {false, false, false};
  bool hasBadTwoGRFSrc = false;
  bool compOpt = false,
       forceEvenSplit =
           (execSize == 32 && inst->opcode() == G4_sel && inst->getCondMod()) ||
           packedByteDst;
  uint8_t numInFirstMov = 0;
  bool useFlag = inst->getPredicate() || inst->getCondMod() ||
                 (!bb->isAllLaneActive() && !inst->isWriteEnableInst());
  bool evenSplitDst = false;

  // separate the checks for BDW to make it more maintainable
  // For CM use pre-BDW region rules due to HW bugs.
  if (kernel.getKernelType() != VISA_CM &&
      (genX == GENX_BDW || genX == GENX_CHV)) {
    // for BDW we check the following rules:
    // Rule 3D
    // [DevBDW/DevCHV]: When an instruction has a source region spanning two
    // registers and a destination regioning contained in one register, one of
    // the following must be true: The destination region is entirely contained
    // in the lower Oword of a register. The destination region is entirely
    // contained in the upper Oword of a register. The destination elements are
    // evenly split between the two OWords of a register AND evenly split
    // between the two source registers. Rule 3G [DevBDW]: When destination
    // spans two registers, the source may be one or two registers. The
    // destination elements must be evenly split between the two registers.

    bool mayUseIntAcc = op == G4_pseudo_sada2;
    if (crossGRFDst) {
      // rule 3G
      goodTwoGRFDst = inst->goodTwoGRFDst(evenSplitDst) &&
                      !specialCondForComprInst && !mayUseIntAcc;
      splitOp = !goodTwoGRFDst;
    } else {
      // rule 3D
      G4_Operand *srcs[3];
      uint8_t eleInFirstGRF[3];

      for (int i = 0; i < inst->getNumSrc(); i++) {
        srcs[i] = inst->getSrc(i);
        if (srcs[i] && srcs[i]->isSrcRegRegion()) {
          bool indirectSrc =
              srcs[i]->asSrcRegRegion()->getRegAccess() != Direct;
          if (!indirectSrc && srcs[i]->asSrcRegRegion()->isScalar()) {
            continue;
          }

          if (inst->opcode() == G4_pln && i == 1) {
            // src1 for plane may touch multiple GRFs as there's a hidden source
            continue;
          }

          if (srcs[i]->crossGRF(builder)) {
            twoGRFSrc[i] = true;

            if (!nullDst && dst) {
              // check if dst can be entirely contained in one oword
              int dstRegionSize =
                  dst->getRightBound() - dst->getLeftBound() + 1;

              if (dstRegionSize <= 16) {
                // see if we can make the dst fit in one oword
                unsigned short dstOffset = 0;
                bool dstOwordAligned = false;
                int dstAlign = Round_Up_Pow2(dstRegionSize);
                dstOwordAligned =
                    builder.tryToAlignOperand(dst, dstOffset, dstAlign);
                if (!dstOwordAligned) {
                  // If we can align dst to its size, it must fit in one OWord
                  // if we can't, it may still be in OWord (e.g., for size < 16)
                  dstOffset %= kernel.numEltPerGRF<Type_UB>();
                  bool fitInOword =
                      !(dstOffset < 16 && (dstOffset + dstRegionSize) > 16);
                  if (!fitInOword) {
                    // technically if dst and src are both evenly split the
                    // instruction is still ok, but this case should be rare so
                    // we ignore it
                    G4_DstRegRegion *newDst =
                        insertMovAfter(iter, dst, dst->getType(), bb);
                    [[maybe_unused]] bool alignTmpDst =
                        builder.tryToAlignOperand(newDst, dstOffset, 16);
                    vISA_ASSERT(alignTmpDst,
                                 "must be able to oword align tmp dst");
                    inst->setDest(newDst);
                    return true;
                  }
                }
              } else {
                // dst does not fit in one oword, but is guaranteed to be evenly
                // split (it does not cross GRF). check if src is evenly split
                // across the two GRFs
                bool sameSubregOff, vertCrossGRF, contRegion;
                evenTwoGRFSrc[i] =
                    srcs[i]->asSrcRegRegion()->evenlySplitCrossGRF(
                        builder, execSize, sameSubregOff, vertCrossGRF,
                        contRegion, eleInFirstGRF[i]);
                bool coverTwoGRF =
                    srcs[i]->asSrcRegRegion()->coverTwoGRF(builder);
                const RegionDesc *rd = srcs[i]->asSrcRegRegion()->getRegion();
                uint16_t stride = 0;
                fullTwoGRFSrc[i] =
                    coverTwoGRF &&
                    rd->isSingleStride(inst->getExecSize(), stride) &&
                    (stride == 1);

                if (!evenTwoGRFSrc[i]) {
                  // evenly split would be the best approach here, but
                  // unfortunately we can't do it if the instruction is
                  // predicated
                  splitOp = true;
                  // compensation OPT
                  if (!forceEvenSplit && !hasBadTwoGRFSrc && minExSize == 1 &&
                      goodOneGRFDst && contRegion &&
                      eleInFirstGRF[i] > (execSize >> 1)) {
                    if (!useFlag && ((!compOpt && numInFirstMov == 0) ||
                                     numInFirstMov == eleInFirstGRF[i])) {
                      compOpt = true;
                      numInFirstMov = eleInFirstGRF[i];
                    } else {
                      compOpt = false;
                      hasBadTwoGRFSrc = true;
                      badTwoGRFSrc[i] = true;
                    }
                  } else {
                    hasBadTwoGRFSrc = true;
                    badTwoGRFSrc[i] = true;
                  }
                }
              }
            }
            // nothing needs to be done when dst is null???
          }
        }
      }
    }

    // the only reason for split is due to 32-bit flag
    // split inst into two SIMD 16 instructions
    if (!splitOp && execSize == 32 &&
        (inst->getPredicate() || inst->getCondMod())) {
      if (forceEvenSplit) {
        splitSIMD32Inst(iter, bb);
        return insertMOV;
      }
    }
  } else {
    // pre-BDW checks

    // Check if the instruction will use int ACC later. if yes, compressed
    // instruction is split into 2 one-GRF instructions.

