LoongArchISelLowering.cpp

Go to the documentation of this file.

//=- LoongArchISelLowering.cpp - LoongArch DAG Lowering Implementation  ---===//
//
// 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
//
//===----------------------------------------------------------------------===//
//
// This file defines the interfaces that LoongArch uses to lower LLVM code into
// a selection DAG.
//
//===----------------------------------------------------------------------===//
 
#include "LoongArchISelLowering.h"
#include "LoongArch.h"
#include "LoongArchMachineFunctionInfo.h"
#include "LoongArchRegisterInfo.h"
#include "LoongArchSelectionDAGInfo.h"
#include "LoongArchSubtarget.h"
#include "MCTargetDesc/LoongArchBaseInfo.h"
#include "MCTargetDesc/LoongArchMCTargetDesc.h"
#include "MCTargetDesc/LoongArchMatInt.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/CodeGen/ISDOpcodes.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/RuntimeLibcallUtil.h"
#include "llvm/CodeGen/SelectionDAGNodes.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/IntrinsicsLoongArch.h"
#include "llvm/Support/CodeGen.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/MathExtras.h"
#include <llvm/Analysis/VectorUtils.h>
 
using namespace llvm;
 
#define DEBUG_TYPE "loongarch-isel-lowering"
 
STATISTIC(NumTailCalls, "Number of tail calls");
 
enum MaterializeFPImm {
  NoMaterializeFPImm = 0,
  MaterializeFPImm2Ins = 2,
  MaterializeFPImm3Ins = 3,
  MaterializeFPImm4Ins = 4,
  MaterializeFPImm5Ins = 5,
  MaterializeFPImm6Ins = 6
};
 
static cl::opt<MaterializeFPImm> MaterializeFPImmInsNum(
    "loongarch-materialize-float-imm", cl::Hidden,
    cl::desc("Maximum number of instructions used (including code sequence "
             "to generate the value and moving the value to FPR) when "
             "materializing floating-point immediates (default = 3)"),
    cl::init(MaterializeFPImm3Ins),
    cl::values(clEnumValN(NoMaterializeFPImm, "0", "Use constant pool"),
               clEnumValN(MaterializeFPImm2Ins, "2",
                          "Materialize FP immediate within 2 instructions"),
               clEnumValN(MaterializeFPImm3Ins, "3",
                          "Materialize FP immediate within 3 instructions"),
               clEnumValN(MaterializeFPImm4Ins, "4",
                          "Materialize FP immediate within 4 instructions"),
               clEnumValN(MaterializeFPImm5Ins, "5",
                          "Materialize FP immediate within 5 instructions"),
               clEnumValN(MaterializeFPImm6Ins, "6",
                          "Materialize FP immediate within 6 instructions "
                          "(behaves same as 5 on loongarch64)")));
 
static cl::opt<bool> ZeroDivCheck("loongarch-check-zero-division", cl::Hidden,
                                  cl::desc("Trap on integer division by zero."),
                                  cl::init(false));
 
LoongArchTargetLowering::LoongArchTargetLowering(const TargetMachine &TM,
                                                 const LoongArchSubtarget &STI)
    : TargetLowering(TM, STI), Subtarget(STI) {
 
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  // Set up the register classes.
 
  addRegisterClass(GRLenVT, &LoongArch::GPRRegClass);
  if (Subtarget.hasBasicF())
    addRegisterClass(MVT::f32, &LoongArch::FPR32RegClass);
  if (Subtarget.hasBasicD())
    addRegisterClass(MVT::f64, &LoongArch::FPR64RegClass);
 
  static const MVT::SimpleValueType LSXVTs[] = {
      MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, MVT::v4f32, MVT::v2f64};
  static const MVT::SimpleValueType LASXVTs[] = {
      MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64, MVT::v8f32, MVT::v4f64};
 
  if (Subtarget.hasExtLSX())
    for (MVT VT : LSXVTs)
      addRegisterClass(VT, &LoongArch::LSX128RegClass);
 
  if (Subtarget.hasExtLASX())
    for (MVT VT : LASXVTs)
      addRegisterClass(VT, &LoongArch::LASX256RegClass);
 
  // Set operations for LA32 and LA64.
 
  setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, GRLenVT,
                   MVT::i1, Promote);
 
  setOperationAction(ISD::SHL_PARTS, GRLenVT, Custom);
  setOperationAction(ISD::SRA_PARTS, GRLenVT, Custom);
  setOperationAction(ISD::SRL_PARTS, GRLenVT, Custom);
  setOperationAction(ISD::FP_TO_SINT, GRLenVT, Custom);
  setOperationAction(ISD::ROTL, GRLenVT, Expand);
  setOperationAction(ISD::CTPOP, GRLenVT, Expand);
 
  setOperationAction({ISD::GlobalAddress, ISD::BlockAddress, ISD::ConstantPool,
                      ISD::JumpTable, ISD::GlobalTLSAddress},
                     GRLenVT, Custom);
 
  setOperationAction(ISD::EH_DWARF_CFA, GRLenVT, Custom);
 
  setOperationAction(ISD::DYNAMIC_STACKALLOC, GRLenVT, Custom);
  setOperationAction({ISD::STACKSAVE, ISD::STACKRESTORE}, MVT::Other, Expand);
  setOperationAction(ISD::VASTART, MVT::Other, Custom);
  setOperationAction({ISD::VAARG, ISD::VACOPY, ISD::VAEND}, MVT::Other, Expand);
 
  setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
  setOperationAction(ISD::TRAP, MVT::Other, Legal);
 
  setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
  setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
 
  setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
 
  // BITREV/REVB requires the 32S feature.
  if (STI.has32S()) {
    // Expand bitreverse.i16 with native-width bitrev and shift for now, before
    // we get to know which of sll and revb.2h is faster.
    setOperationAction(ISD::BITREVERSE, MVT::i8, Custom);
    setOperationAction(ISD::BITREVERSE, GRLenVT, Legal);
 
    // LA32 does not have REVB.2W and REVB.D due to the 64-bit operands, and
    // the narrower REVB.W does not exist. But LA32 does have REVB.2H, so i16
    // and i32 could still be byte-swapped relatively cheaply.
    setOperationAction(ISD::BSWAP, MVT::i16, Custom);
  } else {
    setOperationAction(ISD::BSWAP, GRLenVT, Expand);
    setOperationAction(ISD::CTTZ, GRLenVT, Expand);
    setOperationAction(ISD::CTLZ, GRLenVT, Expand);
    setOperationAction(ISD::ROTR, GRLenVT, Expand);
    setOperationAction(ISD::SELECT, GRLenVT, Custom);
    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
  }
 
  setOperationAction(ISD::BR_JT, MVT::Other, Expand);
  setOperationAction(ISD::BR_CC, GRLenVT, Expand);
  setOperationAction(ISD::BRCOND, MVT::Other, Custom);
  setOperationAction(ISD::SELECT_CC, GRLenVT, Expand);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
  setOperationAction({ISD::SMUL_LOHI, ISD::UMUL_LOHI}, GRLenVT, Expand);
 
  setOperationAction(ISD::FP_TO_UINT, GRLenVT, Custom);
  setOperationAction(ISD::UINT_TO_FP, GRLenVT, Expand);
 
  // Set operations for LA64 only.
 
  if (Subtarget.is64Bit()) {
    setOperationAction(ISD::ADD, MVT::i32, Custom);
    setOperationAction(ISD::SUB, MVT::i32, Custom);
    setOperationAction(ISD::SHL, MVT::i32, Custom);
    setOperationAction(ISD::SRA, MVT::i32, Custom);
    setOperationAction(ISD::SRL, MVT::i32, Custom);
    setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
    setOperationAction(ISD::BITCAST, MVT::i32, Custom);
    setOperationAction(ISD::ROTR, MVT::i32, Custom);
    setOperationAction(ISD::ROTL, MVT::i32, Custom);
    setOperationAction(ISD::CTTZ, MVT::i32, Custom);
    setOperationAction(ISD::CTLZ, MVT::i32, Custom);
    setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom);
    setOperationAction(ISD::READ_REGISTER, MVT::i32, Custom);
    setOperationAction(ISD::WRITE_REGISTER, MVT::i32, Custom);
    setOperationAction(ISD::INTRINSIC_VOID, MVT::i32, Custom);
    setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i32, Custom);
    setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i32, Custom);
 
    setOperationAction(ISD::BITREVERSE, MVT::i32, Custom);
    setOperationAction(ISD::BSWAP, MVT::i32, Custom);
    setOperationAction({ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM}, MVT::i32,
                       Custom);
    setOperationAction(ISD::LROUND, MVT::i32, Custom);
  }
 
  // Set operations for LA32 only.
 
  if (!Subtarget.is64Bit()) {
    setOperationAction(ISD::READ_REGISTER, MVT::i64, Custom);
    setOperationAction(ISD::WRITE_REGISTER, MVT::i64, Custom);
    setOperationAction(ISD::INTRINSIC_VOID, MVT::i64, Custom);
    setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
    setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
    if (Subtarget.hasBasicD())
      setOperationAction(ISD::BITCAST, MVT::i64, Custom);
  }
 
  setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
 
  static const ISD::CondCode FPCCToExpand[] = {
      ISD::SETOGT, ISD::SETOGE, ISD::SETUGT, ISD::SETUGE,
      ISD::SETGE,  ISD::SETNE,  ISD::SETGT};
 
  // Set operations for 'F' feature.
 
