//===- HexagonPatternsHVX.td - Selection Patterns for HVX --*- tablegen -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//


def SDTVecUnaryOp:
  SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>]>;

def SDTVecBinOp:
  SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>, SDTCisSameAs<1,2>]>;

def SDTHexagonVEXTRACTW: SDTypeProfile<1, 2,
  [SDTCisVT<0, i32>, SDTCisVec<1>, SDTCisVT<2, i32>]>;
def HexagonVEXTRACTW : SDNode<"HexagonISD::VEXTRACTW", SDTHexagonVEXTRACTW>;

def SDTHexagonVINSERTW0: SDTypeProfile<1, 2,
  [SDTCisVec<0>, SDTCisSameAs<0, 1>, SDTCisVT<2, i32>]>;
def HexagonVINSERTW0: SDNode<"HexagonISD::VINSERTW0", SDTHexagonVINSERTW0>;

def HwLen2: SDNodeXForm<imm, [{
  const auto &ST = static_cast<const HexagonSubtarget&>(CurDAG->getSubtarget());
  return CurDAG->getTargetConstant(ST.getVectorLength()/2, SDLoc(N), MVT::i32);
}]>;

def Q2V: OutPatFrag<(ops node:$Qs), (V6_vandqrt $Qs, (A2_tfrsi -1))>;

def Combinev: OutPatFrag<(ops node:$Vs, node:$Vt),
  (REG_SEQUENCE HvxWR, $Vs, vsub_hi, $Vt, vsub_lo)>;

def Combineq: OutPatFrag<(ops node:$Qs, node:$Qt),
  (V6_vandvrt
    (V6_vor
      (V6_vror (V6_vpackeb (V6_vd0), (Q2V $Qs)),
               (A2_tfrsi (HwLen2 (i32 0)))),  // Half the vector length
      (V6_vpackeb (V6_vd0), (Q2V $Qt))),
    (A2_tfrsi -1))>;

def LoVec: OutPatFrag<(ops node:$Vs), (EXTRACT_SUBREG $Vs, vsub_lo)>;
def HiVec: OutPatFrag<(ops node:$Vs), (EXTRACT_SUBREG $Vs, vsub_hi)>;

def HexagonQCAT:       SDNode<"HexagonISD::QCAT",       SDTVecBinOp>;
def HexagonQTRUE:      SDNode<"HexagonISD::QTRUE",      SDTVecLeaf>;
def HexagonQFALSE:     SDNode<"HexagonISD::QFALSE",     SDTVecLeaf>;
def HexagonVPACKL:     SDNode<"HexagonISD::VPACKL",     SDTVecUnaryOp>;
def HexagonVUNPACK:    SDNode<"HexagonISD::VUNPACK",    SDTVecUnaryOp>;
def HexagonVUNPACKU:   SDNode<"HexagonISD::VUNPACKU",   SDTVecUnaryOp>;

def vzero:  PatFrag<(ops), (splat_vector (i32 0))>;
def qtrue:  PatFrag<(ops), (HexagonQTRUE)>;
def qfalse: PatFrag<(ops), (HexagonQFALSE)>;
def qcat:   PatFrag<(ops node:$Qs, node:$Qt),
                    (HexagonQCAT node:$Qs, node:$Qt)>;

def qnot:     PatFrag<(ops node:$Qs), (xor node:$Qs, qtrue)>;
def vpackl:   PatFrag<(ops node:$Vs), (HexagonVPACKL node:$Vs)>;
def vunpack:  PatFrag<(ops node:$Vs), (HexagonVUNPACK node:$Vs)>;
def vunpacku: PatFrag<(ops node:$Vs), (HexagonVUNPACKU node:$Vs)>;

def VSxtb: OutPatFrag<(ops node:$Vs), (V6_vunpackb  $Vs)>;
def VSxth: OutPatFrag<(ops node:$Vs), (V6_vunpackh  $Vs)>;
def VZxtb: OutPatFrag<(ops node:$Vs), (V6_vunpackub $Vs)>;
def VZxth: OutPatFrag<(ops node:$Vs), (V6_vunpackuh $Vs)>;

def IsVecOff : PatLeaf<(i32 imm), [{
  int32_t V = N->getSExtValue();
  int32_t VecSize = HRI->getSpillSize(Hexagon::HvxVRRegClass);
  assert(isPowerOf2_32(VecSize));
  if ((uint32_t(V) & (uint32_t(VecSize)-1)) != 0)
    return false;
  int32_t L = Log2_32(VecSize);
  return isInt<4>(V >> L);
}]>;


def alignedload: PatFrag<(ops node:$a), (load $a), [{
  return isAlignedMemNode(dyn_cast<MemSDNode>(N));
}]>;

def unalignedload: PatFrag<(ops node:$a), (load $a), [{
  return !isAlignedMemNode(dyn_cast<MemSDNode>(N));
}]>;

def alignedstore: PatFrag<(ops node:$v, node:$a), (store $v, $a), [{
  return isAlignedMemNode(dyn_cast<MemSDNode>(N));
}]>;

def unalignedstore: PatFrag<(ops node:$v, node:$a), (store $v, $a), [{
  return !isAlignedMemNode(dyn_cast<MemSDNode>(N));
}]>;