    bool mayUseIntAcc = op == G4_pseudo_sada2 ||
                        (op == G4_mul && IS_DTYPE(inst->getSrc(0)->getType()) &&
                         IS_DTYPE(inst->getSrc(1)->getType()));

    if (crossGRFDst) {
      // rule 3D
      goodTwoGRFDst = inst->goodTwoGRFDst(evenSplitDst) &&
                      !specialCondForComprInst && !mayUseIntAcc;
      splitOp = !goodTwoGRFDst;
    }

    G4_Operand *srcs[3];
    uint8_t eleInFirstGRF[3];
    for (int i = 0; i < inst->getNumSrc(); i++) {
      srcs[i] = inst->getSrc(i);

      if (srcs[i] && srcs[i]->isSrcRegRegion() &&
          !(inst->opcode() == G4_math && i == 1 && srcs[i]->isNullReg())) {
        bool indirectSrc =
            (srcs[i]->isSrcRegRegion() &&
             srcs[i]->asSrcRegRegion()->getRegAccess() != Direct);

        if (!indirectSrc && srcs[i]->asSrcRegRegion()->isScalar()) {
          continue;
        }
        if (inst->opcode() == G4_pln && i == 1) {
          continue;
        }

        if (crossGRFDst && indirectSrc) {
          // Assumption: all indirect operand follow GenX requirement (no
          // cross-GRF indexing ...) pre_BDW rule 6D: When a Vx1 or a VxH
          // addressing mode is used on src0, the destination must use ONLY one
          // register. Vx1 is not handled now. only vxh is considered here if
          // (srcs[i]->asSrcRegRegion()->getRegion()->isRegionWH())
          { splitOp = true; }
        } else if (srcs[i]->crossGRF(builder)) {
          twoGRFSrc[i] = true;
          bool sameSubregOff, vertCrossGRF, contRegion;
          evenTwoGRFSrc[i] = srcs[i]->asSrcRegRegion()->evenlySplitCrossGRF(
              builder, execSize, sameSubregOff, vertCrossGRF, contRegion,
              eleInFirstGRF[i]);
          bool coverTwoGRF = srcs[i]->asSrcRegRegion()->coverTwoGRF(builder);
          const RegionDesc *rd = srcs[i]->asSrcRegRegion()->getRegion();
          uint16_t stride = 0;
          fullTwoGRFSrc[i] = coverTwoGRF &&
                             rd->isSingleStride(inst->getExecSize(), stride) &&
                             (stride == 1);

          if (dst && !crossGRFDst) {
            // destination requirements are:
            // -- The destination region is entirely contained in the lower
            // OWord of a register.
            // -- The destination region is entirely contained in the upper
            // OWord of a register.
            // -- The destination elements are evenly split between the two
            // OWords of a register.
            int dstRegionSize = dst->getRightBound() - dst->getLeftBound() + 1;

            // round up dst region size to next power of two
            int dstAlign = Round_Up_Pow2(dstRegionSize);

            bool dstOwordAligned = false;
            dstOwordAligned = builder.tryToAlignOperand(dst, dstAlign);
            if (dstOwordAligned) {
              // If we can align dst to its size, it must fit in one OWord
              goodOneGRFDst = true;
            } else {
              // if we can't, it may still be in one OWord or evenly split
              goodOneGRFDst = dst->goodOneGRFDst(builder, execSize);
            }
          }

          // region can be fixed later in fixCompressedInst().
          // rule 3E and 3F
          // 2-GRF src should follow below implicit rules, no matter the dst
          // size: pre-BDW
          // 1. Data must be evenly split between source registers.
          // 2. Same subregister number in the two GRFs(occupy whole two GRFs)
          // if dst is two-GRF.
          if (!evenTwoGRFSrc[i] ||
              (((goodTwoGRFDst || (goodOneGRFDst && !contRegion)) &&
                !sameSubregOff) ||
               (goodTwoGRFDst && IS_WTYPE(srcs[i]->getType()) &&
                !(srcs[i]->asSrcRegRegion()->checkGRFAlign(builder) &&
                  coverTwoGRF)))) {
            splitOp = true;
            // compensation OPT
            if (!forceEvenSplit && !hasBadTwoGRFSrc && minExSize == 1 &&
                goodOneGRFDst && contRegion &&
                eleInFirstGRF[i] > (execSize >> 1)) {
              if (!useFlag && ((!compOpt && numInFirstMov == 0) ||
                               numInFirstMov == eleInFirstGRF[i])) {
                compOpt = true;
                numInFirstMov = eleInFirstGRF[i];
              } else {
                compOpt = false;
                hasBadTwoGRFSrc = true;
                badTwoGRFSrc[i] = true;
              }
            } else {
              hasBadTwoGRFSrc = true;
              badTwoGRFSrc[i] = true;
            }
          }
          // rule 3C and 3D
          // mul (4) r8.3<1>:f r2.3<4;4,1>:f r31.0<8;2,4>:f {Align1}
          else if (dst && !crossGRFDst && !goodOneGRFDst) {
            splitOp = true;
          }
        }
      }
    }

    // the only reason for split is due to 32-bit flag
    // split inst into two SIMD 16 instructions
    if (!splitOp && execSize == 32 &&
        (packedByteDst || (inst->getPredicate() || inst->getCondMod()))) {
      if (forceEvenSplit) {
        // FIXME: try to use evenlySplitInst() instead.
        splitSIMD32Inst(iter, bb);
        return insertMOV;
      }
    }

    // You will need to do this ONLY when destination spans 2 registers, src1 is
    // a word or byte and you expect channels to be turned off !! currrently for
    // instruction with pred or emask on pre-BDW
    bool specialCondForShootDown =
        (dst && goodTwoGRFDst &&
         (inst->getPredicate() ||
          (!bb->isAllLaneActive() && !inst->isWriteEnableInst())) &&
         oneGRFSrc[1] &&
         (IS_BTYPE(srcs[1]->getType()) || IS_WTYPE(srcs[1]->getType())));
    if (specialCondForShootDown) {
      splitOp = true;
    }
  }

  if (!splitOp) {
    return insertMOV;
  }

  vISA_ASSERT((inst->opcode() != G4_smov),
               "Error in splitting smov instruction");