  if (Subtarget.hasBasicF()) {
    setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
    setTruncStoreAction(MVT::f32, MVT::f16, Expand);
    setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::bf16, Expand);
    setTruncStoreAction(MVT::f32, MVT::bf16, Expand);
    setCondCodeAction(FPCCToExpand, MVT::f32, Expand);
 
    setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
    setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
    setOperationAction(ISD::BR_CC, MVT::f32, Expand);
    setOperationAction(ISD::FMA, MVT::f32, Legal);
    setOperationAction(ISD::FMINNUM_IEEE, MVT::f32, Legal);
    setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
    setOperationAction(ISD::FMAXNUM_IEEE, MVT::f32, Legal);
    setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
    setOperationAction(ISD::FCANONICALIZE, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FSETCCS, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FSETCC, MVT::f32, Legal);
    setOperationAction(ISD::IS_FPCLASS, MVT::f32, Legal);
    setOperationAction(ISD::FSIN, MVT::f32, Expand);
    setOperationAction(ISD::FCOS, MVT::f32, Expand);
    setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
    setOperationAction(ISD::FPOW, MVT::f32, Expand);
    setOperationAction(ISD::FREM, MVT::f32, LibCall);
    setOperationAction(ISD::FP16_TO_FP, MVT::f32,
                       Subtarget.isSoftFPABI() ? LibCall : Custom);
    setOperationAction(ISD::FP_TO_FP16, MVT::f32,
                       Subtarget.isSoftFPABI() ? LibCall : Custom);
    setOperationAction(ISD::BF16_TO_FP, MVT::f32, Custom);
    setOperationAction(ISD::FP_TO_BF16, MVT::f32,
                       Subtarget.isSoftFPABI() ? LibCall : Custom);
    setOperationAction(ISD::SET_ROUNDING, MVT::Other, Custom);
    setOperationAction(ISD::GET_ROUNDING, GRLenVT, Custom);
 
    if (Subtarget.is64Bit())
      setOperationAction(ISD::FRINT, MVT::f32, Legal);
 
    if (!Subtarget.hasBasicD()) {
      setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
      if (Subtarget.is64Bit()) {
        setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
        setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
      }
    }
  }
 
  // Set operations for 'D' feature.
 
  if (Subtarget.hasBasicD()) {
    setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
    setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
    setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::bf16, Expand);
    setTruncStoreAction(MVT::f64, MVT::bf16, Expand);
    setTruncStoreAction(MVT::f64, MVT::f16, Expand);
    setTruncStoreAction(MVT::f64, MVT::f32, Expand);
    setCondCodeAction(FPCCToExpand, MVT::f64, Expand);
 
    setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
    setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
    setOperationAction(ISD::BR_CC, MVT::f64, Expand);
    setOperationAction(ISD::STRICT_FSETCCS, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_FSETCC, MVT::f64, Legal);
    setOperationAction(ISD::FMA, MVT::f64, Legal);
    setOperationAction(ISD::FMINNUM_IEEE, MVT::f64, Legal);
    setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
    setOperationAction(ISD::FMAXNUM_IEEE, MVT::f64, Legal);
    setOperationAction(ISD::FCANONICALIZE, MVT::f64, Legal);
    setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
    setOperationAction(ISD::IS_FPCLASS, MVT::f64, Legal);
    setOperationAction(ISD::FSIN, MVT::f64, Expand);
    setOperationAction(ISD::FCOS, MVT::f64, Expand);
    setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
    setOperationAction(ISD::FPOW, MVT::f64, Expand);
    setOperationAction(ISD::FREM, MVT::f64, LibCall);
    setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
    setOperationAction(ISD::FP_TO_FP16, MVT::f64,
                       Subtarget.isSoftFPABI() ? LibCall : Custom);
    setOperationAction(ISD::BF16_TO_FP, MVT::f64, Custom);
    setOperationAction(ISD::FP_TO_BF16, MVT::f64,
                       Subtarget.isSoftFPABI() ? LibCall : Custom);
 
    if (Subtarget.is64Bit())
      setOperationAction(ISD::FRINT, MVT::f64, Legal);
  }
 
  // Set operations for 'LSX' feature.
 
  if (Subtarget.hasExtLSX()) {
    for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
      // Expand all truncating stores and extending loads.
      for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) {
        setTruncStoreAction(VT, InnerVT, Expand);
        setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
        setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
        setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
      }
      // By default everything must be expanded. Then we will selectively turn
      // on ones that can be effectively codegen'd.
      for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op)
        setOperationAction(Op, VT, Expand);
    }
 
    for (MVT VT : LSXVTs) {
      setOperationAction({ISD::LOAD, ISD::STORE}, VT, Legal);
      setOperationAction(ISD::BITCAST, VT, Legal);
      setOperationAction({ISD::UNDEF, ISD::POISON}, VT, Legal);
 
      setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Legal);
      setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
 
      setOperationAction(ISD::SETCC, VT, Legal);
      setOperationAction(ISD::VSELECT, VT, Legal);
      setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
      setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
    }
    for (MVT VT : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) {
      setOperationAction({ISD::ADD, ISD::SUB}, VT, Legal);
      setOperationAction({ISD::UMAX, ISD::UMIN, ISD::SMAX, ISD::SMIN}, VT,
                         Legal);
      setOperationAction({ISD::MUL, ISD::SDIV, ISD::SREM, ISD::UDIV, ISD::UREM},
                         VT, Legal);
      setOperationAction({ISD::AND, ISD::OR, ISD::XOR}, VT, Legal);
      setOperationAction({ISD::SHL, ISD::SRA, ISD::SRL}, VT, Legal);
      setOperationAction({ISD::CTPOP, ISD::CTLZ}, VT, Legal);
      setOperationAction({ISD::MULHS, ISD::MULHU}, VT, Legal);
      setCondCodeAction(
          {ISD::SETNE, ISD::SETGE, ISD::SETGT, ISD::SETUGE, ISD::SETUGT}, VT,
          Expand);
      setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
      setOperationAction(ISD::ABS, VT, Legal);
      setOperationAction(ISD::ABDS, VT, Legal);
      setOperationAction(ISD::ABDU, VT, Legal);
      setOperationAction(ISD::SADDSAT, VT, Legal);
      setOperationAction(ISD::SSUBSAT, VT, Legal);
      setOperationAction(ISD::UADDSAT, VT, Legal);
      setOperationAction(ISD::USUBSAT, VT, Legal);
      setOperationAction(ISD::ROTL, VT, Custom);
      setOperationAction(ISD::ROTR, VT, Custom);
      setOperationAction(ISD::AVGFLOORS, VT, Legal);
      setOperationAction(ISD::AVGFLOORU, VT, Legal);
      setOperationAction(ISD::AVGCEILS, VT, Legal);
      setOperationAction(ISD::AVGCEILU, VT, Legal);
    }
    for (MVT VT : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
      setOperationAction(ISD::BITREVERSE, VT, Custom);
    for (MVT VT : {MVT::v8i16, MVT::v4i32, MVT::v2i64})
      setOperationAction(ISD::BSWAP, VT, Legal);
    for (MVT VT : {MVT::v4i32, MVT::v2i64}) {
      setOperationAction({ISD::SINT_TO_FP, ISD::UINT_TO_FP}, VT, Legal);
      setOperationAction({ISD::FP_TO_SINT, ISD::FP_TO_UINT}, VT, Legal);
    }
    setOperationAction(ISD::UINT_TO_FP, GRLenVT, Custom);
    for (MVT VT : {MVT::v4f32, MVT::v2f64}) {
      setOperationAction({ISD::FADD, ISD::FSUB}, VT, Legal);
      setOperationAction({ISD::FMUL, ISD::FDIV}, VT, Legal);
      setOperationAction(ISD::FMA, VT, Legal);
      setOperationAction(ISD::FSQRT, VT, Legal);
      setOperationAction(ISD::FNEG, VT, Legal);
      setCondCodeAction({ISD::SETGE, ISD::SETGT, ISD::SETOGE, ISD::SETOGT,
                         ISD::SETUGE, ISD::SETUGT},
                        VT, Expand);
      setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Legal);
      setOperationAction(ISD::FCEIL, VT, Legal);
      setOperationAction(ISD::FFLOOR, VT, Legal);
      setOperationAction(ISD::FTRUNC, VT, Legal);
      setOperationAction(ISD::FROUNDEVEN, VT, Legal);
      setOperationAction(ISD::FMINNUM, VT, Legal);
      setOperationAction(ISD::FMAXNUM, VT, Legal);
    }
    setOperationAction(ISD::CTPOP, GRLenVT, Legal);
    setOperationAction(ISD::FCEIL, {MVT::f32, MVT::f64}, Legal);
    setOperationAction(ISD::FFLOOR, {MVT::f32, MVT::f64}, Legal);
    setOperationAction(ISD::FTRUNC, {MVT::f32, MVT::f64}, Legal);
    setOperationAction(ISD::FROUNDEVEN, {MVT::f32, MVT::f64}, Legal);
 
    for (MVT VT :
         {MVT::v16i8, MVT::v8i8, MVT::v4i8, MVT::v2i8, MVT::v8i16, MVT::v4i16,
          MVT::v2i16, MVT::v4i32, MVT::v2i32, MVT::v2i64}) {
      setOperationAction(ISD::TRUNCATE, VT, Custom);
      setOperationAction(ISD::VECREDUCE_ADD, VT, Custom);
      setOperationAction(ISD::VECREDUCE_AND, VT, Custom);
      setOperationAction(ISD::VECREDUCE_OR, VT, Custom);
      setOperationAction(ISD::VECREDUCE_XOR, VT, Custom);
      setOperationAction(ISD::VECREDUCE_SMAX, VT, Custom);
      setOperationAction(ISD::VECREDUCE_SMIN, VT, Custom);
      setOperationAction(ISD::VECREDUCE_UMAX, VT, Custom);
      setOperationAction(ISD::VECREDUCE_UMIN, VT, Custom);
    }
    setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
    setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
    // We want to legalize this to an f64 load rather than an i64 load.
    setOperationAction(ISD::LOAD, MVT::v2f32, Custom);
    for (MVT VT : {MVT::v2i64, MVT::v4i32, MVT::v8i16})
      setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Custom);
    for (MVT VT : {MVT::v16i16, MVT::v8i32, MVT::v4i64, MVT::v16i32, MVT::v8i64,
                   MVT::v16i64})
      setOperationAction(ISD::SIGN_EXTEND, VT, Custom);
  }
 