// HVX loads

multiclass HvxLdfi_pat<InstHexagon MI, PatFrag Load, ValueType ResType,
                       PatFrag ImmPred> {
  def: Pat<(ResType (Load (add (i32 AddrFI:$fi), ImmPred:$Off))),
           (MI AddrFI:$fi, imm:$Off)>;
  def: Pat<(ResType (Load (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off))),
           (MI AddrFI:$fi, imm:$Off)>;
  def: Pat<(ResType (Load AddrFI:$fi)), (ResType (MI AddrFI:$fi, 0))>;
}

multiclass HvxLdgi_pat<InstHexagon MI, PatFrag Load, ValueType ResType,
                     PatFrag ImmPred> {
  def: Pat<(ResType (Load (add I32:$Rt, ImmPred:$Off))),
           (MI I32:$Rt, imm:$Off)>;
  def: Pat<(ResType (Load I32:$Rt)),
           (MI I32:$Rt, 0)>;
}

multiclass HvxLdc_pat<InstHexagon MI, PatFrag Load, ValueType ResType> {
  // The HVX selection code for shuffles can generate vector constants.
  // Calling "Select" on the resulting loads from CP fails without these
  // patterns.
  def: Pat<(ResType (Load (HexagonCP tconstpool:$Addr))),
           (MI (A2_tfrsi imm:$Addr), 0)>;
  def: Pat<(ResType (Load (HexagonAtPcrel tconstpool:$Addr))),
           (MI (C4_addipc imm:$Addr), 0)>;
}

multiclass HvxLd_pat<InstHexagon MI, PatFrag Load, ValueType ResType,
                     PatFrag ImmPred> {
  defm: HvxLdfi_pat<MI, Load, ResType, ImmPred>;
  defm: HvxLdgi_pat<MI, Load, ResType, ImmPred>;
  defm: HvxLdc_pat <MI, Load, ResType>;
}

// Aligned loads: everything, plus loads with valignaddr node.
multiclass HvxLda_pat<InstHexagon MI, PatFrag Load, ValueType ResType,
                      PatFrag ImmPred> {
  let AddedComplexity = 50 in {
    def: Pat<(ResType (Load (valignaddr I32:$Rt))),
             (MI I32:$Rt, 0)>;
    def: Pat<(ResType (Load (add (valignaddr I32:$Rt), ImmPred:$Off))),
             (MI I32:$Rt, imm:$Off)>;
  }
  defm: HvxLd_pat<MI, Load, ResType, ImmPred>;
}

let Predicates = [UseHVX] in {
  // alignedload will match a non-temporal load as well, so try non-temporal
  // first.
  defm: HvxLda_pat<V6_vL32b_nt_ai, alignednontemporalload, VecI8,  IsVecOff>;
  defm: HvxLda_pat<V6_vL32b_nt_ai, alignednontemporalload, VecI16, IsVecOff>;
  defm: HvxLda_pat<V6_vL32b_nt_ai, alignednontemporalload, VecI32, IsVecOff>;
  defm: HvxLda_pat<V6_vL32b_ai,               alignedload, VecI8,  IsVecOff>;
  defm: HvxLda_pat<V6_vL32b_ai,               alignedload, VecI16, IsVecOff>;
  defm: HvxLda_pat<V6_vL32b_ai,               alignedload, VecI32, IsVecOff>;

  defm: HvxLd_pat<V6_vL32Ub_ai,             unalignedload, VecI8,  IsVecOff>;
  defm: HvxLd_pat<V6_vL32Ub_ai,             unalignedload, VecI16, IsVecOff>;
  defm: HvxLd_pat<V6_vL32Ub_ai,             unalignedload, VecI32, IsVecOff>;
}


// HVX stores

multiclass HvxStfi_pat<InstHexagon MI, PatFrag Store, PatFrag Value,
                       PatFrag ImmPred> {
  def: Pat<(Store Value:$Vs, (add (i32 AddrFI:$fi), ImmPred:$Off)),
           (MI AddrFI:$fi, imm:$Off, Value:$Vs)>;
  def: Pat<(Store Value:$Vs, (IsOrAdd (i32 AddrFI:$fi), ImmPred:$Off)),
           (MI AddrFI:$fi, imm:$Off, Value:$Vs)>;
  def: Pat<(Store Value:$Vs, AddrFI:$fi),
           (MI AddrFI:$fi, 0, Value:$Vs)>;
}

multiclass HvxStgi_pat<InstHexagon MI, PatFrag Store, PatFrag Value,
                       PatFrag ImmPred> {
  def: Pat<(Store Value:$Vs, (add I32:$Rt, ImmPred:$Off)),
           (MI I32:$Rt, imm:$Off, Value:$Vs)>;
  def: Pat<(Store Value:$Vs, (IsOrAdd I32:$Rt, ImmPred:$Off)),
           (MI I32:$Rt, imm:$Off, Value:$Vs)>;
  def: Pat<(Store Value:$Vs, I32:$Rt),
           (MI I32:$Rt, 0, Value:$Vs)>;
}

multiclass HvxSt_pat<InstHexagon MI, PatFrag Store, PatFrag Value,
                     PatFrag ImmPred> {
  defm: HvxStfi_pat<MI, Store, Value, ImmPred>;
  defm: HvxStgi_pat<MI, Store, Value, ImmPred>;
}

let Predicates = [UseHVX] in {
  // alignedstore will match a non-temporal store as well, so try non-temporal
  // first.
  defm: HvxSt_pat<V6_vS32b_nt_ai, alignednontemporalstore,  HVI8, IsVecOff>;
  defm: HvxSt_pat<V6_vS32b_nt_ai, alignednontemporalstore, HVI16, IsVecOff>;
  defm: HvxSt_pat<V6_vS32b_nt_ai, alignednontemporalstore, HVI32, IsVecOff>;
  defm: HvxSt_pat<V6_vS32b_ai,               alignedstore,  HVI8, IsVecOff>;
  defm: HvxSt_pat<V6_vS32b_ai,               alignedstore, HVI16, IsVecOff>;
  defm: HvxSt_pat<V6_vS32b_ai,               alignedstore, HVI32, IsVecOff>;
  defm: HvxSt_pat<V6_vS32Ub_ai,            unalignedstore,  HVI8, IsVecOff>;
  defm: HvxSt_pat<V6_vS32Ub_ai,            unalignedstore, HVI16, IsVecOff>;
  defm: HvxSt_pat<V6_vS32Ub_ai,            unalignedstore, HVI32, IsVecOff>;
}