  // split instruction like:
  // mad (8) V24(2,0)<2> V20(2,0)<16;8,2> V20(2,0)<16;8,2> V23(2,0)<16;8,2>
  if (splitOp && crossGRFDst && evenSplitDst && !hasBadTwoGRFSrc &&
      (execSize >= 16 || !useFlag)) {
    if (minExSize == 1 || execSize > minExSize) {
      evenlySplitInst(iter, bb);
      return insertMOV;
    }
  }

  // For inst with pred, condMod, or with mask in SIMDCF BB, we insert MOVs with
  // nomask for src/dst to avoid instruction splitting. inserted MOVs may be
  // split into multiple instructions. ATTN: We do not include sel here because
  // the condMod generated by sel is never used.
  if (useFlag && !(inst->opcode() == G4_sel && !(inst->getPredicate()) &&
                   inst->getCondMod())) {
    // if there is predicate or cond modifier, we keep the original instruction
    // and perform splitting on new MOV instructions.
    if (!nullDst && !crossGRFDst && !goodOneGRFDst) {
      // try to move 2-GRF src into 1GRF tmp to avoid splitting.
      // this is unnecessary in non-SIMDCF/nonPred/nonCondMod cases because we
      // can do compensation.
      for (int i = 0; i < inst->getNumSrc(); i++) {
        if (twoGRFSrc[i] && !fullTwoGRFSrc[i]) {
          moveSrcToGRF(iter, i, 1, bb);
          twoGRFSrc[i] = false;
          badTwoGRFSrc[i] = false;
          INST_LIST_ITER tmpIter = iter;
          tmpIter--;
          reduceExecSize(tmpIter, bb);
        }
      }
      if (!fullTwoGRFSrc[0] && !fullTwoGRFSrc[1] && !fullTwoGRFSrc[2]) {
        return insertMOV;
      }
    }
    // FIXME: another option is that if original exec size is 16 and will be
    // split into
    //  two simd8, we can use quarter control in some cases.

    if (!nullDst && dst &&
        ((!crossGRFDst && !goodOneGRFDst) || (crossGRFDst && !goodTwoGRFDst) ||
         (goodTwoGRFDst && specialCondForComprInst))) {
      // TODO: NULL dst
      // use temp dst.
      // case 1: SIMD CF
      // mov (16) r3.5<1>:b r1.0<8;8,1>:d
      // ==>
      // mov (16) r6.0<2>:b r3.5<16;16,1>:b {nomask}
      // mov (16) r6.0<2>:b r1.0<8;8,1>:d  -- other dst alignment fix will take
      // care of dst.
      // mov (16) r3.5<1>:b r6.0<32;16,2>:b {nomask}

      // case 2, no SIMDCF
      // (f0.0) mov (16) r3.5<1>:b r1.0<8;8,1>:d
      // ==>
      // mov (16) r6.0<2>:b r1.0<8;8,1>:d  -- other dst alignment fix will take
      // care of dst.
      // (f0.0) mov (16) r3.5<1>:b r6.0<32;16,2>:b

      uint8_t scale = TypeSize(instExecType) / dst->getTypeSize();

      if (scale > 1 && TypeSize(instExecType) * (unsigned)execSize >
                           kernel.numEltPerGRF<Type_UB>()) {
        scale = kernel.numEltPerGRF<Type_UB>() / dst->getTypeSize() / execSize;
      } else if (scale == 0) {
        scale = 1;
      }

      // can't split if inst is in SIMD flow and is not NoMask, or the inst has
      // predicate Have to introduce a temp that supports splitting instead
      if ((!bb->isAllLaneActive() && !inst->isWriteEnableInst()) ||
          inst->getPredicate()) {
        saveDst(iter, scale, bb);
        INST_LIST_ITER tmpIter = iter;
        tmpIter--;
        // Fix up the move to load dst.  We can split the move instruction as it
        // is NoMask
        reduceExecSize(tmpIter, bb);

        // source may also be bad, so we have to call reduceExecSize() on iter
        // again
        reduceExecSize(iter, bb);

        // generate MOV after inst
        // if the dst is bad, it will be fixed by the next call to
        // reduceExecSize()
        restoreDst(iter, dst, bb);
      } else {
        insertMovAfter(iter, scale, bb);
      }
      goodOneGRFDst = true;
      goodTwoGRFDst = true;
      crossGRFDst = false;
      insertMOV = true;
    }

    removeBadSrc(iter, bb, crossGRFDst, oneGRFSrc, badTwoGRFSrc);
    return insertMOV;
  }

  if (!nullDst && dst && !crossGRFDst && !goodOneGRFDst && !hasBadTwoGRFSrc) {
    // insert a temp dst and a MOV
    // example:
    // add (8) r5.3<1>:w r2.0<16;8,2>:w 0x1:w
    // ==>
    // add (8) r6.0<1>:w r2.0<8;8,1>:d 0x1:w
    // mov (8) r5.3<1>:b r6.0<8;8,1>:w
    // In some cases splitting the instruction generates the same number of
    // instruction without dependency, but needs more analysis.
    inst->setDest(insertMovAfter(iter, dst, dst->getType(), bb));
    return true;
  }

  // only two kinds of instruction use ACC operands:
  // 1. instructions generated in ARCTAN intrinsic translation.
  // they do not need splitting
  // 2. instructions generated in MAC opt. there is a check to make
  // sure only evenly splitting will happen to them.
  if (useAcc) {
    evenlySplitInst(iter, bb);
    return insertMOV;
  }
  // split the instruction into a list of instructions
  splitInstruction(iter, bb, compOpt, numInFirstMov, false, true);
  return true;
}

// split a SIMD32 inst into two SIMD16.
// there is predicate/conditional modifier used in this inst.
//
// Result:
//    Inst refered to by 'iter' is split into two simd16 instrcutions.
//    One is inserted right before 'iter', the other is to reuse 'iter'.
// And the caller of this function can access two new instructions via '--iter'
// and 'iter' !
void HWConformity::splitSIMD32Inst(INST_LIST_ITER iter, G4_BB *bb) {
  G4_INST *inst = *iter;
  G4_opcode op = inst->opcode();
  G4_Operand *srcs[3] = {nullptr};
  int numSrc = inst->getNumSrc();