  // Set operations for 'LASX' feature.
 
  if (Subtarget.hasExtLASX()) {
    for (MVT VT : LASXVTs) {
      setOperationAction({ISD::LOAD, ISD::STORE}, VT, Legal);
      setOperationAction(ISD::BITCAST, VT, Legal);
      setOperationAction({ISD::UNDEF, ISD::POISON}, VT, Legal);
 
      setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
      setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR, VT, Legal);
 
      setOperationAction(ISD::SETCC, VT, Custom);
      setOperationAction(ISD::VSELECT, VT, Legal);
      setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
    }
    for (MVT VT : {MVT::v4i64, MVT::v8i32, MVT::v16i16, MVT::v32i8}) {
      setOperationAction({ISD::ADD, ISD::SUB}, VT, Legal);
      setOperationAction({ISD::UMAX, ISD::UMIN, ISD::SMAX, ISD::SMIN}, VT,
                         Legal);
      setOperationAction({ISD::MUL, ISD::SDIV, ISD::SREM, ISD::UDIV, ISD::UREM},
                         VT, Legal);
      setOperationAction({ISD::AND, ISD::OR, ISD::XOR}, VT, Legal);
      setOperationAction({ISD::SHL, ISD::SRA, ISD::SRL}, VT, Legal);
      setOperationAction({ISD::CTPOP, ISD::CTLZ}, VT, Legal);
      setOperationAction({ISD::MULHS, ISD::MULHU}, VT, Legal);
      setCondCodeAction(
          {ISD::SETNE, ISD::SETGE, ISD::SETGT, ISD::SETUGE, ISD::SETUGT}, VT,
          Expand);
      setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
      setOperationAction(ISD::ABS, VT, Legal);
      setOperationAction(ISD::ABDS, VT, Legal);
      setOperationAction(ISD::ABDU, VT, Legal);
      setOperationAction(ISD::SADDSAT, VT, Legal);
      setOperationAction(ISD::SSUBSAT, VT, Legal);
      setOperationAction(ISD::UADDSAT, VT, Legal);
      setOperationAction(ISD::USUBSAT, VT, Legal);
      setOperationAction(ISD::VECREDUCE_ADD, VT, Custom);
      setOperationAction(ISD::ROTL, VT, Custom);
      setOperationAction(ISD::ROTR, VT, Custom);
      setOperationAction(ISD::AVGFLOORS, VT, Legal);
      setOperationAction(ISD::AVGFLOORU, VT, Legal);
      setOperationAction(ISD::AVGCEILS, VT, Legal);
      setOperationAction(ISD::AVGCEILU, VT, Legal);
    }
    for (MVT VT : {MVT::v32i8, MVT::v16i16, MVT::v8i32})
      setOperationAction(ISD::BITREVERSE, VT, Custom);
    for (MVT VT : {MVT::v16i16, MVT::v8i32, MVT::v4i64})
      setOperationAction(ISD::BSWAP, VT, Legal);
    for (MVT VT : {MVT::v8i32, MVT::v4i32, MVT::v4i64}) {
      setOperationAction({ISD::FP_TO_SINT, ISD::FP_TO_UINT}, VT, Legal);
      setOperationAction(ISD::SINT_TO_FP, VT, Legal);
      setOperationAction(ISD::UINT_TO_FP, VT, Custom);
    }
    for (MVT VT : {MVT::v8f32, MVT::v4f64}) {
      setOperationAction({ISD::FADD, ISD::FSUB}, VT, Legal);
      setOperationAction({ISD::FMUL, ISD::FDIV}, VT, Legal);
      setOperationAction(ISD::FMA, VT, Legal);
      setOperationAction(ISD::FSQRT, VT, Legal);
      setOperationAction(ISD::FNEG, VT, Legal);
      setCondCodeAction({ISD::SETGE, ISD::SETGT, ISD::SETOGE, ISD::SETOGT,
                         ISD::SETUGE, ISD::SETUGT},
                        VT, Expand);
      setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Legal);
      setOperationAction(ISD::FCEIL, VT, Legal);
      setOperationAction(ISD::FFLOOR, VT, Legal);
      setOperationAction(ISD::FTRUNC, VT, Legal);
      setOperationAction(ISD::FROUNDEVEN, VT, Legal);
      setOperationAction(ISD::FMINNUM, VT, Legal);
      setOperationAction(ISD::FMAXNUM, VT, Legal);
    }
    setOperationAction(ISD::FP_ROUND, MVT::v4f32, Custom);
    setOperationAction(ISD::FP_EXTEND, MVT::v4f64, Custom);
    for (MVT VT : {MVT::v4i64, MVT::v8i32, MVT::v16i16}) {
      setOperationAction(ISD::SIGN_EXTEND, VT, Legal);
      setOperationAction(ISD::ZERO_EXTEND, VT, Legal);
      setOperationAction(ISD::ANY_EXTEND, VT, Custom);
    }
    for (MVT VT :
         {MVT::v2i64, MVT::v4i32, MVT::v4i64, MVT::v8i16, MVT::v8i32}) {
      setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Legal);
      setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Legal);
    }
    for (MVT VT : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
      setOperationAction(ISD::TRUNCATE, VT, Legal);
  }
 
  // Set DAG combine for LA32 and LA64.
  if (Subtarget.hasBasicF()) {
    setTargetDAGCombine(ISD::SINT_TO_FP);
  }
 
  setTargetDAGCombine(ISD::AND);
  setTargetDAGCombine(ISD::OR);
  setTargetDAGCombine(ISD::SRL);
  setTargetDAGCombine(ISD::SETCC);
 
  // Set DAG combine for 'LSX' feature.
 
  if (Subtarget.hasExtLSX()) {
    setTargetDAGCombine(ISD::ADD);
    setTargetDAGCombine(ISD::SUB);
    setTargetDAGCombine(ISD::SHL);
    setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
    setTargetDAGCombine(ISD::BITCAST);
    setTargetDAGCombine(ISD::VSELECT);
    setTargetDAGCombine(ISD::FP_TO_SINT);
    setTargetDAGCombine(ISD::FP_TO_UINT);
    setTargetDAGCombine(ISD::UINT_TO_FP);
    setTargetDAGCombine(ISD::ZERO_EXTEND);
    setTargetDAGCombine(ISD::SIGN_EXTEND);
  }
 
  // Set DAG combine for 'LASX' feature.
  if (Subtarget.hasExtLASX()) {
    setTargetDAGCombine(ISD::ANY_EXTEND);
    setTargetDAGCombine(ISD::CONCAT_VECTORS);
  }
 
  // Compute derived properties from the register classes.
  computeRegisterProperties(Subtarget.getRegisterInfo());
 
  setStackPointerRegisterToSaveRestore(LoongArch::R3);
 
  setBooleanContents(ZeroOrOneBooleanContent);
  setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
 
  setMaxAtomicSizeInBitsSupported(Subtarget.getGRLen());
 
  setMinCmpXchgSizeInBits(32);
 
  // Function alignments.
  setMinFunctionAlignment(Align(4));
  // Set preferred alignments.
  setPrefFunctionAlignment(Subtarget.getPrefFunctionAlignment());
  setPrefLoopAlignment(Subtarget.getPrefLoopAlignment());
  setMaxBytesForAlignment(Subtarget.getMaxBytesForAlignment());
 
  // cmpxchg sizes down to 8 bits become legal if LAMCAS is available.
  if (Subtarget.hasLAMCAS())
    setMinCmpXchgSizeInBits(8);
 
  if (Subtarget.hasSCQ()) {
    setMaxAtomicSizeInBitsSupported(128);
    setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, Custom);
  }
 
  // Disable strict node mutation.
  IsStrictFPEnabled = true;
}
 
bool LoongArchTargetLowering::isOffsetFoldingLegal(
    const GlobalAddressSDNode *GA) const {
  // In order to maximise the opportunity for common subexpression elimination,
  // keep a separate ADD node for the global address offset instead of folding
  // it in the global address node. Later peephole optimisations may choose to
  // fold it back in when profitable.
  return false;
}
 