// Bitcasts between same-size vector types are no-ops, except for the
// actual type change.
let Predicates = [UseHVX] in {
  defm: NopCast_pat<VecI8,   VecI16,  HvxVR>;
  defm: NopCast_pat<VecI8,   VecI32,  HvxVR>;
  defm: NopCast_pat<VecI16,  VecI32,  HvxVR>;

  defm: NopCast_pat<VecPI8,  VecPI16, HvxWR>;
  defm: NopCast_pat<VecPI8,  VecPI32, HvxWR>;
  defm: NopCast_pat<VecPI16, VecPI32, HvxWR>;
}

let Predicates = [UseHVX] in {
  let AddedComplexity = 100 in {
    // These should be preferred over a vsplat of 0.
    def: Pat<(VecI8   vzero), (V6_vd0)>;
    def: Pat<(VecI16  vzero), (V6_vd0)>;
    def: Pat<(VecI32  vzero), (V6_vd0)>;
    def: Pat<(VecPI8  vzero), (PS_vdd0)>;
    def: Pat<(VecPI16 vzero), (PS_vdd0)>;
    def: Pat<(VecPI32 vzero), (PS_vdd0)>;

    def: Pat<(concat_vectors  (VecI8 vzero),  (VecI8 vzero)), (PS_vdd0)>;
    def: Pat<(concat_vectors (VecI16 vzero), (VecI16 vzero)), (PS_vdd0)>;
    def: Pat<(concat_vectors (VecI32 vzero), (VecI32 vzero)), (PS_vdd0)>;
  }

  def: Pat<(VecPI8 (concat_vectors HVI8:$Vs, HVI8:$Vt)),
           (Combinev HvxVR:$Vt, HvxVR:$Vs)>;
  def: Pat<(VecPI16 (concat_vectors HVI16:$Vs, HVI16:$Vt)),
           (Combinev HvxVR:$Vt, HvxVR:$Vs)>;
  def: Pat<(VecPI32 (concat_vectors HVI32:$Vs, HVI32:$Vt)),
           (Combinev HvxVR:$Vt, HvxVR:$Vs)>;

  def: Pat<(VecQ8  (qcat HQ16:$Qs, HQ16:$Qt)), (Combineq $Qt, $Qs)>;
  def: Pat<(VecQ16 (qcat HQ32:$Qs, HQ32:$Qt)), (Combineq $Qt, $Qs)>;

  def: Pat<(HexagonVEXTRACTW HVI8:$Vu, I32:$Rs),
           (V6_extractw HvxVR:$Vu, I32:$Rs)>;
  def: Pat<(HexagonVEXTRACTW HVI16:$Vu, I32:$Rs),
           (V6_extractw HvxVR:$Vu, I32:$Rs)>;
  def: Pat<(HexagonVEXTRACTW HVI32:$Vu, I32:$Rs),
           (V6_extractw HvxVR:$Vu, I32:$Rs)>;

  def: Pat<(HexagonVINSERTW0 HVI8:$Vu,  I32:$Rt),
           (V6_vinsertwr HvxVR:$Vu, I32:$Rt)>;
  def: Pat<(HexagonVINSERTW0 HVI16:$Vu, I32:$Rt),
           (V6_vinsertwr HvxVR:$Vu, I32:$Rt)>;
  def: Pat<(HexagonVINSERTW0 HVI32:$Vu, I32:$Rt),
           (V6_vinsertwr HvxVR:$Vu, I32:$Rt)>;
}

// Splats for HvxV60
def V60splatib: OutPatFrag<(ops node:$V),  (V6_lvsplatw (ToI32 (SplatB $V)))>;
def V60splatih: OutPatFrag<(ops node:$V),  (V6_lvsplatw (ToI32 (SplatH $V)))>;
def V60splatiw: OutPatFrag<(ops node:$V),  (V6_lvsplatw (ToI32 $V))>;
def V60splatrb: OutPatFrag<(ops node:$Rs), (V6_lvsplatw (S2_vsplatrb $Rs))>;
def V60splatrh: OutPatFrag<(ops node:$Rs),
                           (V6_lvsplatw (A2_combine_ll $Rs, $Rs))>;
def V60splatrw: OutPatFrag<(ops node:$Rs), (V6_lvsplatw $Rs)>;

// Splats for HvxV62+
def V62splatib: OutPatFrag<(ops node:$V),  (V6_lvsplatb (ToI32 $V))>;
def V62splatih: OutPatFrag<(ops node:$V),  (V6_lvsplath (ToI32 $V))>;
def V62splatiw: OutPatFrag<(ops node:$V),  (V6_lvsplatw (ToI32 $V))>;
def V62splatrb: OutPatFrag<(ops node:$Rs), (V6_lvsplatb $Rs)>;
def V62splatrh: OutPatFrag<(ops node:$Rs), (V6_lvsplath $Rs)>;
def V62splatrw: OutPatFrag<(ops node:$Rs), (V6_lvsplatw $Rs)>;

def Rep: OutPatFrag<(ops node:$N), (Combinev $N, $N)>;