  // check dst/src dependency
  checkSrcDstOverlap(iter, bb, false);
  for (int i = 0; i < numSrc; i++) {
    srcs[i] = inst->getSrc(i);
  }

  // compute max exeuction size.
  // boundary is GRF-boundary and HS change, but for Dst, elements should be
  // symetric if half-GRF boundary is crossed.
  G4_DstRegRegion *dst = inst->getDst();
  vASSERT(nullptr != dst);
  bool nullDst = inst->hasNULLDst();
  G4_ExecSize instExSize = inst->getExecSize(),
              currExSize = G4_ExecSize(instExSize / 2);
  for (int i = 0; i < instExSize; i += currExSize) {
    // create new Oprands. Acc should not be split since we generate it in
    // jitter and can control this. create new condMod and predicate
    G4_CondMod *newCondMod = inst->getCondMod();
    if (newCondMod) {
      newCondMod = builder.createCondMod(newCondMod->getMod(),
                                         newCondMod->getBase(), i == 0 ? 0 : 1);
    }

    G4_Predicate *newPredOpnd = inst->getPredicate();
    if (newPredOpnd) {
      newPredOpnd = builder.createPredicate(
          newPredOpnd->getState(), newPredOpnd->getBase(), i == 0 ? 0 : 1,
          newPredOpnd->getControl());
    }

    G4_DstRegRegion *newDst;
    if (!nullDst) {
      newDst = builder.createSubDstOperand(dst, (uint16_t)i, currExSize);
    } else {
      newDst = dst;
    }
    // generate new inst
    G4_INST *newInst;
    if ((i + currExSize) < instExSize) {
      // 1st simd16 (M0)
      newInst = builder.makeSplittingInst(inst, currExSize);
      newInst->setDest(newDst);
      newInst->setPredicate(newPredOpnd);
      newInst->setCondMod(newCondMod);
      bb->insertBefore(iter, newInst);
    } else {
      // 2nd simd16 (M16). reuse the original inst and may need to reset mask
      // offset.
      if (!inst->isWriteEnableInst() || newPredOpnd || newCondMod) {
        inst->setMaskOption(InstOpt_M16);
      }
      newInst = inst;
      newInst->setExecSize(currExSize);
      newInst->setDest(newDst);
      newInst->setPredicate(newPredOpnd);
      newInst->setCondMod(newCondMod);
    }

    for (int j = 0; j < numSrc; j++) {
      if (srcs[j]) {
        // src1 for single source math should be arc reg null.
        if (srcs[j]->isImm() ||
            (inst->opcode() == G4_math && j == 1 && srcs[j]->isNullReg())) {
          newInst->setSrc(srcs[j], j);
        } else if (srcs[j]->asSrcRegRegion()->isScalar() ||
                   (j == 0 && op == G4_line)) {
          newInst->setSrc(builder.duplicateOperand(srcs[j]), j);
        } else {
          newInst->setSrc(
              builder.createSubSrcOperand(
                  srcs[j]->asSrcRegRegion(), (uint16_t)i, currExSize,
                  (uint8_t)(srcs[j]->asSrcRegRegion()->getRegion()->vertStride),
                  (uint8_t)(srcs[j]->asSrcRegRegion()->getRegion()->width)),
              j);
        }
      }
    }

    // maintain def-use chain
    if (newInst == inst) {
      newInst->trimDefInstList();
    } else {
      // Defs (uses) of this new instruction will be a subset of the
      // original instruction's defs (uses).
      inst->copyDefsTo(newInst, true);
      inst->copyUsesTo(newInst, true);
    }
  }
}

void HWConformity::splitInstruction(INST_LIST_ITER iter, G4_BB *bb,
                                    bool compOpt, uint8_t numInFirstMov,
                                    bool rule4_11, bool canSrcCrossGRF) {
  G4_INST *inst = *iter;
  G4_opcode op = inst->opcode();
  G4_Operand *srcs[3] = {nullptr};

  // check dst/src dependency
  checkSrcDstOverlap(iter, bb, compOpt);

  int numSrcs = inst->getNumSrc();

  for (int i = 0; i < numSrcs; i++) {
    srcs[i] = inst->getSrc(i);
  }

  uint8_t minExSize = checkMinExecSize(op);
  // compute max exeuction size.
  // boundary is GRF-boundary and HS change, but for Dst, elements should be
  // symetric if half-GRF boundary is crossed.
  G4_DstRegRegion *dst = inst->getDst();
  bool nullDst = inst->hasNULLDst();
  G4_ExecSize instExSize = inst->getExecSize();
  G4_ExecSize currExSize;
  uint16_t vs[3] = {0}, wd[3] = {0};

  G4_Predicate *instPred = inst->getPredicate();

  // first, produce mask if needed
  // mov (16) r2.0<1>:uw 0:uw {Align1, NoMask}   // 0:uw
  // mov (16) r2.0<1>:uw 0x1:uw {Align1}   // 1:uw
  // this part is currently not used since we do not split inst with predicate
  // or emask
  bool isSIMDCFInst = !bb->isAllLaneActive() && !inst->isWriteEnableInst();
  G4_Declare *maskDcl = NULL;
  if (instPred || isSIMDCFInst) {
    maskDcl = builder.createTempVar(instExSize, Type_UW, Eight_Word);
    G4_DstRegRegion *tmpMaskOpnd =
        builder.createDst(maskDcl->getRegVar(), 0, 0, 1, Type_UW);

    G4_INST *firstMov = builder.createMov(instExSize, tmpMaskOpnd,
                                          builder.createImm(0, Type_UW),
                                          inst->getOption(), false);