SDValue LoongArchTargetLowering::LowerOperation(SDValue Op,
                                                SelectionDAG &DAG) const {
  switch (Op.getOpcode()) {
  case ISD::ATOMIC_FENCE:
    return lowerATOMIC_FENCE(Op, DAG);
  case ISD::EH_DWARF_CFA:
    return lowerEH_DWARF_CFA(Op, DAG);
  case ISD::GlobalAddress:
    return lowerGlobalAddress(Op, DAG);
  case ISD::GlobalTLSAddress:
    return lowerGlobalTLSAddress(Op, DAG);
  case ISD::INTRINSIC_WO_CHAIN:
    return lowerINTRINSIC_WO_CHAIN(Op, DAG);
  case ISD::INTRINSIC_W_CHAIN:
    return lowerINTRINSIC_W_CHAIN(Op, DAG);
  case ISD::INTRINSIC_VOID:
    return lowerINTRINSIC_VOID(Op, DAG);
  case ISD::BlockAddress:
    return lowerBlockAddress(Op, DAG);
  case ISD::JumpTable:
    return lowerJumpTable(Op, DAG);
  case ISD::SHL_PARTS:
    return lowerShiftLeftParts(Op, DAG);
  case ISD::SRA_PARTS:
    return lowerShiftRightParts(Op, DAG, true);
  case ISD::SRL_PARTS:
    return lowerShiftRightParts(Op, DAG, false);
  case ISD::ConstantPool:
    return lowerConstantPool(Op, DAG);
  case ISD::FP_TO_SINT:
    return lowerFP_TO_SINT(Op, DAG);
  case ISD::FP_TO_UINT:
    return lowerFP_TO_UINT(Op, DAG);
  case ISD::BITCAST:
    return lowerBITCAST(Op, DAG);
  case ISD::UINT_TO_FP:
    return lowerUINT_TO_FP(Op, DAG);
  case ISD::SINT_TO_FP:
    return lowerSINT_TO_FP(Op, DAG);
  case ISD::VASTART:
    return lowerVASTART(Op, DAG);
  case ISD::FRAMEADDR:
    return lowerFRAMEADDR(Op, DAG);
  case ISD::RETURNADDR:
    return lowerRETURNADDR(Op, DAG);
  case ISD::SET_ROUNDING:
    return lowerSET_ROUNDING(Op, DAG);
  case ISD::GET_ROUNDING:
    return lowerGET_ROUNDING(Op, DAG);
  case ISD::WRITE_REGISTER:
    return lowerWRITE_REGISTER(Op, DAG);
  case ISD::INSERT_VECTOR_ELT:
    return lowerINSERT_VECTOR_ELT(Op, DAG);
  case ISD::EXTRACT_VECTOR_ELT:
    return lowerEXTRACT_VECTOR_ELT(Op, DAG);
  case ISD::BUILD_VECTOR:
    return lowerBUILD_VECTOR(Op, DAG);
  case ISD::CONCAT_VECTORS:
    return lowerCONCAT_VECTORS(Op, DAG);
  case ISD::VECTOR_SHUFFLE:
    return lowerVECTOR_SHUFFLE(Op, DAG);
  case ISD::BITREVERSE:
    return lowerBITREVERSE(Op, DAG);
  case ISD::SCALAR_TO_VECTOR:
    return lowerSCALAR_TO_VECTOR(Op, DAG);
  case ISD::PREFETCH:
    return lowerPREFETCH(Op, DAG);
  case ISD::SELECT:
    return lowerSELECT(Op, DAG);
  case ISD::BRCOND:
    return lowerBRCOND(Op, DAG);
  case ISD::FP_TO_FP16:
    return lowerFP_TO_FP16(Op, DAG);
  case ISD::FP16_TO_FP:
    return lowerFP16_TO_FP(Op, DAG);
  case ISD::FP_TO_BF16:
    return lowerFP_TO_BF16(Op, DAG);
  case ISD::BF16_TO_FP:
    return lowerBF16_TO_FP(Op, DAG);
  case ISD::VECREDUCE_ADD:
    return lowerVECREDUCE_ADD(Op, DAG);
  case ISD::ROTL:
  case ISD::ROTR:
    return lowerRotate(Op, DAG);
  case ISD::VECREDUCE_AND:
  case ISD::VECREDUCE_OR:
  case ISD::VECREDUCE_XOR:
  case ISD::VECREDUCE_SMAX:
  case ISD::VECREDUCE_SMIN:
  case ISD::VECREDUCE_UMAX:
  case ISD::VECREDUCE_UMIN:
    return lowerVECREDUCE(Op, DAG);
  case ISD::ConstantFP:
    return lowerConstantFP(Op, DAG);
  case ISD::SETCC:
    return lowerSETCC(Op, DAG);
  case ISD::FP_ROUND:
    return lowerFP_ROUND(Op, DAG);
  case ISD::FP_EXTEND:
    return lowerFP_EXTEND(Op, DAG);
  case ISD::SIGN_EXTEND_VECTOR_INREG:
    return lowerSIGN_EXTEND_VECTOR_INREG(Op, DAG);
  case ISD::DYNAMIC_STACKALLOC:
    return lowerDYNAMIC_STACKALLOC(Op, DAG);
  case ISD::ANY_EXTEND:
    return lowerANY_EXTEND(Op, DAG);
  }
  return SDValue();
}
 
// Helper to attempt to return a cheaper, bit-inverted version of \p V.
static SDValue isNOT(SDValue V, SelectionDAG &DAG) {
  // TODO: don't always ignore oneuse constraints.
  V = peekThroughBitcasts(V);
  EVT VT = V.getValueType();
 
  // Match not(xor X, -1) -> X.
  if (V.getOpcode() == ISD::XOR &&
      (ISD::isBuildVectorAllOnes(V.getOperand(1).getNode()) ||
       isAllOnesConstant(V.getOperand(1))))
    return V.getOperand(0);
 
  // Match not(extract_subvector(not(X)) -> extract_subvector(X).
  if (V.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
      (isNullConstant(V.getOperand(1)) || V.getOperand(0).hasOneUse())) {
    if (SDValue Not = isNOT(V.getOperand(0), DAG)) {
      Not = DAG.getBitcast(V.getOperand(0).getValueType(), Not);
      return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(Not), VT, Not,
                         V.getOperand(1));
    }
  }
 
  // Match not(SplatVector(not(X)) -> SplatVector(X).
  if (V.getOpcode() == ISD::BUILD_VECTOR) {
    if (SDValue SplatValue =
            cast<BuildVectorSDNode>(V.getNode())->getSplatValue()) {
      if (!V->isOnlyUserOf(SplatValue.getNode()))
        return SDValue();
 
      if (SDValue Not = isNOT(SplatValue, DAG)) {
        Not = DAG.getBitcast(V.getOperand(0).getValueType(), Not);
        return DAG.getSplat(VT, SDLoc(Not), Not);
      }
    }
  }
 
  // Match not(or(not(X),not(Y))) -> and(X, Y).
  if (V.getOpcode() == ISD::OR && DAG.getTargetLoweringInfo().isTypeLegal(VT) &&
      V.getOperand(0).hasOneUse() && V.getOperand(1).hasOneUse()) {
    // TODO: Handle cases with single NOT operand -> VANDN
    if (SDValue Op1 = isNOT(V.getOperand(1), DAG))
      if (SDValue Op0 = isNOT(V.getOperand(0), DAG))
        return DAG.getNode(ISD::AND, SDLoc(V), VT, DAG.getBitcast(VT, Op0),
                           DAG.getBitcast(VT, Op1));
  }
 
  // TODO: Add more matching patterns. Such as,
  // not(concat_vectors(not(X), not(Y))) -> concat_vectors(X, Y).
  // not(slt(C, X)) -> slt(X - 1, C)
  return SDValue();
}
 
// Combine two ISD::FP_ROUND / LoongArchISD::VFCVT nodes with same type to
// LoongArchISD::VFCVT. For example:
//   x1 = fp_round x, 0
//   y1 = fp_round y, 0
//   z = concat_vectors x1, y1
// Or
//   x1 = LoongArch::VFCVT undef, x
//   y1 = LoongArch::VFCVT undef, y
//   z = LoongArchISD::VPACKEV y1, x1; or LoongArchISD::VPERMI y1, x1, 68
// can be combined to:
//   z = LoongArch::VFCVT y, x
static SDValue combineFP_ROUND(SDValue N, const SDLoc &DL, SelectionDAG &DAG,
                               const LoongArchSubtarget &Subtarget) {
  assert(((N->getOpcode() == ISD::CONCAT_VECTORS && N->getNumOperands() == 2) ||
          (N->getOpcode() == LoongArchISD::VPACKEV) ||
          (N->getOpcode() == LoongArchISD::VPERMI)) &&
         "Invalid Node");
 
  SDValue Op0 = peekThroughBitcasts(N->getOperand(0));
  SDValue Op1 = peekThroughBitcasts(N->getOperand(1));
  unsigned Opcode0 = Op0.getOpcode();
  unsigned Opcode1 = Op1.getOpcode();
  if (Opcode0 != Opcode1)
    return SDValue();
 
  if (Opcode0 != ISD::FP_ROUND && Opcode0 != LoongArchISD::VFCVT)
    return SDValue();
 
  // Check if two nodes have only one use.
  if (!Op0.hasOneUse() || !Op1.hasOneUse())
    return SDValue();
 
  EVT VT = N.getValueType();
  EVT SVT0 = Op0.getValueType();
  EVT SVT1 = Op1.getValueType();
  // Check if two nodes have the same result type.
  if (SVT0 != SVT1)
    return SDValue();
 
  // Check if two nodes have the same operand type.
  EVT SSVT0 = Op0.getOperand(0).getValueType();
  EVT SSVT1 = Op1.getOperand(0).getValueType();
  if (SSVT0 != SSVT1)
    return SDValue();
 
  if (N->getOpcode() == ISD::CONCAT_VECTORS && Opcode0 == ISD::FP_ROUND) {
    if (Subtarget.hasExtLASX() && VT.is256BitVector() && SVT0 == MVT::v4f32 &&
        SSVT0 == MVT::v4f64) {
      // A vector_shuffle is required in the final step, as xvfcvt instruction
      // operates on each 128-bit segament as a lane.
      SDValue Res = DAG.getNode(LoongArchISD::VFCVT, DL, MVT::v8f32,
                                Op1.getOperand(0), Op0.getOperand(0));
      SDValue Undef = DAG.getUNDEF(Res.getValueType());
      // After VFCVT, the high part of Res comes from the high parts of Op0 and
      // Op1, and the low part comes from the low parts of Op0 and Op1. However,
      // the desired order requires Op0 to fully occupy the lower half and Op1
      // the upper half of Res. The Mask reorders the elements of Res to achieve
      // this:
      // - The first four elements (0, 1, 4, 5) come from Op0.
      // - The next four elements (2, 3, 6, 7) come from Op1.
      SmallVector<int, 8> Mask = {0, 1, 4, 5, 2, 3, 6, 7};
      Res = DAG.getVectorShuffle(Res.getValueType(), DL, Res, Undef, Mask);
      return DAG.getBitcast(VT, Res);
    }
  }
 
  if ((N->getOpcode() == LoongArchISD::VPACKEV ||
       N->getOpcode() == LoongArchISD::VPERMI) &&
      Opcode0 == LoongArchISD::VFCVT) {
    // For VPACKEV or VPERMI, check if the first operation of VFCVT is undef.
    if (!Op0.getOperand(0).isUndef() || !Op1.getOperand(0).isUndef())
      return SDValue();
 
    if (!Subtarget.hasExtLSX() || SVT0 != MVT::v4f32 || SSVT0 != MVT::v2f64)
      return SDValue();
 
    if (N->getOpcode() == LoongArchISD::VPACKEV &&
        (VT == MVT::v2i64 || VT == MVT::v2f64)) {
      SDValue Res = DAG.getNode(LoongArchISD::VFCVT, DL, MVT::v4f32,
                                Op0.getOperand(1), Op1.getOperand(1));
      return DAG.getBitcast(VT, Res);
    }
 
    if (N->getOpcode() == LoongArchISD::VPERMI && VT == MVT::v4f32) {
      int64_t Imm = cast<ConstantSDNode>(N->getOperand(2))->getSExtValue();
      if (Imm != 68)
        return SDValue();
      return DAG.getNode(LoongArchISD::VFCVT, DL, MVT::v4f32, Op0.getOperand(1),
                         Op1.getOperand(1));
    }
  }
 
  return SDValue();
}
 
SDValue LoongArchTargetLowering::lowerFP_ROUND(SDValue Op,
                                               SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SDValue In = Op.getOperand(0);
  MVT VT = Op.getSimpleValueType();
  MVT SVT = In.getSimpleValueType();
 
  if (VT == MVT::v4f32 && SVT == MVT::v4f64) {
    SDValue Lo, Hi;
    std::tie(Lo, Hi) = DAG.SplitVector(In, DL);
    return DAG.getNode(LoongArchISD::VFCVT, DL, VT, Hi, Lo);
  }
 
  return SDValue();
}
 
SDValue LoongArchTargetLowering::lowerFP_EXTEND(SDValue Op,
                                                SelectionDAG &DAG) const {
 