let Predicates = [UseHVX,UseHVXV60] in {
  let AddedComplexity = 10 in {
    def: Pat<(VecI8   (splat_vector u8_0ImmPred:$V)),  (V60splatib $V)>;
    def: Pat<(VecI16  (splat_vector u16_0ImmPred:$V)), (V60splatih $V)>;
    def: Pat<(VecI32  (splat_vector anyimm:$V)),       (V60splatiw $V)>;
    def: Pat<(VecPI8  (splat_vector u8_0ImmPred:$V)),  (Rep (V60splatib $V))>;
    def: Pat<(VecPI16 (splat_vector u16_0ImmPred:$V)), (Rep (V60splatih $V))>;
    def: Pat<(VecPI32 (splat_vector anyimm:$V)),       (Rep (V60splatiw $V))>;
  }
  def: Pat<(VecI8   (splat_vector I32:$Rs)), (V60splatrb $Rs)>;
  def: Pat<(VecI16  (splat_vector I32:$Rs)), (V60splatrh $Rs)>;
  def: Pat<(VecI32  (splat_vector I32:$Rs)), (V60splatrw $Rs)>;
  def: Pat<(VecPI8  (splat_vector I32:$Rs)), (Rep (V60splatrb $Rs))>;
  def: Pat<(VecPI16 (splat_vector I32:$Rs)), (Rep (V60splatrh $Rs))>;
  def: Pat<(VecPI32 (splat_vector I32:$Rs)), (Rep (V60splatrw $Rs))>;
}
let Predicates = [UseHVX,UseHVXV62] in {
  let AddedComplexity = 30 in {
    def: Pat<(VecI8   (splat_vector u8_0ImmPred:$V)),  (V62splatib imm:$V)>;
    def: Pat<(VecI16  (splat_vector u16_0ImmPred:$V)), (V62splatih imm:$V)>;
    def: Pat<(VecI32  (splat_vector anyimm:$V)),       (V62splatiw imm:$V)>;
    def: Pat<(VecPI8  (splat_vector u8_0ImmPred:$V)),
             (Rep (V62splatib imm:$V))>;
    def: Pat<(VecPI16 (splat_vector u16_0ImmPred:$V)),
             (Rep (V62splatih imm:$V))>;
    def: Pat<(VecPI32 (splat_vector anyimm:$V)),
             (Rep (V62splatiw imm:$V))>;
  }
  let AddedComplexity = 20 in {
    def: Pat<(VecI8   (splat_vector I32:$Rs)), (V62splatrb $Rs)>;
    def: Pat<(VecI16  (splat_vector I32:$Rs)), (V62splatrh $Rs)>;
    def: Pat<(VecI32  (splat_vector I32:$Rs)), (V62splatrw $Rs)>;
    def: Pat<(VecPI8  (splat_vector I32:$Rs)), (Rep (V62splatrb $Rs))>;
    def: Pat<(VecPI16 (splat_vector I32:$Rs)), (Rep (V62splatrh $Rs))>;
    def: Pat<(VecPI32 (splat_vector I32:$Rs)), (Rep (V62splatrw $Rs))>;
  }
}

class Vneg1<ValueType VecTy>
  : PatFrag<(ops), (VecTy (splat_vector (i32 -1)))>;

class Vnot<ValueType VecTy>
  : PatFrag<(ops node:$Vs), (xor $Vs, Vneg1<VecTy>)>;

let Predicates = [UseHVX] in {
  let AddedComplexity = 200 in {
    def: Pat<(Vnot<VecI8>   HVI8:$Vs), (V6_vnot HvxVR:$Vs)>;
    def: Pat<(Vnot<VecI16> HVI16:$Vs), (V6_vnot HvxVR:$Vs)>;
    def: Pat<(Vnot<VecI32> HVI32:$Vs), (V6_vnot HvxVR:$Vs)>;
  }

  def: OpR_RR_pat<V6_vaddb,    Add,   VecI8,  HVI8>;
  def: OpR_RR_pat<V6_vaddh,    Add,  VecI16, HVI16>;
  def: OpR_RR_pat<V6_vaddw,    Add,  VecI32, HVI32>;
  def: OpR_RR_pat<V6_vaddb_dv, Add,  VecPI8,  HWI8>;
  def: OpR_RR_pat<V6_vaddh_dv, Add, VecPI16, HWI16>;
  def: OpR_RR_pat<V6_vaddw_dv, Add, VecPI32, HWI32>;
  def: OpR_RR_pat<V6_vsubb,    Sub,   VecI8,  HVI8>;
  def: OpR_RR_pat<V6_vsubh,    Sub,  VecI16, HVI16>;
  def: OpR_RR_pat<V6_vsubw,    Sub,  VecI32, HVI32>;
  def: OpR_RR_pat<V6_vsubb_dv, Sub,  VecPI8,  HWI8>;
  def: OpR_RR_pat<V6_vsubh_dv, Sub, VecPI16, HWI16>;
  def: OpR_RR_pat<V6_vsubw_dv, Sub, VecPI32, HWI32>;
  def: OpR_RR_pat<V6_vand,     And,   VecI8,  HVI8>;
  def: OpR_RR_pat<V6_vand,     And,  VecI16, HVI16>;
  def: OpR_RR_pat<V6_vand,     And,  VecI32, HVI32>;
  def: OpR_RR_pat<V6_vor,       Or,   VecI8,  HVI8>;
  def: OpR_RR_pat<V6_vor,       Or,  VecI16, HVI16>;
  def: OpR_RR_pat<V6_vor,       Or,  VecI32, HVI32>;
  def: OpR_RR_pat<V6_vxor,     Xor,   VecI8,  HVI8>;
  def: OpR_RR_pat<V6_vxor,     Xor,  VecI16, HVI16>;
  def: OpR_RR_pat<V6_vxor,     Xor,  VecI32, HVI32>;

  def: OpR_RR_pat<V6_vminb,   Smin,   VecI8,  HVI8>;
  def: OpR_RR_pat<V6_vmaxb,   Smax,   VecI8,  HVI8>;
  def: OpR_RR_pat<V6_vminub,  Umin,   VecI8,  HVI8>;
  def: OpR_RR_pat<V6_vmaxub,  Umax,   VecI8,  HVI8>;
  def: OpR_RR_pat<V6_vminh,   Smin,  VecI16, HVI16>;
  def: OpR_RR_pat<V6_vmaxh,   Smax,  VecI16, HVI16>;
  def: OpR_RR_pat<V6_vminuh,  Umin,  VecI16, HVI16>;
  def: OpR_RR_pat<V6_vmaxuh,  Umax,  VecI16, HVI16>;
  def: OpR_RR_pat<V6_vminw,   Smin,  VecI32, HVI32>;
  def: OpR_RR_pat<V6_vmaxw,   Smax,  VecI32, HVI32>;