    G4_Predicate *pred = builder.duplicateOperand(inst->getPredicate());
    auto movInst = builder.createMov(instExSize, tmpMaskOpnd,
                                     builder.createImm(1, Type_UW),
                                     inst->getOption(), false);
    movInst->setPredicate(pred);

    if (isSIMDCFInst) {
      firstMov->setNoMask(true);
    }
  }

  bool needsMaskOffset =
      instPred || isSIMDCFInst || inst->getCondMod() != nullptr;

  for (uint8_t i = 0; i < instExSize; i += currExSize) {
    if (compOpt && i == 0) {
      currExSize = G4_ExecSize(numInFirstMov);
      G4_INST *newInst = builder.makeSplittingInst(inst, instExSize);
      newInst->setDest(builder.duplicateOperand(inst->getDst()));
      newInst->setPredicate(builder.duplicateOperand(inst->getPredicate()));
      newInst->setCondMod(builder.duplicateOperand(inst->getCondMod()));
      for (int j = 0; j < inst->getNumSrc(); j++) {
        newInst->setSrc(builder.duplicateOperand(srcs[j]), j);
      }
      // update def-use chain
      inst->copyDefsTo(newInst, true);
      inst->copyDefsTo(newInst, true);
      bb->insertBefore(iter, newInst);
      continue;
    }

    // this stores the max allowed exec size for each operand (0 -- dst, 1 --
    // src0, and so on)
    uint8_t opndExSize[4] = {0, 0, 0, 0};
    currExSize = G4_ExecSize(roundDownPow2(instExSize - i));

    bool crossGRFsrc = false;
    for (int j = 0; j < numSrcs; j++) {
      if (!srcs[j] || !srcs[j]->isSrcRegRegion() || srcs[j]->isNullReg() ||
          (j == 0 && op == G4_line)) {
        opndExSize[j + 1] = currExSize;
        continue;
      }
      bool twoGRFsrc = false;
      opndExSize[j + 1] = srcs[j]->asSrcRegRegion()->getMaxExecSize(
          builder, i, currExSize, canSrcCrossGRF, vs[j], wd[j], twoGRFsrc);

      if (opndExSize[j + 1] > 8 && rule4_11) {
        opndExSize[j + 1] = 8;
      }

      crossGRFsrc |= twoGRFsrc;
      if (minExSize == 1) {
        currExSize = G4_ExecSize(opndExSize[j + 1]);
      }
    }

    vASSERT(dst);
    if (!nullDst) {
      opndExSize[0] = dst->getMaxExecSize(builder, i, currExSize, crossGRFsrc);

      if (opndExSize[0] > 8 && rule4_11)
        opndExSize[0] = 8;
    } else {
      // dst essentially does not affect the splitting decision
      opndExSize[0] = currExSize;
    }

    if (minExSize == 1) {
      currExSize = G4_ExecSize(opndExSize[0]);
    }

    bool needMov = false;
    if (minExSize > 1) {
      // find minimal execsize. if it is not less than minExSize, use it
      // to avoid dependency
      // FIXME: optimize this part by avoiding MOVs
      uint8_t currMinExSize = 64;
      currExSize = G4_ExecSize(0);
      for (int j = 0; j <= numSrcs; j++) {
        // use max possible exsize
        if (opndExSize[j] > currExSize) {
          currExSize = G4_ExecSize(opndExSize[j]);
        }
        if (opndExSize[j] != 0 && opndExSize[j] < currMinExSize) {
          currMinExSize = opndExSize[j];
        }
      }

      if (currMinExSize >= minExSize) {
        currExSize = G4_ExecSize(currMinExSize);
      } else {
        for (int j = 0; j <= numSrcs; j++) {
          if (opndExSize[j] != 0 && opndExSize[j] < currExSize) {
            needMov = true;
          }
        }
      }
    }

    vISA_ASSERT(currExSize != 0,
                 "illegal execution size in instruction splitting");
    // create new Oprands. Acc should not be split since we generate it in
    // jitter and can control this.
    G4_DstRegRegion *newDst =
        !nullDst ? builder.createSubDstOperand(dst, (uint16_t)i, currExSize)
                 : dst;

    // generate new inst
    G4_INST *newInst;
    INST_LIST_ITER newInstIter;
    if ((i + currExSize) < instExSize) {
      newInst = builder.makeSplittingInst(inst, currExSize);
      newInst->setDest(newDst);
      newInst->setPredicate(builder.duplicateOperand(inst->getPredicate()));
      newInst->setCondMod(builder.duplicateOperand(inst->getCondMod()));
      bb->insertBefore(iter, newInst);
      newInstIter = iter;
      newInstIter--;
    } else {
      // reuse the original inst
      newInst = inst;
      newInst->setDest(newDst);
      newInst->setExecSize(currExSize);
      newInstIter = iter;
    }

    for (int j = 0; j < inst->getNumSrc(); j++) {
      if (srcs[j]) {
        // src1 for single source math should be arc reg null.
        if (srcs[j]->isImm() ||
            (inst->opcode() == G4_math && j == 1 && srcs[j]->isNullReg())) {
          newInst->setSrc(srcs[j], j);
        } else if (srcs[j]->asSrcRegRegion()->isScalar() ||
                   (j == 0 && op == G4_line)) {
          newInst->setSrc(builder.duplicateOperand(srcs[j]), j);
        } else {
          if (srcs[j]->isAddrExp()) {
            G4_AddrExp *addExp =
                builder.createAddrExp(srcs[j]->asAddrExp()->getRegVar(),
                                      srcs[j]->asAddrExp()->getOffset(),
                                      srcs[j]->asAddrExp()->getType());
            newInst->setSrc(addExp, j);
          } else {
            uint16_t start = i;
            newInst->setSrc(
                builder.createSubSrcOperand(srcs[j]->asSrcRegRegion(), start,
                                            currExSize, vs[j], wd[j]),
                j);
          }
        }
      }
    }

    if (instExSize == 16 && currExSize == 8 && needsMaskOffset) {
      if (instPred) {
        G4_Predicate *tPred = builder.duplicateOperand(instPred);
        tPred->setInst(newInst);
        newInst->setPredicate(tPred);
      }

      if (newInst->getMaskOffset() == 0) {
        newInst->setMaskOption(i == 0 ? InstOpt_M0 : InstOpt_M8);
      } else {
        newInst->setMaskOption(i == 0 ? InstOpt_M16 : InstOpt_M24);
      }
    }