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  SDValue Src = Op->getOperand(0);
  EVT SVT = Src.getValueType();
 
  bool V2F32ToV2F64 =
      VT == MVT::v2f64 && SVT == MVT::v2f32 && Subtarget.hasExtLSX();
  bool V4F32ToV4F64 =
      VT == MVT::v4f64 && SVT == MVT::v4f32 && Subtarget.hasExtLASX();
  if (!V2F32ToV2F64 && !V4F32ToV4F64)
    return SDValue();
 
  // Check if Op is the high part of vector.
  auto CheckVecHighPart = [](SDValue Op) {
    Op = peekThroughBitcasts(Op);
    if (Op.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
      SDValue SOp = Op.getOperand(0);
      EVT SVT = SOp.getValueType();
      if (!SVT.isVector() || (SVT.getVectorNumElements() % 2 != 0))
        return SDValue();
 
      const uint64_t Imm = Op.getConstantOperandVal(1);
      if (Imm == SVT.getVectorNumElements() / 2)
        return SOp;
      return SDValue();
    }
    return SDValue();
  };
 
  unsigned Opcode;
  SDValue VFCVTOp;
  EVT WideOpVT = SVT.getSimpleVT().getDoubleNumVectorElementsVT();
  SDValue ZeroIdx = DAG.getVectorIdxConstant(0, DL);
 
  // If the operand of ISD::FP_EXTEND comes from the high part of vector,
  // generate LoongArchISD::VFCVTH, otherwise LoongArchISD::VFCVTL.
  if (SDValue V = CheckVecHighPart(Src)) {
    assert(V.getValueSizeInBits() == WideOpVT.getSizeInBits() &&
           "Unexpected wide vector");
    Opcode = LoongArchISD::VFCVTH;
    VFCVTOp = DAG.getBitcast(WideOpVT, V);
  } else {
    Opcode = LoongArchISD::VFCVTL;
    VFCVTOp = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideOpVT,
                          DAG.getUNDEF(WideOpVT), Src, ZeroIdx);
  }
 
  // v2f64 = fp_extend v2f32
  if (V2F32ToV2F64)
    return DAG.getNode(Opcode, DL, VT, VFCVTOp);
 
  // v4f64 = fp_extend v4f32
  if (V4F32ToV4F64) {
    // XVFCVT instruction operates on each 128-bit segment as a lane, so a
    // vector_shuffle is required firstly.
    SmallVector<int, 8> Mask = {0, 1, 4, 5, 2, 3, 6, 7};
    SDValue Res = DAG.getVectorShuffle(WideOpVT, DL, VFCVTOp,
                                       DAG.getUNDEF(WideOpVT), Mask);
    Res = DAG.getNode(Opcode, DL, VT, Res);
    return Res;
  }
 
  return SDValue();
}
 
SDValue LoongArchTargetLowering::lowerConstantFP(SDValue Op,
                                                 SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
  const APFloat &FPVal = CFP->getValueAPF();
  SDLoc DL(CFP);
 
  assert((VT == MVT::f32 && Subtarget.hasBasicF()) ||
         (VT == MVT::f64 && Subtarget.hasBasicD()));
 
  // If value is 0.0 or -0.0, just ignore it.
  if (FPVal.isZero())
    return SDValue();
 
  // If lsx enabled, use cheaper 'vldi' instruction if possible.
  if (isFPImmVLDILegal(FPVal, VT))
    return SDValue();
 
  // Construct as integer, and move to float register.
  APInt INTVal = FPVal.bitcastToAPInt();
 
  // If more than MaterializeFPImmInsNum instructions will be used to
  // generate the INTVal and move it to float register, fallback to
  // use floating point load from the constant pool.
  auto Seq = LoongArchMatInt::generateInstSeq(INTVal.getSExtValue());
  int InsNum = Seq.size() + ((VT == MVT::f64 && !Subtarget.is64Bit()) ? 2 : 1);
  if (InsNum > MaterializeFPImmInsNum && !FPVal.isOne())
    return SDValue();
 
  switch (VT.getSimpleVT().SimpleTy) {
  default:
    llvm_unreachable("Unexpected floating point type!");
    break;
  case MVT::f32: {
    SDValue NewVal = DAG.getConstant(INTVal, DL, MVT::i32);
    if (Subtarget.is64Bit())
      NewVal = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, NewVal);
    return DAG.getNode(Subtarget.is64Bit() ? LoongArchISD::MOVGR2FR_W_LA64
                                           : LoongArchISD::MOVGR2FR_W,
                       DL, VT, NewVal);
  }
  case MVT::f64: {
    if (Subtarget.is64Bit()) {
      SDValue NewVal = DAG.getConstant(INTVal, DL, MVT::i64);
      return DAG.getNode(LoongArchISD::MOVGR2FR_D, DL, VT, NewVal);
    }
    SDValue Lo = DAG.getConstant(INTVal.trunc(32), DL, MVT::i32);
    SDValue Hi = DAG.getConstant(INTVal.lshr(32).trunc(32), DL, MVT::i32);
    return DAG.getNode(LoongArchISD::MOVGR2FR_D_LO_HI, DL, VT, Lo, Hi);
  }
  }
 
  return SDValue();
}
 
// Ensure SETCC result and operand have the same bit width; isel does not
// support mismatched widths.
SDValue LoongArchTargetLowering::lowerSETCC(SDValue Op,
                                            SelectionDAG &DAG) const {
  SDLoc DL(Op);
  EVT ResultVT = Op.getValueType();
  EVT OperandVT = Op.getOperand(0).getValueType();
 
  EVT SetCCResultVT =
      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), OperandVT);
 
  if (ResultVT == SetCCResultVT)
    return Op;
 
  assert(Op.getOperand(0).getValueType() == Op.getOperand(1).getValueType() &&
         "SETCC operands must have the same type!");
 
  SDValue SetCCNode =
      DAG.getNode(ISD::SETCC, DL, SetCCResultVT, Op.getOperand(0),
                  Op.getOperand(1), Op.getOperand(2));
 
  if (ResultVT.bitsGT(SetCCResultVT))
    SetCCNode = DAG.getNode(ISD::SIGN_EXTEND, DL, ResultVT, SetCCNode);
  else if (ResultVT.bitsLT(SetCCResultVT))
    SetCCNode = DAG.getNode(ISD::TRUNCATE, DL, ResultVT, SetCCNode);
 
  return SetCCNode;
}
 
// Lower sext_invec using vslti instructions.
// For example:
//  %b = sext <4 x i16> %a to <4 x i32>
// can be lowered to:
//  VSLTI_H vr2, vr1, 0
//  VILVL.H vr1, vr2, vr1
SDValue LoongArchTargetLowering::lowerSIGN_EXTEND_VECTOR_INREG(
    SDValue Op, SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SDValue Src = Op.getOperand(0);
  MVT SrcVT = Src.getSimpleValueType();
  MVT DstVT = Op.getSimpleValueType();
 
  if (!SrcVT.is128BitVector())
    return SDValue();
 
  // lower to VSLTI + VILVL if extend could be done in single step.
  if (DstVT.getScalarSizeInBits() / SrcVT.getScalarSizeInBits() == 2) {
    SDValue Zero = DAG.getConstant(0, DL, SrcVT);
    SDValue Mask = DAG.getNode(ISD::SETCC, DL, SrcVT, Src, Zero,
                               DAG.getCondCode(ISD::SETLT));
    SDValue LoInterleaved =
        DAG.getNode(LoongArchISD::VILVL, DL, SrcVT, Mask, Src);
 
    return DAG.getBitcast(DstVT, LoInterleaved);
  }
 
  return SDValue();
}
 
// ANY_EXTEND can be replaced by ZERO_EXTEND when LASX is enabled.
SDValue LoongArchTargetLowering::lowerANY_EXTEND(SDValue Op,
                                                 SelectionDAG &DAG) const {
  assert(Subtarget.hasExtLASX());
  // We don't have corresponding instrunction for ANY_EXTEND, lowering it to
  // ZERO_EXTEND won't break its semantics, while avoid scalar extract/insert.
  return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(Op), Op.getValueType(),
                     Op.getOperand(0));
}
 
// Lower vecreduce_add using vhaddw instructions.
// For Example:
//  call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %a)
// can be lowered to:
//  VHADDW_D_W    vr0, vr0, vr0
//  VHADDW_Q_D    vr0, vr0, vr0
//  VPICKVE2GR_D  a0,  vr0, 0
//  ADDI_W        a0,  a0,  0
SDValue LoongArchTargetLowering::lowerVECREDUCE_ADD(SDValue Op,
                                                    SelectionDAG &DAG) const {
 