  def: Pat<(vselect HQ8:$Qu, HVI8:$Vs, HVI8:$Vt),
           (V6_vmux HvxQR:$Qu, HvxVR:$Vs, HvxVR:$Vt)>;
  def: Pat<(vselect HQ16:$Qu, HVI16:$Vs, HVI16:$Vt),
           (V6_vmux HvxQR:$Qu, HvxVR:$Vs, HvxVR:$Vt)>;
  def: Pat<(vselect HQ32:$Qu, HVI32:$Vs, HVI32:$Vt),
           (V6_vmux HvxQR:$Qu, HvxVR:$Vs, HvxVR:$Vt)>;

  def: Pat<(vselect (qnot HQ8:$Qu), HVI8:$Vs, HVI8:$Vt),
           (V6_vmux HvxQR:$Qu, HvxVR:$Vt, HvxVR:$Vs)>;
  def: Pat<(vselect (qnot HQ16:$Qu), HVI16:$Vs, HVI16:$Vt),
           (V6_vmux HvxQR:$Qu, HvxVR:$Vt, HvxVR:$Vs)>;
  def: Pat<(vselect (qnot HQ32:$Qu), HVI32:$Vs, HVI32:$Vt),
           (V6_vmux HvxQR:$Qu, HvxVR:$Vt, HvxVR:$Vs)>;
}

let Predicates = [UseHVX] in {
  // For i8 vectors Vs = (a0, a1, ...), Vt = (b0, b1, ...),
  // V6_vmpybv Vs, Vt produces a pair of i16 vectors Hi:Lo,
  // where Lo = (a0*b0, a2*b2, ...), Hi = (a1*b1, a3*b3, ...).
  def: Pat<(mul HVI8:$Vs, HVI8:$Vt),
           (V6_vshuffeb (HiVec (V6_vmpybv HvxVR:$Vs, HvxVR:$Vt)),
                        (LoVec (V6_vmpybv HvxVR:$Vs, HvxVR:$Vt)))>;
  def: Pat<(mul HVI16:$Vs, HVI16:$Vt),
           (V6_vmpyih HvxVR:$Vs, HvxVR:$Vt)>;
  def: Pat<(mul HVI32:$Vs, HVI32:$Vt),
           (V6_vmpyiewuh_acc (V6_vmpyieoh HvxVR:$Vs, HvxVR:$Vt),
                             HvxVR:$Vs, HvxVR:$Vt)>;
}

let Predicates = [UseHVX] in {
  def: Pat<(VecPI16 (sext HVI8:$Vs)),  (VSxtb $Vs)>;
  def: Pat<(VecPI32 (sext HVI16:$Vs)), (VSxth $Vs)>;
  def: Pat<(VecPI16 (zext HVI8:$Vs)),  (VZxtb $Vs)>;
  def: Pat<(VecPI32 (zext HVI16:$Vs)), (VZxth $Vs)>;

  def: Pat<(VecI16 (sext_invec HVI8:$Vs)),  (LoVec (VSxtb $Vs))>;
  def: Pat<(VecI32 (sext_invec HVI16:$Vs)), (LoVec (VSxth $Vs))>;
  def: Pat<(VecI32 (sext_invec HVI8:$Vs)),
           (LoVec (VSxth (LoVec (VSxtb $Vs))))>;
  def: Pat<(VecPI16 (sext_invec HWI8:$Vss)),  (VSxtb (LoVec $Vss))>;
  def: Pat<(VecPI32 (sext_invec HWI16:$Vss)), (VSxth (LoVec $Vss))>;
  def: Pat<(VecPI32 (sext_invec HWI8:$Vss)),
           (VSxth (LoVec (VSxtb (LoVec $Vss))))>;

  def: Pat<(VecI16 (zext_invec HVI8:$Vs)),  (LoVec (VZxtb $Vs))>;
  def: Pat<(VecI32 (zext_invec HVI16:$Vs)), (LoVec (VZxth $Vs))>;
  def: Pat<(VecI32 (zext_invec HVI8:$Vs)),
           (LoVec (VZxth (LoVec (VZxtb $Vs))))>;
  def: Pat<(VecPI16 (zext_invec HWI8:$Vss)),  (VZxtb (LoVec $Vss))>;
  def: Pat<(VecPI32 (zext_invec HWI16:$Vss)), (VZxth (LoVec $Vss))>;
  def: Pat<(VecPI32 (zext_invec HWI8:$Vss)),
           (VZxth (LoVec (VZxtb (LoVec $Vss))))>;

  def: Pat<(VecI8 (trunc HWI16:$Vss)),
           (V6_vpackeb (HiVec $Vss), (LoVec $Vss))>;
  def: Pat<(VecI16 (trunc HWI32:$Vss)),
           (V6_vpackeh (HiVec $Vss), (LoVec $Vss))>;

  def: Pat<(VecQ8 (trunc HVI8:$Vs)),
           (V6_vandvrt HvxVR:$Vs, (A2_tfrsi 0x01010101))>;
  def: Pat<(VecQ16 (trunc HVI16:$Vs)),
           (V6_vandvrt HvxVR:$Vs, (A2_tfrsi 0x01010101))>;
  def: Pat<(VecQ32 (trunc HVI32:$Vs)),
           (V6_vandvrt HvxVR:$Vs, (A2_tfrsi 0x01010101))>;
}

let Predicates = [UseHVX] in {
  // The "source" types are not legal, and there are no parameterized
  // definitions for them, but they are length-specific.
  let Predicates = [UseHVX,UseHVX64B] in {
    def: Pat<(VecI16 (sext_inreg HVI16:$Vs, v32i8)),
             (V6_vasrh (V6_vaslh HVI16:$Vs, (A2_tfrsi 8)), (A2_tfrsi 8))>;
    def: Pat<(VecI32 (sext_inreg HVI32:$Vs, v16i8)),
             (V6_vasrw (V6_vaslw HVI32:$Vs, (A2_tfrsi 24)), (A2_tfrsi 24))>;
    def: Pat<(VecI32 (sext_inreg HVI32:$Vs, v16i16)),
             (V6_vasrw (V6_vaslw HVI32:$Vs, (A2_tfrsi 16)), (A2_tfrsi 16))>;
  }
  let Predicates = [UseHVX,UseHVX128B] in {
    def: Pat<(VecI16 (sext_inreg HVI16:$Vs, v64i8)),
             (V6_vasrh (V6_vaslh HVI16:$Vs, (A2_tfrsi 8)), (A2_tfrsi 8))>;
    def: Pat<(VecI32 (sext_inreg HVI32:$Vs, v32i8)),
             (V6_vasrw (V6_vaslw HVI32:$Vs, (A2_tfrsi 24)), (A2_tfrsi 24))>;
    def: Pat<(VecI32 (sext_inreg HVI32:$Vs, v32i16)),
             (V6_vasrw (V6_vaslw HVI32:$Vs, (A2_tfrsi 16)), (A2_tfrsi 16))>;
  }