    // maintain def-use chain
    if (newInst == inst) {
      newInst->trimDefInstList();
    } else {
      inst->copyDefsTo(newInst, /*checked*/ true);
      inst->copyUsesTo(newInst, /*checked*/ true);
    }

    // the following code is to keep minimal execution size for some opcode, for
    // example, DP4 insert mov if needed
    if (needMov) {
      for (int j = 0; j < inst->getNumSrc(); j++) {
        if (opndExSize[j + 1] < currExSize) {
          newInst->setSrc(
              insertMovBefore(newInstIter, j, srcs[j]->getType(), bb), j);
          // reducing exec size for new MOV
          INST_LIST_ITER newMovIter = newInstIter;
          newMovIter--;
          reduceExecSize(newMovIter, bb);
        }
      }
    }
    // dst
    if (needMov && opndExSize[0] < currExSize) {
      (*newInstIter)
          ->setDest(insertMovAfter(newInstIter, inst->getDst(),
                                   inst->getDst()->getType(), bb));
      INST_LIST_ITER newMovIter = newInstIter;
      newMovIter++;
      reduceExecSize(newMovIter, bb);
    }
  }
}

// evenly split an inst into two instructions with half execution size.
// this is used to split a simd16 math into two simd8 before other reducing
// exeuction size actions
//
// This will has two instructions: one is right before "iter", the other is to
// re-use "iter". The caller is safe to use "--iter" and "iter" to refer those
// two instructions.
bool HWConformity::evenlySplitInst(INST_LIST_ITER iter, G4_BB *bb,
                                   bool checkOverlap) {
  G4_INST *inst = *iter;
  G4_opcode op = inst->opcode();
  G4_Operand *srcs[3];
  int origMaskOffset = inst->getMaskOffset();
  bool extraMov = false;
  const int numSrc = inst->getNumSrc();

  // check dst/src dependency
  if (checkOverlap) {
    extraMov = checkSrcDstOverlap(iter, bb, false);
  }

  bool useARF = false;
  for (int i = 0; i < numSrc; i++) {
    srcs[i] = inst->getSrc(i);
  }

  // compute max exeuction size.
  // boundary is GRF-boundary and HS change, but for Dst, elements should be
  // symetric if half-GRF boundary is crossed.

  G4_DstRegRegion *dst = inst->getDst();
  vASSERT(nullptr != dst);
  bool nullDst = inst->hasNULLDst();
  G4_ExecSize instExSize = inst->getExecSize(),
              currExSize = G4_ExecSize(instExSize / 2);

  G4_Predicate *newPred = NULL;
  if (inst->getPredicate()) {
    newPred = inst->getPredicate();
    newPred->splitPred();
  }

  G4_CondMod *newCond = NULL;
  if (inst->getCondMod()) {
    newCond = inst->getCondMod();
    newCond->splitCondMod();
  }

  G4_SrcRegRegion *accSrcRegion = NULL;
  if (inst->getImplAccSrc()) {
    accSrcRegion = inst->getImplAccSrc()->asSrcRegRegion();
  }

  G4_DstRegRegion *accDstRegion = NULL;
  if (inst->getImplAccDst()) {
    accDstRegion = inst->getImplAccDst();
  }

  if (accSrcRegion || accDstRegion || newPred || newCond) {
    useARF = true;
  }

  for (int i = 0; i < instExSize; i += currExSize) {
    // create new Oprands.
    G4_DstRegRegion *newDst;
    if (!nullDst) {
      newDst = builder.createSubDstOperand(dst, (uint16_t)i, currExSize);
    } else {
      newDst = dst;
    }
    // generate new inst
    G4_INST *newInst;
    if ((i + currExSize) < instExSize) {
      newInst = builder.makeSplittingInst(inst, currExSize);
      if (accDstRegion)
        newInst->setImplAccDst(builder.duplicateOperand(accDstRegion));
      if (accSrcRegion)
        newInst->setImplAccSrc(builder.duplicateOperand(accSrcRegion));
      newInst->setDest(newDst);
      newInst->setPredicate(builder.duplicateOperand(newPred));
      newInst->setCondMod(builder.duplicateOperand(newCond));
      newInst->setEvenlySplitInst(true);
      bb->insertBefore(iter, newInst);
    } else {
      // reuse the original inst
      newInst = inst;
      newInst->setExecSize(currExSize);
      newInst->setDest(newDst);
      if (newPred) {
        inst->setPredicate(builder.duplicateOperand(newPred));
      }
      if (newCond) {
        inst->setCondMod(builder.duplicateOperand(newCond));
      }
      if (accSrcRegion) {
        newInst->setImplAccSrc(builder.createSrcRegRegion(*accSrcRegion));
      }
      if (accDstRegion) {
        newInst->setImplAccDst(builder.createDstRegRegion(*accDstRegion));
      }
    }

    for (int j = 0; j < numSrc; j++) {
      if (srcs[j]) {
        if (srcs[j]->isImm() || srcs[j]->isNullReg()) {
          newInst->setSrc(srcs[j], j);
        } else if (srcs[j]->isScalarSrc() || (j == 0 && op == G4_line)) {
          // no need to split, but need to duplicate
          newInst->setSrc(builder.duplicateOperand(srcs[j]), j);
        } else if (op == G4_movi) {
          // we create temp region which is in VxH format
          RegionDesc VxHregionDesc = RegionDesc(UNDEFINED_SHORT, 1, 0);

          auto tempRegion = G4_SrcRegRegion(*srcs[j]->asSrcRegRegion());
          tempRegion.setRegion(builder, &VxHregionDesc);

          newInst->setSrc(builder.createSubSrcOperand(
                              &tempRegion, (uint16_t)i, currExSize,
                              (uint8_t)(tempRegion.getRegion()->vertStride),
                              (uint8_t)(tempRegion.getRegion()->width)),
                          j);