  SDLoc DL(Op);
  MVT OpVT = Op.getSimpleValueType();
  SDValue Val = Op.getOperand(0);
 
  unsigned NumEles = Val.getSimpleValueType().getVectorNumElements();
  unsigned EleBits = Val.getSimpleValueType().getScalarSizeInBits();
  unsigned ResBits = OpVT.getScalarSizeInBits();
 
  unsigned LegalVecSize = 128;
  bool isLASX256Vector =
      Subtarget.hasExtLASX() && Val.getValueSizeInBits() == 256;
 
  // Ensure operand type legal or enable it legal.
  while (!isTypeLegal(Val.getSimpleValueType())) {
    Val = DAG.WidenVector(Val, DL);
  }
 
  // NumEles is designed for iterations count, v4i32 for LSX
  // and v8i32 for LASX should have the same count.
  if (isLASX256Vector) {
    NumEles /= 2;
    LegalVecSize = 256;
  }
 
  EleBits *= 2;
  for (unsigned i = 1; i < NumEles; i *= 2, EleBits *= 2) {
    EleBits = std::min(EleBits, 64u);
    MVT IntTy = MVT::getIntegerVT(EleBits);
    MVT VecTy = MVT::getVectorVT(IntTy, LegalVecSize / EleBits);
    Val = DAG.getNode(LoongArchISD::VHADDW, DL, VecTy, Val, Val);
  }
 
  if (isLASX256Vector) {
    SDValue Tmp = DAG.getNode(LoongArchISD::XVPERMI, DL, MVT::v4i64, Val,
                              DAG.getConstant(2, DL, Subtarget.getGRLenVT()));
    Val = DAG.getNode(ISD::ADD, DL, MVT::v4i64, Tmp, Val);
  }
 
  Val = DAG.getBitcast(MVT::getVectorVT(OpVT, LegalVecSize / ResBits), Val);
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, OpVT, Val,
                     DAG.getConstant(0, DL, Subtarget.getGRLenVT()));
}
 
// Lower vecreduce_and/or/xor/[s/u]max/[s/u]min.
// For Example:
//  call i32 @llvm.vector.reduce.smax.v4i32(<4 x i32> %a)
// can be lowered to:
//  VBSRL_V       vr1, vr0, 8
//  VMAX_W        vr0, vr1, vr0
//  VBSRL_V       vr1, vr0, 4
//  VMAX_W        vr0, vr1, vr0
//  VPICKVE2GR_W  a0,  vr0, 0
// For 256 bit vector, it is illegal and will be spilt into
// two 128 bit vector by default then processed by this.
SDValue LoongArchTargetLowering::lowerVECREDUCE(SDValue Op,
                                                SelectionDAG &DAG) const {
  SDLoc DL(Op);
 
  MVT OpVT = Op.getSimpleValueType();
  SDValue Val = Op.getOperand(0);
 
  unsigned NumEles = Val.getSimpleValueType().getVectorNumElements();
  unsigned EleBits = Val.getSimpleValueType().getScalarSizeInBits();
 
  // Ensure operand type legal or enable it legal.
  while (!isTypeLegal(Val.getSimpleValueType())) {
    Val = DAG.WidenVector(Val, DL);
  }
 
  unsigned Opcode = ISD::getVecReduceBaseOpcode(Op.getOpcode());
  MVT VecTy = Val.getSimpleValueType();
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  for (int i = NumEles; i > 1; i /= 2) {
    SDValue ShiftAmt = DAG.getConstant(i * EleBits / 16, DL, GRLenVT);
    SDValue Tmp = DAG.getNode(LoongArchISD::VBSRL, DL, VecTy, Val, ShiftAmt);
    Val = DAG.getNode(Opcode, DL, VecTy, Tmp, Val);
  }
 
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, OpVT, Val,
                     DAG.getConstant(0, DL, GRLenVT));
}
 
SDValue LoongArchTargetLowering::lowerPREFETCH(SDValue Op,
                                               SelectionDAG &DAG) const {
  unsigned IsData = Op.getConstantOperandVal(4);
 
  // We don't support non-data prefetch.
  // Just preserve the chain.
  if (!IsData)
    return Op.getOperand(0);
 
  return Op;
}
 
SDValue LoongArchTargetLowering::lowerRotate(SDValue Op,
                                             SelectionDAG &DAG) const {
  MVT VT = Op.getSimpleValueType();
  assert(VT.isVector() && "Unexpected type");
 
  SDLoc DL(Op);
  SDValue R = Op.getOperand(0);
  SDValue Amt = Op.getOperand(1);
  unsigned Opcode = Op.getOpcode();
  unsigned EltSizeInBits = VT.getScalarSizeInBits();
 
  auto checkCstSplat = [](SDValue V, APInt &CstSplatValue) {
    if (V.getOpcode() != ISD::BUILD_VECTOR)
      return false;
    if (SDValue SplatValue =
            cast<BuildVectorSDNode>(V.getNode())->getSplatValue()) {
      if (auto *C = dyn_cast<ConstantSDNode>(SplatValue)) {
        CstSplatValue = C->getAPIntValue();
        return true;
      }
    }
    return false;
  };
 
  // Check for constant splat rotation amount.
  APInt CstSplatValue;
  bool IsCstSplat = checkCstSplat(Amt, CstSplatValue);
  bool isROTL = Opcode == ISD::ROTL;
 
  // Check for splat rotate by zero.
  if (IsCstSplat && CstSplatValue.urem(EltSizeInBits) == 0)
    return R;
 
  // LoongArch targets always prefer ISD::ROTR.
  if (isROTL) {
    SDValue Zero = DAG.getConstant(0, DL, VT);
    return DAG.getNode(ISD::ROTR, DL, VT, R,
                       DAG.getNode(ISD::SUB, DL, VT, Zero, Amt));
  }
 
  // Rotate by a immediate.
  if (IsCstSplat) {
    // ISD::ROTR: Attemp to rotate by a positive immediate.
    SDValue Bits = DAG.getConstant(EltSizeInBits, DL, VT);
    if (SDValue Urem =
            DAG.FoldConstantArithmetic(ISD::UREM, DL, VT, {Amt, Bits}))
      return DAG.getNode(Opcode, DL, VT, R, Urem);
  }
 
  return Op;
}
 
// Return true if Val is equal to (setcc LHS, RHS, CC).
// Return false if Val is the inverse of (setcc LHS, RHS, CC).
// Otherwise, return std::nullopt.
static std::optional<bool> matchSetCC(SDValue LHS, SDValue RHS,
                                      ISD::CondCode CC, SDValue Val) {
  assert(Val->getOpcode() == ISD::SETCC);
  SDValue LHS2 = Val.getOperand(0);
  SDValue RHS2 = Val.getOperand(1);
  ISD::CondCode CC2 = cast<CondCodeSDNode>(Val.getOperand(2))->get();
 
  if (LHS == LHS2 && RHS == RHS2) {
    if (CC == CC2)
      return true;
    if (CC == ISD::getSetCCInverse(CC2, LHS2.getValueType()))
      return false;
  } else if (LHS == RHS2 && RHS == LHS2) {
    CC2 = ISD::getSetCCSwappedOperands(CC2);
    if (CC == CC2)
      return true;
    if (CC == ISD::getSetCCInverse(CC2, LHS2.getValueType()))
      return false;
  }
 
  return std::nullopt;
}
 
static SDValue combineSelectToBinOp(SDNode *N, SelectionDAG &DAG,
                                    const LoongArchSubtarget &Subtarget) {
  SDValue CondV = N->getOperand(0);
  SDValue TrueV = N->getOperand(1);
  SDValue FalseV = N->getOperand(2);
  MVT VT = N->getSimpleValueType(0);
  SDLoc DL(N);
 
  // (select c, -1, y) -> -c | y
  if (isAllOnesConstant(TrueV)) {
    SDValue Neg = DAG.getNegative(CondV, DL, VT);
    return DAG.getNode(ISD::OR, DL, VT, Neg, DAG.getFreeze(FalseV));
  }
  // (select c, y, -1) -> (c-1) | y
  if (isAllOnesConstant(FalseV)) {
    SDValue Neg =
        DAG.getNode(ISD::ADD, DL, VT, CondV, DAG.getAllOnesConstant(DL, VT));
    return DAG.getNode(ISD::OR, DL, VT, Neg, DAG.getFreeze(TrueV));
  }
 
  // (select c, 0, y) -> (c-1) & y
  if (isNullConstant(TrueV)) {
    SDValue Neg =
        DAG.getNode(ISD::ADD, DL, VT, CondV, DAG.getAllOnesConstant(DL, VT));
    return DAG.getNode(ISD::AND, DL, VT, Neg, DAG.getFreeze(FalseV));
  }
  // (select c, y, 0) -> -c & y
  if (isNullConstant(FalseV)) {
    SDValue Neg = DAG.getNegative(CondV, DL, VT);
    return DAG.getNode(ISD::AND, DL, VT, Neg, DAG.getFreeze(TrueV));
  }
 
  // select c, ~x, x --> xor -c, x
  if (isa<ConstantSDNode>(TrueV) && isa<ConstantSDNode>(FalseV)) {
    const APInt &TrueVal = TrueV->getAsAPIntVal();
    const APInt &FalseVal = FalseV->getAsAPIntVal();
    if (~TrueVal == FalseVal) {
      SDValue Neg = DAG.getNegative(CondV, DL, VT);
      return DAG.getNode(ISD::XOR, DL, VT, Neg, FalseV);
    }
  }
 
  // Try to fold (select (setcc lhs, rhs, cc), truev, falsev) into bitwise ops
  // when both truev and falsev are also setcc.
  if (CondV.getOpcode() == ISD::SETCC && TrueV.getOpcode() == ISD::SETCC &&
      FalseV.getOpcode() == ISD::SETCC) {
    SDValue LHS = CondV.getOperand(0);
    SDValue RHS = CondV.getOperand(1);
    ISD::CondCode CC = cast<CondCodeSDNode>(CondV.getOperand(2))->get();
 