  // Take a pair of vectors Vt:Vs and shift them towards LSB by (Rt & HwLen).
  def: Pat<(VecI8 (valign HVI8:$Vt, HVI8:$Vs, I32:$Rt)),
           (LoVec (V6_valignb HvxVR:$Vt, HvxVR:$Vs, I32:$Rt))>;
  def: Pat<(VecI16 (valign HVI16:$Vt, HVI16:$Vs, I32:$Rt)),
           (LoVec (V6_valignb HvxVR:$Vt, HvxVR:$Vs, I32:$Rt))>;
  def: Pat<(VecI32 (valign HVI32:$Vt, HVI32:$Vs, I32:$Rt)),
           (LoVec (V6_valignb HvxVR:$Vt, HvxVR:$Vs, I32:$Rt))>;

  def: Pat<(HexagonVASL HVI8:$Vs, I32:$Rt),
           (V6_vpackeb (V6_vaslh (HiVec (VZxtb HvxVR:$Vs)), I32:$Rt),
                       (V6_vaslh (LoVec (VZxtb HvxVR:$Vs)), I32:$Rt))>;
  def: Pat<(HexagonVASR HVI8:$Vs, I32:$Rt),
           (V6_vpackeb (V6_vasrh (HiVec (VSxtb HvxVR:$Vs)), I32:$Rt),
                       (V6_vasrh (LoVec (VSxtb HvxVR:$Vs)), I32:$Rt))>;
  def: Pat<(HexagonVLSR HVI8:$Vs, I32:$Rt),
           (V6_vpackeb (V6_vlsrh (HiVec (VZxtb HvxVR:$Vs)), I32:$Rt),
                       (V6_vlsrh (LoVec (VZxtb HvxVR:$Vs)), I32:$Rt))>;

  def: Pat<(HexagonVASL HVI16:$Vs, I32:$Rt), (V6_vaslh HvxVR:$Vs, I32:$Rt)>;
  def: Pat<(HexagonVASL HVI32:$Vs, I32:$Rt), (V6_vaslw HvxVR:$Vs, I32:$Rt)>;
  def: Pat<(HexagonVASR HVI16:$Vs, I32:$Rt), (V6_vasrh HvxVR:$Vs, I32:$Rt)>;
  def: Pat<(HexagonVASR HVI32:$Vs, I32:$Rt), (V6_vasrw HvxVR:$Vs, I32:$Rt)>;
  def: Pat<(HexagonVLSR HVI16:$Vs, I32:$Rt), (V6_vlsrh HvxVR:$Vs, I32:$Rt)>;
  def: Pat<(HexagonVLSR HVI32:$Vs, I32:$Rt), (V6_vlsrw HvxVR:$Vs, I32:$Rt)>;

  def: Pat<(add HVI32:$Vx, (HexagonVASL HVI32:$Vu, I32:$Rt)),
           (V6_vaslw_acc HvxVR:$Vx, HvxVR:$Vu, I32:$Rt)>;
  def: Pat<(add HVI32:$Vx, (HexagonVASR HVI32:$Vu, I32:$Rt)),
           (V6_vasrw_acc HvxVR:$Vx, HvxVR:$Vu, I32:$Rt)>;

  def: Pat<(shl HVI16:$Vs, HVI16:$Vt), (V6_vaslhv HvxVR:$Vs, HvxVR:$Vt)>;
  def: Pat<(shl HVI32:$Vs, HVI32:$Vt), (V6_vaslwv HvxVR:$Vs, HvxVR:$Vt)>;
  def: Pat<(sra HVI16:$Vs, HVI16:$Vt), (V6_vasrhv HvxVR:$Vs, HvxVR:$Vt)>;
  def: Pat<(sra HVI32:$Vs, HVI32:$Vt), (V6_vasrwv HvxVR:$Vs, HvxVR:$Vt)>;
  def: Pat<(srl HVI16:$Vs, HVI16:$Vt), (V6_vlsrhv HvxVR:$Vs, HvxVR:$Vt)>;
  def: Pat<(srl HVI32:$Vs, HVI32:$Vt), (V6_vlsrwv HvxVR:$Vs, HvxVR:$Vt)>;

  // Vpackl is a pseudo-op that is used when legalizing widened truncates.
  // It should never be produced with a register pair in the output, but
  // it can happen to have a pair as an input.
  def: Pat<(VecI8  (vpackl HVI16:$Vs)), (V6_vdealb HvxVR:$Vs)>;
  def: Pat<(VecI8  (vpackl HVI32:$Vs)), (V6_vdealb4w (IMPLICIT_DEF), HvxVR:$Vs)>;
  def: Pat<(VecI16 (vpackl HVI32:$Vs)), (V6_vdealh HvxVR:$Vs)>;
  def: Pat<(VecI8  (vpackl HWI16:$Vs)), (V6_vpackeb (HiVec $Vs), (LoVec $Vs))>;
  def: Pat<(VecI8  (vpackl HWI32:$Vs)),
           (V6_vpackeb (IMPLICIT_DEF), (V6_vpackeh (HiVec $Vs), (LoVec $Vs)))>;
  def: Pat<(VecI16 (vpackl HWI32:$Vs)), (V6_vpackeh (HiVec $Vs), (LoVec $Vs))>;