          // restore the original region
          const RegionDesc *rd = builder.getRegionStride1();
          newInst->getSrc(j)->asSrcRegRegion()->setRegion(builder, rd);
        } else {
          newInst->setSrc(
              builder.createSubSrcOperand(
                  srcs[j]->asSrcRegRegion(), (uint16_t)i, currExSize,
                  (uint8_t)(srcs[j]->asSrcRegRegion()->getRegion()->vertStride),
                  (uint8_t)(srcs[j]->asSrcRegRegion()->getRegion()->width)),
              j);
        }
      }
    }

    // set mask
    bool needsMaskOffset =
        useARF || (!bb->isAllLaneActive() && !inst->isWriteEnableInst());
    if (needsMaskOffset) {
      int newMaskOffset = origMaskOffset + (i == 0 ? 0 : currExSize);
      bool nibOk =
          builder.hasNibCtrl() && (inst->getDst()->getTypeSize() == 8 ||
                                   TypeSize(inst->getExecType()) == 8);
      G4_InstOption newMask =
          G4_INST::offsetToMask(currExSize, newMaskOffset, nibOk);
      if (newMask == InstOpt_NoOpt) {
        [[maybe_unused]] bool useMask = inst->getPredicate() || inst->getCondModBase() ||
                       (!bb->isAllLaneActive() && !inst->isWriteEnableInst());
        vISA_ASSERT(!useMask,
                     "no legal emask found for the split instruction");
      } else {
        newInst->setMaskOption(newMask);
      }
    }

    // maintain def-use chain
    if (newInst == inst) {
      newInst->trimDefInstList();
    } else {
      inst->copyDefsTo(newInst, /*checked*/ true);
      inst->copyUsesTo(newInst, /*checked*/ true);
    }
  }

  return extraMov;
}

// this is specifically for math instruction
// assumption: the input math function is a compressed instruction and need
// split
bool HWConformity::reduceExecSizeForMath(INST_LIST_ITER iter, G4_BB *bb) {
  // split the instruction into two first
  evenlySplitInst(iter, bb);
  // fix execution size for each one
  INST_LIST_ITER firstIter = iter;
  firstIter--;
  reduceExecSize(firstIter, bb);
  return reduceExecSize(iter, bb);
}
// check overlap between src and dst
// if overlap exists, insert to MOV to eliminate it
// how about replicate regions?<0;4,1>
bool HWConformity::checkSrcDstOverlap(INST_LIST_ITER iter, G4_BB *bb,
                                      bool compOpt) {
  G4_INST *inst = *iter;
  G4_Operand *srcs[3];
  bool hasOverlap = false;

  for (int i = 0; i < inst->getNumSrc(); i++) {
    srcs[i] = inst->getSrc(i);
  }
  // check dst/src dependency
  // how about replicate regions?<0;4,1>
  if (inst->getDst() && !inst->hasNULLDst()) {
    for (int i = 0; i < inst->getNumSrc(); i++) {
      bool useTmp = false;
      if (srcs[i] &&
          (IS_VINTTYPE(srcs[i]->getType()) || IS_VFTYPE(srcs[i]->getType()))) {
        useTmp = true;
      } else {
        G4_CmpRelation rel = inst->getDst()->compareOperand(srcs[i], builder);
        if (rel != Rel_disjoint) {
          useTmp = (rel != Rel_eq) || compOpt ||
                   srcs[i]->asSrcRegRegion()->getRegion()->isRepeatRegion(
                       inst->getExecSize());
        }
      }
      if (useTmp) {
        // insert mov
        inst->setSrc(
            insertMovBefore(iter, i,
                            G4_Operand::GetNonVectorImmType(srcs[i]->getType()),
                            bb),
            i);
        srcs[i] = inst->getSrc(i);
        // reducing exec size for new MOV
        INST_LIST_ITER newMovIter = iter;
        newMovIter--;
        reduceExecSize(newMovIter, bb);
        hasOverlap = true;
      }
    }
  }

  return hasOverlap;
}

// move source operand to one or two GRF
// tmp dst use the same type as source.
// this MOV does not need further resucing execsize
void HWConformity::moveSrcToGRF(INST_LIST_ITER it, uint32_t srcNum,
                                uint16_t numGRF, G4_BB *bb) {
  G4_INST *inst = *it;
  G4_ExecSize execSize = inst->getExecSize();

  G4_Operand *src = inst->getSrc(srcNum);
  uint32_t srcTypeSize = src->getTypeSize();
  uint16_t dclSize = (kernel.numEltPerGRF<Type_UB>() * numGRF) / srcTypeSize;
  uint16_t hs = dclSize / execSize;
  uint16_t wd = execSize;
  uint16_t vs = hs * wd;
  const RegionDesc *region = builder.createRegionDesc(vs, wd, hs);

  // look up in MOV table to see if there is already inserted MOV for this
  // source.
  G4_INST *def_inst = NULL;
  def_inst = checkSrcDefInst(inst, def_inst, srcNum);

  G4_Type tmpType = G4_Operand::GetNonVectorImmType(src->getType());

  if (def_inst && def_inst->getDst()->getType() == tmpType &&
      (def_inst->getExecSize() == execSize) &&
      def_inst->getDst()->coverGRF(builder, numGRF, execSize) &&
      def_inst->getDst()->checkGRFAlign(builder) &&
      (bb->isAllLaneActive() || def_inst->isWriteEnableInst())) {
    G4_DstRegRegion *existing_def = def_inst->getDst();
    G4_SrcRegRegion *newSrc =
        builder.createSrc(existing_def->getBase(), existing_def->getRegOff(),
                          existing_def->getSubRegOff(), region, src->getType());
    inst->setSrc(newSrc, srcNum);
  }

  G4_Declare *dcl =
      builder.createTempVar(dclSize, src->getType(), builder.getGRFAlign());
  G4_DstRegRegion *dstRegion =
      builder.createDst(dcl->getRegVar(), 0, 0, hs, dcl->getElemType());
  G4_INST *newInst = builder.createMov(
      execSize, dstRegion, src,
      (!bb->isAllLaneActive() ? InstOpt_WriteEnable : InstOpt_NoOpt), false);