    // (select x, x, y) -> x | y
    // (select !x, x, y) -> x & y
    if (std::optional<bool> MatchResult = matchSetCC(LHS, RHS, CC, TrueV)) {
      return DAG.getNode(*MatchResult ? ISD::OR : ISD::AND, DL, VT, TrueV,
                         DAG.getFreeze(FalseV));
    }
    // (select x, y, x) -> x & y
    // (select !x, y, x) -> x | y
    if (std::optional<bool> MatchResult = matchSetCC(LHS, RHS, CC, FalseV)) {
      return DAG.getNode(*MatchResult ? ISD::AND : ISD::OR, DL, VT,
                         DAG.getFreeze(TrueV), FalseV);
    }
  }
 
  return SDValue();
}
 
// Transform `binOp (select cond, x, c0), c1` where `c0` and `c1` are constants
// into `select cond, binOp(x, c1), binOp(c0, c1)` if profitable.
// For now we only consider transformation profitable if `binOp(c0, c1)` ends up
// being `0` or `-1`. In such cases we can replace `select` with `and`.
// TODO: Should we also do this if `binOp(c0, c1)` is cheaper to materialize
// than `c0`?
static SDValue
foldBinOpIntoSelectIfProfitable(SDNode *BO, SelectionDAG &DAG,
                                const LoongArchSubtarget &Subtarget) {
  unsigned SelOpNo = 0;
  SDValue Sel = BO->getOperand(0);
  if (Sel.getOpcode() != ISD::SELECT || !Sel.hasOneUse()) {
    SelOpNo = 1;
    Sel = BO->getOperand(1);
  }
 
  if (Sel.getOpcode() != ISD::SELECT || !Sel.hasOneUse())
    return SDValue();
 
  unsigned ConstSelOpNo = 1;
  unsigned OtherSelOpNo = 2;
  if (!isa<ConstantSDNode>(Sel->getOperand(ConstSelOpNo))) {
    ConstSelOpNo = 2;
    OtherSelOpNo = 1;
  }
  SDValue ConstSelOp = Sel->getOperand(ConstSelOpNo);
  ConstantSDNode *ConstSelOpNode = dyn_cast<ConstantSDNode>(ConstSelOp);
  if (!ConstSelOpNode || ConstSelOpNode->isOpaque())
    return SDValue();
 
  SDValue ConstBinOp = BO->getOperand(SelOpNo ^ 1);
  ConstantSDNode *ConstBinOpNode = dyn_cast<ConstantSDNode>(ConstBinOp);
  if (!ConstBinOpNode || ConstBinOpNode->isOpaque())
    return SDValue();
 
  SDLoc DL(Sel);
  EVT VT = BO->getValueType(0);
 
  SDValue NewConstOps[2] = {ConstSelOp, ConstBinOp};
  if (SelOpNo == 1)
    std::swap(NewConstOps[0], NewConstOps[1]);
 
  SDValue NewConstOp =
      DAG.FoldConstantArithmetic(BO->getOpcode(), DL, VT, NewConstOps);
  if (!NewConstOp)
    return SDValue();
 
  const APInt &NewConstAPInt = NewConstOp->getAsAPIntVal();
  if (!NewConstAPInt.isZero() && !NewConstAPInt.isAllOnes())
    return SDValue();
 
  SDValue OtherSelOp = Sel->getOperand(OtherSelOpNo);
  SDValue NewNonConstOps[2] = {OtherSelOp, ConstBinOp};
  if (SelOpNo == 1)
    std::swap(NewNonConstOps[0], NewNonConstOps[1]);
  SDValue NewNonConstOp = DAG.getNode(BO->getOpcode(), DL, VT, NewNonConstOps);
 
  SDValue NewT = (ConstSelOpNo == 1) ? NewConstOp : NewNonConstOp;
  SDValue NewF = (ConstSelOpNo == 1) ? NewNonConstOp : NewConstOp;
  return DAG.getSelect(DL, VT, Sel.getOperand(0), NewT, NewF);
}
 
// Changes the condition code and swaps operands if necessary, so the SetCC
// operation matches one of the comparisons supported directly by branches
// in the LoongArch ISA. May adjust compares to favor compare with 0 over
// compare with 1/-1.
static void translateSetCCForBranch(const SDLoc &DL, SDValue &LHS, SDValue &RHS,
                                    ISD::CondCode &CC, SelectionDAG &DAG) {
  // If this is a single bit test that can't be handled by ANDI, shift the
  // bit to be tested to the MSB and perform a signed compare with 0.
  if (isIntEqualitySetCC(CC) && isNullConstant(RHS) &&
      LHS.getOpcode() == ISD::AND && LHS.hasOneUse() &&
      isa<ConstantSDNode>(LHS.getOperand(1))) {
    uint64_t Mask = LHS.getConstantOperandVal(1);
    if ((isPowerOf2_64(Mask) || isMask_64(Mask)) && !isInt<12>(Mask)) {
      unsigned ShAmt = 0;
      if (isPowerOf2_64(Mask)) {
        CC = CC == ISD::SETEQ ? ISD::SETGE : ISD::SETLT;
        ShAmt = LHS.getValueSizeInBits() - 1 - Log2_64(Mask);
      } else {
        ShAmt = LHS.getValueSizeInBits() - llvm::bit_width(Mask);
      }
 
      LHS = LHS.getOperand(0);
      if (ShAmt != 0)
        LHS = DAG.getNode(ISD::SHL, DL, LHS.getValueType(), LHS,
                          DAG.getConstant(ShAmt, DL, LHS.getValueType()));
      return;
    }
  }
 
  if (auto *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
    int64_t C = RHSC->getSExtValue();
    switch (CC) {
    default:
      break;
    case ISD::SETGT:
      // Convert X > -1 to X >= 0.
      if (C == -1) {
        RHS = DAG.getConstant(0, DL, RHS.getValueType());
        CC = ISD::SETGE;
        return;
      }
      break;
    case ISD::SETLT:
      // Convert X < 1 to 0 >= X.
      if (C == 1) {
        RHS = LHS;
        LHS = DAG.getConstant(0, DL, RHS.getValueType());
        CC = ISD::SETGE;
        return;
      }
      break;
    }
  }
 
  switch (CC) {
  default:
    break;
  case ISD::SETGT:
  case ISD::SETLE:
  case ISD::SETUGT:
  case ISD::SETULE:
    CC = ISD::getSetCCSwappedOperands(CC);
    std::swap(LHS, RHS);
    break;
  }
}
 
SDValue LoongArchTargetLowering::lowerSELECT(SDValue Op,
                                             SelectionDAG &DAG) const {
  SDValue CondV = Op.getOperand(0);
  SDValue TrueV = Op.getOperand(1);
  SDValue FalseV = Op.getOperand(2);
  SDLoc DL(Op);
  MVT VT = Op.getSimpleValueType();
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  if (SDValue V = combineSelectToBinOp(Op.getNode(), DAG, Subtarget))
    return V;
 
  if (Op.hasOneUse()) {
    unsigned UseOpc = Op->user_begin()->getOpcode();
    if (isBinOp(UseOpc) && DAG.isSafeToSpeculativelyExecute(UseOpc)) {
      SDNode *BinOp = *Op->user_begin();
      if (SDValue NewSel = foldBinOpIntoSelectIfProfitable(*Op->user_begin(),
                                                           DAG, Subtarget)) {
        DAG.ReplaceAllUsesWith(BinOp, &NewSel);
        // Opcode check is necessary because foldBinOpIntoSelectIfProfitable
        // may return a constant node and cause crash in lowerSELECT.
        if (NewSel.getOpcode() == ISD::SELECT)
          return lowerSELECT(NewSel, DAG);
        return NewSel;
      }
    }
  }
 
  // If the condition is not an integer SETCC which operates on GRLenVT, we need
  // to emit a LoongArchISD::SELECT_CC comparing the condition to zero. i.e.:
  // (select condv, truev, falsev)
  // -> (loongarchisd::select_cc condv, zero, setne, truev, falsev)
  if (CondV.getOpcode() != ISD::SETCC ||
      CondV.getOperand(0).getSimpleValueType() != GRLenVT) {
    SDValue Zero = DAG.getConstant(0, DL, GRLenVT);
    SDValue SetNE = DAG.getCondCode(ISD::SETNE);
 
    SDValue Ops[] = {CondV, Zero, SetNE, TrueV, FalseV};
 
    return DAG.getNode(LoongArchISD::SELECT_CC, DL, VT, Ops);
  }
 
  // If the CondV is the output of a SETCC node which operates on GRLenVT
  // inputs, then merge the SETCC node into the lowered LoongArchISD::SELECT_CC
  // to take advantage of the integer compare+branch instructions. i.e.: (select
  // (setcc lhs, rhs, cc), truev, falsev)
  // -> (loongarchisd::select_cc lhs, rhs, cc, truev, falsev)
  SDValue LHS = CondV.getOperand(0);
  SDValue RHS = CondV.getOperand(1);
  ISD::CondCode CCVal = cast<CondCodeSDNode>(CondV.getOperand(2))->get();
 
  // Special case for a select of 2 constants that have a difference of 1.
  // Normally this is done by DAGCombine, but if the select is introduced by
  // type legalization or op legalization, we miss it. Restricting to SETLT
  // case for now because that is what signed saturating add/sub need.
  // FIXME: We don't need the condition to be SETLT or even a SETCC,
  // but we would probably want to swap the true/false values if the condition
  // is SETGE/SETLE to avoid an XORI.
  if (isa<ConstantSDNode>(TrueV) && isa<ConstantSDNode>(FalseV) &&
      CCVal == ISD::SETLT) {
    const APInt &TrueVal = TrueV->getAsAPIntVal();
    const APInt &FalseVal = FalseV->getAsAPIntVal();
    if (TrueVal - 1 == FalseVal)
      return DAG.getNode(ISD::ADD, DL, VT, CondV, FalseV);
    if (TrueVal + 1 == FalseVal)
      return DAG.getNode(ISD::SUB, DL, VT, FalseV, CondV);
  }
 
  translateSetCCForBranch(DL, LHS, RHS, CCVal, DAG);
  // 1 < x ? x : 1 -> 0 < x ? x : 1
  if (isOneConstant(LHS) && (CCVal == ISD::SETLT || CCVal == ISD::SETULT) &&
      RHS == TrueV && LHS == FalseV) {
    LHS = DAG.getConstant(0, DL, VT);
    // 0 <u x is the same as x != 0.
    if (CCVal == ISD::SETULT) {
      std::swap(LHS, RHS);
      CCVal = ISD::SETNE;
    }
  }
 