  def: Pat<(VecI16  (vunpack   HVI8:$Vs)), (LoVec (VSxtb $Vs))>;
  def: Pat<(VecI32  (vunpack   HVI8:$Vs)), (LoVec (VSxth (LoVec (VSxtb $Vs))))>;
  def: Pat<(VecI32  (vunpack  HVI16:$Vs)), (LoVec (VSxth $Vs))>;
  def: Pat<(VecPI16 (vunpack   HVI8:$Vs)), (VSxtb $Vs)>;
  def: Pat<(VecPI32 (vunpack   HVI8:$Vs)), (VSxth (LoVec (VSxtb $Vs)))>;
  def: Pat<(VecPI32 (vunpack  HVI32:$Vs)), (VSxth $Vs)>;

  def: Pat<(VecI16  (vunpacku  HVI8:$Vs)), (LoVec (VZxtb $Vs))>;
  def: Pat<(VecI32  (vunpacku  HVI8:$Vs)), (LoVec (VZxth (LoVec (VZxtb $Vs))))>;
  def: Pat<(VecI32  (vunpacku HVI16:$Vs)), (LoVec (VZxth $Vs))>;
  def: Pat<(VecPI16 (vunpacku  HVI8:$Vs)), (VZxtb $Vs)>;
  def: Pat<(VecPI32 (vunpacku  HVI8:$Vs)), (VZxth (LoVec (VZxtb $Vs)))>;
  def: Pat<(VecPI32 (vunpacku HVI32:$Vs)), (VZxth $Vs)>;

  let Predicates = [UseHVX,UseHVXV60] in {
    def: Pat<(VecI16 (bswap HVI16:$Vs)),
             (V6_vdelta HvxVR:$Vs, (V60splatib (i32 0x01)))>;
    def: Pat<(VecI32 (bswap HVI32:$Vs)),
             (V6_vdelta HvxVR:$Vs, (V60splatib (i32 0x03)))>;
  }
  let Predicates = [UseHVX,UseHVXV62], AddedComplexity = 10 in {
    def: Pat<(VecI16 (bswap HVI16:$Vs)),
             (V6_vdelta HvxVR:$Vs, (V62splatib (i32 0x01)))>;
    def: Pat<(VecI32 (bswap HVI32:$Vs)),
             (V6_vdelta HvxVR:$Vs, (V62splatib (i32 0x03)))>;
  }

  def: Pat<(VecI8 (ctpop HVI8:$Vs)),
           (V6_vpackeb (V6_vpopcounth (HiVec (V6_vunpackub HvxVR:$Vs))),
                       (V6_vpopcounth (LoVec (V6_vunpackub HvxVR:$Vs))))>;
  def: Pat<(VecI16 (ctpop HVI16:$Vs)), (V6_vpopcounth HvxVR:$Vs)>;
  def: Pat<(VecI32 (ctpop HVI32:$Vs)),
           (V6_vaddw (LoVec (V6_vzh (V6_vpopcounth HvxVR:$Vs))),
                     (HiVec (V6_vzh (V6_vpopcounth HvxVR:$Vs))))>;

  let Predicates = [UseHVX,UseHVXV60] in
  def: Pat<(VecI8 (ctlz HVI8:$Vs)),
           (V6_vsubb (V6_vpackeb (V6_vcl0h (HiVec (V6_vunpackub HvxVR:$Vs))),
                                 (V6_vcl0h (LoVec (V6_vunpackub HvxVR:$Vs)))),
                     (V60splatib (i32 0x08)))>;
  let Predicates = [UseHVX,UseHVXV62], AddedComplexity = 10 in
  def: Pat<(VecI8 (ctlz HVI8:$Vs)),
           (V6_vsubb (V6_vpackeb (V6_vcl0h (HiVec (V6_vunpackub HvxVR:$Vs))),
                                 (V6_vcl0h (LoVec (V6_vunpackub HvxVR:$Vs)))),
                     (V62splatib (i32 0x08)))>;

  def: Pat<(VecI16 (ctlz HVI16:$Vs)), (V6_vcl0h HvxVR:$Vs)>;
  def: Pat<(VecI32 (ctlz HVI32:$Vs)), (V6_vcl0w HvxVR:$Vs)>;
}

class HvxSel_pat<InstHexagon MI, PatFrag RegPred>
  : Pat<(select I1:$Pu, RegPred:$Vs, RegPred:$Vt),
        (MI I1:$Pu, RegPred:$Vs, RegPred:$Vt)>;

let Predicates = [UseHVX] in {
  def: HvxSel_pat<PS_vselect, HVI8>;
  def: HvxSel_pat<PS_vselect, HVI16>;
  def: HvxSel_pat<PS_vselect, HVI32>;
  def: HvxSel_pat<PS_wselect, HWI8>;
  def: HvxSel_pat<PS_wselect, HWI16>;
  def: HvxSel_pat<PS_wselect, HWI32>;
}

let Predicates = [UseHVX] in {
  def: Pat<(VecQ8   (qtrue)), (PS_qtrue)>;
  def: Pat<(VecQ16  (qtrue)), (PS_qtrue)>;
  def: Pat<(VecQ32  (qtrue)), (PS_qtrue)>;
  def: Pat<(VecQ8  (qfalse)), (PS_qfalse)>;
  def: Pat<(VecQ16 (qfalse)), (PS_qfalse)>;
  def: Pat<(VecQ32 (qfalse)), (PS_qfalse)>;

  def: Pat<(vnot  HQ8:$Qs), (V6_pred_not HvxQR:$Qs)>;
  def: Pat<(vnot HQ16:$Qs), (V6_pred_not HvxQR:$Qs)>;
  def: Pat<(vnot HQ32:$Qs), (V6_pred_not HvxQR:$Qs)>;
  def: Pat<(qnot  HQ8:$Qs), (V6_pred_not HvxQR:$Qs)>;
  def: Pat<(qnot HQ16:$Qs), (V6_pred_not HvxQR:$Qs)>;
  def: Pat<(qnot HQ32:$Qs), (V6_pred_not HvxQR:$Qs)>;