  // insert instruction and maintain def-use chain
  bb->insertBefore(it, newInst);
  inst->transferDef(newInst, Gen4_Operand_Number(srcNum + 1), Opnd_src0);
  newInst->addDefUse(inst, Gen4_Operand_Number(srcNum + 1));

  G4_SrcRegRegion *newSrc =
      builder.createSrc(dcl->getRegVar(), 0, 0, region, dcl->getElemType());
  inst->setSrc(newSrc, srcNum);
}

/*
 *  create a new mov instruction and insert it before iter
 *  mov (esize) tmpDst dst (nomask)
 *  add (esize) tmpDst ...
 *  where esize is "inst"'s execution size
 *
 */
void HWConformity::saveDst(INST_LIST_ITER &it, uint8_t stride, G4_BB *bb) {
  G4_INST *inst = *it;
  G4_DstRegRegion *dst = inst->getDst();
  G4_ExecSize execSize = inst->getExecSize();
  G4_Type dstType = dst->getType();
  uint16_t dstWidthBytes = execSize * TypeSize(dstType) * stride;

  G4_SubReg_Align subAlign =
      getDclAlignment(dstWidthBytes, inst, execSize == 1);

  uint32_t numElt = execSize == 1 ? 1 : execSize * stride;
  G4_Declare *dcl = builder.createTempVar(numElt, dstType, subAlign);

  uint16_t hs = dst->getHorzStride();
  const RegionDesc *region =
      builder.createRegionDesc(hs * execSize, execSize, hs);
  G4_SrcRegRegion *srcRegion = builder.createSrc(
      dst->getBase(), dst->getRegOff(), dst->getSubRegOff(), region, dstType);

  G4_DstRegRegion *tmpDstOpnd = builder.createDstRegRegion(dcl, stride);

  unsigned int new_option = inst->getOption();

  G4_INST *newInst =
      builder.createMov(execSize, tmpDstOpnd, srcRegion, new_option, false);
  newInst->setNoMask(true);

  bb->insertBefore(it, newInst);
  inst->setDest(builder.duplicateOperand(tmpDstOpnd));
}

void HWConformity::restoreDst(INST_LIST_ITER &it, G4_DstRegRegion *origDst,
                              G4_BB *bb) {
  G4_INST *inst = *it;
  G4_DstRegRegion *dst = inst->getDst();
  G4_ExecSize execSize = inst->getExecSize();

  uint16_t hs = dst->getHorzStride();
  const RegionDesc *region =
      builder.createRegionDesc(hs * execSize, execSize, hs);
  G4_SrcRegRegion *srcRegion =
      builder.createSrc(dst->getBase(), dst->getRegOff(), dst->getSubRegOff(),
                        region, dst->getType());

  unsigned int new_option = inst->getOption();

  G4_INST *newInst =
      builder.createMov(execSize, origDst, srcRegion, new_option, false);
  newInst->setNoMask(true);

  INST_LIST_ITER iter = it;
  iter++;
  bb->insertBefore(iter, newInst);

  // how about def-use?
  inst->transferUse(newInst);
  inst->addDefUse(newInst, Gen4_Operand_Number::Opnd_src0);
}

/*
 *  create a new mov instruction and insert it after iter
 *  mov (esize) dst tmp:dst_type
 *  where esize is "inst"'s execution size and insert it after "inst"
 *  dst of inst is replaced with the tmp dst using the same type
 */
void HWConformity::insertMovAfter(INST_LIST_ITER &it, uint16_t stride,
                                  G4_BB *bb) {
  G4_INST *inst = *it;
  G4_DstRegRegion *dst = inst->getDst();
  G4_ExecSize execSize = inst->getExecSize();
  G4_Type execType = inst->getExecType(), dstType = dst->getType();
  uint16_t opExecWidthBytes = execSize * TypeSize(execType);
  uint16_t dstWidthBytes = execSize * TypeSize(dstType) * stride;

  G4_SubReg_Align subAlign = getDclAlignment(
      opExecWidthBytes > dstWidthBytes ? opExecWidthBytes : dstWidthBytes, inst,
      execSize == 1);

  G4_Declare *dcl = builder.createTempVar(execSize * stride, dstType, subAlign);

  const RegionDesc *region = builder.createRegionDesc(stride, 1, 0);
  G4_SrcRegRegion *srcRegion = builder.createSrcRegRegion(dcl, region);
  G4_DstRegRegion *tmpDstOpnd = builder.createDstRegRegion(dcl, stride);

  G4_Predicate *pred = NULL;
  if (inst->opcode() != G4_sel) {
    pred = inst->getPredicate();
    inst->setPredicate(NULL);
  }
  unsigned int new_option = inst->getOption();

  G4_INST *newInst =
      builder.createMov(execSize, dst, srcRegion, new_option, false);
  newInst->setPredicate(pred);

  INST_LIST_ITER iter = it;
  iter++;
  bb->insertBefore(iter, newInst);
  // change dst of inst
  inst->setDest(tmpDstOpnd);

  // update propagation info
  if (pred) {
    inst->transferDef(newInst, Opnd_pred, Opnd_pred);
  }

  inst->transferUse(newInst);
  inst->addDefUse(newInst, Opnd_src0);
}

void HWConformity::removeBadSrc(INST_LIST_ITER &iter, G4_BB *bb,
                                bool crossGRFDst, bool oneGRFSrc[3],
                                bool badTwoGRFSrc[3]) {
  G4_INST *inst = *iter;
  G4_Operand *dst = inst->getDst();
  // check source and dst region together
  // get rid of bad two-GRF source
  for (int i = 0; i < inst->getNumSrc(); i++) {

    if (badTwoGRFSrc[i]) {
      if (!crossGRFDst || (dst && IS_DTYPE(dst->getType()) &&
                           IS_WTYPE(inst->getSrc(i)->getType()))) {
        inst->setSrc(insertMovBefore(iter, i, inst->getSrc(i)->getType(), bb),
                     i);
      } else {
        moveSrcToGRF(iter, i, 2, bb);
      }
      badTwoGRFSrc[i] = false;
      INST_LIST_ITER tmpIter = iter;
      tmpIter--;
      reduceExecSize(tmpIter, bb);
    }
  }
}