  def: OpR_RR_pat<V6_pred_and,  And,  VecQ8,   HQ8>;
  def: OpR_RR_pat<V6_pred_and,  And, VecQ16,  HQ16>;
  def: OpR_RR_pat<V6_pred_and,  And, VecQ32,  HQ32>;
  def: OpR_RR_pat<V6_pred_or,    Or,  VecQ8,   HQ8>;
  def: OpR_RR_pat<V6_pred_or,    Or, VecQ16,  HQ16>;
  def: OpR_RR_pat<V6_pred_or,    Or, VecQ32,  HQ32>;
  def: OpR_RR_pat<V6_pred_xor,  Xor,  VecQ8,   HQ8>;
  def: OpR_RR_pat<V6_pred_xor,  Xor, VecQ16,  HQ16>;
  def: OpR_RR_pat<V6_pred_xor,  Xor, VecQ32,  HQ32>;

  def: OpR_RR_pat<V6_pred_and_n,  VNot2<And, qnot>,  VecQ8,   HQ8>;
  def: OpR_RR_pat<V6_pred_and_n,  VNot2<And, qnot>, VecQ16,  HQ16>;
  def: OpR_RR_pat<V6_pred_and_n,  VNot2<And, qnot>, VecQ32,  HQ32>;
  def: OpR_RR_pat<V6_pred_or_n,    VNot2<Or, qnot>,  VecQ8,   HQ8>;
  def: OpR_RR_pat<V6_pred_or_n,    VNot2<Or, qnot>, VecQ16,  HQ16>;
  def: OpR_RR_pat<V6_pred_or_n,    VNot2<Or, qnot>, VecQ32,  HQ32>;

  def: OpR_RR_pat<V6_veqb,      seteq,  VecQ8,  HVI8>;
  def: OpR_RR_pat<V6_veqh,      seteq, VecQ16, HVI16>;
  def: OpR_RR_pat<V6_veqw,      seteq, VecQ32, HVI32>;
  def: OpR_RR_pat<V6_vgtb,      setgt,  VecQ8,  HVI8>;
  def: OpR_RR_pat<V6_vgth,      setgt, VecQ16, HVI16>;
  def: OpR_RR_pat<V6_vgtw,      setgt, VecQ32, HVI32>;
  def: OpR_RR_pat<V6_vgtub,    setugt,  VecQ8,  HVI8>;
  def: OpR_RR_pat<V6_vgtuh,    setugt, VecQ16, HVI16>;
  def: OpR_RR_pat<V6_vgtuw,    setugt, VecQ32, HVI32>;

  def: AccRRR_pat<V6_veqb_and,    And,  seteq,    HQ8,  HVI8,  HVI8>;
  def: AccRRR_pat<V6_veqb_or,      Or,  seteq,    HQ8,  HVI8,  HVI8>;
  def: AccRRR_pat<V6_veqb_xor,    Xor,  seteq,    HQ8,  HVI8,  HVI8>;
  def: AccRRR_pat<V6_veqh_and,    And,  seteq,   HQ16, HVI16, HVI16>;
  def: AccRRR_pat<V6_veqh_or,      Or,  seteq,   HQ16, HVI16, HVI16>;
  def: AccRRR_pat<V6_veqh_xor,    Xor,  seteq,   HQ16, HVI16, HVI16>;
  def: AccRRR_pat<V6_veqw_and,    And,  seteq,   HQ32, HVI32, HVI32>;
  def: AccRRR_pat<V6_veqw_or,      Or,  seteq,   HQ32, HVI32, HVI32>;
  def: AccRRR_pat<V6_veqw_xor,    Xor,  seteq,   HQ32, HVI32, HVI32>;

  def: AccRRR_pat<V6_vgtb_and,    And,  setgt,    HQ8,  HVI8,  HVI8>;
  def: AccRRR_pat<V6_vgtb_or,      Or,  setgt,    HQ8,  HVI8,  HVI8>;
  def: AccRRR_pat<V6_vgtb_xor,    Xor,  setgt,    HQ8,  HVI8,  HVI8>;
  def: AccRRR_pat<V6_vgth_and,    And,  setgt,   HQ16, HVI16, HVI16>;
  def: AccRRR_pat<V6_vgth_or,      Or,  setgt,   HQ16, HVI16, HVI16>;
  def: AccRRR_pat<V6_vgth_xor,    Xor,  setgt,   HQ16, HVI16, HVI16>;
  def: AccRRR_pat<V6_vgtw_and,    And,  setgt,   HQ32, HVI32, HVI32>;
  def: AccRRR_pat<V6_vgtw_or,      Or,  setgt,   HQ32, HVI32, HVI32>;
  def: AccRRR_pat<V6_vgtw_xor,    Xor,  setgt,   HQ32, HVI32, HVI32>;

  def: AccRRR_pat<V6_vgtub_and,   And, setugt,    HQ8,  HVI8,  HVI8>;
  def: AccRRR_pat<V6_vgtub_or,     Or, setugt,    HQ8,  HVI8,  HVI8>;
  def: AccRRR_pat<V6_vgtub_xor,   Xor, setugt,    HQ8,  HVI8,  HVI8>;
  def: AccRRR_pat<V6_vgtuh_and,   And, setugt,   HQ16, HVI16, HVI16>;
  def: AccRRR_pat<V6_vgtuh_or,     Or, setugt,   HQ16, HVI16, HVI16>;
  def: AccRRR_pat<V6_vgtuh_xor,   Xor, setugt,   HQ16, HVI16, HVI16>;
  def: AccRRR_pat<V6_vgtuw_and,   And, setugt,   HQ32, HVI32, HVI32>;
  def: AccRRR_pat<V6_vgtuw_or,     Or, setugt,   HQ32, HVI32, HVI32>;
  def: AccRRR_pat<V6_vgtuw_xor,   Xor, setugt,   HQ32, HVI32, HVI32>;
}