  // x <s -1 ? x : -1 -> x <s 0 ? x : -1
  if (isAllOnesConstant(RHS) && CCVal == ISD::SETLT && LHS == TrueV &&
      RHS == FalseV) {
    RHS = DAG.getConstant(0, DL, VT);
  }
 
  SDValue TargetCC = DAG.getCondCode(CCVal);
 
  if (isa<ConstantSDNode>(TrueV) && !isa<ConstantSDNode>(FalseV)) {
    // (select (setcc lhs, rhs, CC), constant, falsev)
    // -> (select (setcc lhs, rhs, InverseCC), falsev, constant)
    std::swap(TrueV, FalseV);
    TargetCC = DAG.getCondCode(ISD::getSetCCInverse(CCVal, LHS.getValueType()));
  }
 
  SDValue Ops[] = {LHS, RHS, TargetCC, TrueV, FalseV};
  return DAG.getNode(LoongArchISD::SELECT_CC, DL, VT, Ops);
}
 
SDValue LoongArchTargetLowering::lowerBRCOND(SDValue Op,
                                             SelectionDAG &DAG) const {
  SDValue CondV = Op.getOperand(1);
  SDLoc DL(Op);
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  if (CondV.getOpcode() == ISD::SETCC) {
    if (CondV.getOperand(0).getValueType() == GRLenVT) {
      SDValue LHS = CondV.getOperand(0);
      SDValue RHS = CondV.getOperand(1);
      ISD::CondCode CCVal = cast<CondCodeSDNode>(CondV.getOperand(2))->get();
 
      translateSetCCForBranch(DL, LHS, RHS, CCVal, DAG);
 
      SDValue TargetCC = DAG.getCondCode(CCVal);
      return DAG.getNode(LoongArchISD::BR_CC, DL, Op.getValueType(),
                         Op.getOperand(0), LHS, RHS, TargetCC,
                         Op.getOperand(2));
    } else if (CondV.getOperand(0).getValueType().isFloatingPoint()) {
      return DAG.getNode(LoongArchISD::BRCOND, DL, Op.getValueType(),
                         Op.getOperand(0), CondV, Op.getOperand(2));
    }
  }
 
  return DAG.getNode(LoongArchISD::BR_CC, DL, Op.getValueType(),
                     Op.getOperand(0), CondV, DAG.getConstant(0, DL, GRLenVT),
                     DAG.getCondCode(ISD::SETNE), Op.getOperand(2));
}
 
SDValue
LoongArchTargetLowering::lowerSCALAR_TO_VECTOR(SDValue Op,
                                               SelectionDAG &DAG) const {
  SDLoc DL(Op);
  MVT OpVT = Op.getSimpleValueType();
 
  SDValue Vector = DAG.getUNDEF(OpVT);
  SDValue Val = Op.getOperand(0);
  SDValue Idx = DAG.getConstant(0, DL, Subtarget.getGRLenVT());
 
  return DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, OpVT, Vector, Val, Idx);
}
 
SDValue LoongArchTargetLowering::lowerBITREVERSE(SDValue Op,
                                                 SelectionDAG &DAG) const {
  EVT ResTy = Op->getValueType(0);
  SDValue Src = Op->getOperand(0);
  SDLoc DL(Op);
 
  // LoongArchISD::BITREV_8B is not supported on LA32.
  if (!Subtarget.is64Bit() && (ResTy == MVT::v16i8 || ResTy == MVT::v32i8))
    return SDValue();
 
  EVT NewVT = ResTy.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
  unsigned int OrigEltNum = ResTy.getVectorNumElements();
  unsigned int NewEltNum = NewVT.getVectorNumElements();
 
  SDValue NewSrc = DAG.getNode(ISD::BITCAST, DL, NewVT, Src);
 
  SmallVector<SDValue, 8> Ops;
  for (unsigned int i = 0; i < NewEltNum; i++) {
    SDValue Op = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i64, NewSrc,
                             DAG.getConstant(i, DL, Subtarget.getGRLenVT()));
    unsigned RevOp = (ResTy == MVT::v16i8 || ResTy == MVT::v32i8)
                         ? (unsigned)LoongArchISD::BITREV_8B
                         : (unsigned)ISD::BITREVERSE;
    Ops.push_back(DAG.getNode(RevOp, DL, MVT::i64, Op));
  }
  SDValue Res =
      DAG.getNode(ISD::BITCAST, DL, ResTy, DAG.getBuildVector(NewVT, DL, Ops));
 
  switch (ResTy.getSimpleVT().SimpleTy) {
  default:
    return SDValue();
  case MVT::v16i8:
  case MVT::v32i8:
    return Res;
  case MVT::v8i16:
  case MVT::v16i16:
  case MVT::v4i32:
  case MVT::v8i32: {
    SmallVector<int, 32> Mask;
    for (unsigned int i = 0; i < NewEltNum; i++)
      for (int j = OrigEltNum / NewEltNum - 1; j >= 0; j--)
        Mask.push_back(j + (OrigEltNum / NewEltNum) * i);
    return DAG.getVectorShuffle(ResTy, DL, Res, DAG.getUNDEF(ResTy), Mask);
  }
  }
}
 
// Widen element type to get a new mask value (if possible).
// For example:
//  shufflevector <4 x i32> %a, <4 x i32> %b,
//                <4 x i32> <i32 6, i32 7, i32 2, i32 3>
// is equivalent to:
//  shufflevector <2 x i64> %a, <2 x i64> %b, <2 x i32> <i32 3, i32 1>
// can be lowered to:
//  VPACKOD_D vr0, vr0, vr1
static SDValue widenShuffleMask(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                                SDValue V1, SDValue V2, SelectionDAG &DAG) {
  unsigned EltBits = VT.getScalarSizeInBits();
 
  if (EltBits > 32 || EltBits == 1)
    return SDValue();
 
  SmallVector<int, 8> NewMask;
  if (widenShuffleMaskElts(Mask, NewMask)) {
    MVT NewEltVT = VT.isFloatingPoint() ? MVT::getFloatingPointVT(EltBits * 2)
                                        : MVT::getIntegerVT(EltBits * 2);
    MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
    if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
      SDValue NewV1 = DAG.getBitcast(NewVT, V1);
      SDValue NewV2 = DAG.getBitcast(NewVT, V2);
      return DAG.getBitcast(
          VT, DAG.getVectorShuffle(NewVT, DL, NewV1, NewV2, NewMask));
    }
  }
 
  return SDValue();
}
 
/// Attempts to match a shuffle mask against the VBSLL, VBSRL, VSLLI and VSRLI
/// instruction.
// The funciton matches elements from one of the input vector shuffled to the
// left or right with zeroable elements 'shifted in'. It handles both the
// strictly bit-wise element shifts and the byte shfit across an entire 128-bit
// lane.
// Mostly copied from X86.
static int matchShuffleAsShift(MVT &ShiftVT, unsigned &Opcode,
                               unsigned ScalarSizeInBits, ArrayRef<int> Mask,
                               int MaskOffset, const APInt &Zeroable) {
  int Size = Mask.size();
  unsigned SizeInBits = Size * ScalarSizeInBits;
 
  auto CheckZeros = [&](int Shift, int Scale, bool Left) {
    for (int i = 0; i < Size; i += Scale)
      for (int j = 0; j < Shift; ++j)
        if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
          return false;
 
    return true;
  };
 
  auto isSequentialOrUndefInRange = [&](unsigned Pos, unsigned Size, int Low,
                                        int Step = 1) {
    for (unsigned i = Pos, e = Pos + Size; i != e; ++i, Low += Step)
      if (!(Mask[i] == -1 || Mask[i] == Low))
        return false;
    return true;
  };
 
  auto MatchShift = [&](int Shift, int Scale, bool Left) {
    for (int i = 0; i != Size; i += Scale) {
      unsigned Pos = Left ? i + Shift : i;
      unsigned Low = Left ? i : i + Shift;
      unsigned Len = Scale - Shift;
      if (!isSequentialOrUndefInRange(Pos, Len, Low + MaskOffset))
        return -1;
    }
 
    int ShiftEltBits = ScalarSizeInBits * Scale;
    bool ByteShift = ShiftEltBits > 64;
    Opcode = Left ? (ByteShift ? LoongArchISD::VBSLL : LoongArchISD::VSLLI)
                  : (ByteShift ? LoongArchISD::VBSRL : LoongArchISD::VSRLI);
    int ShiftAmt = Shift * ScalarSizeInBits / (ByteShift ? 8 : 1);
 
    // Normalize the scale for byte shifts to still produce an i64 element
    // type.
    Scale = ByteShift ? Scale / 2 : Scale;
 
    // We need to round trip through the appropriate type for the shift.
    MVT ShiftSVT = MVT::getIntegerVT(ScalarSizeInBits * Scale);
    ShiftVT = ByteShift ? MVT::getVectorVT(MVT::i8, SizeInBits / 8)
                        : MVT::getVectorVT(ShiftSVT, Size / Scale);
    return (int)ShiftAmt;
  };
 
  unsigned MaxWidth = 128;
  for (int Scale = 2; Scale * ScalarSizeInBits <= MaxWidth; Scale *= 2)
    for (int Shift = 1; Shift != Scale; ++Shift)
      for (bool Left : {true, false})
        if (CheckZeros(Shift, Scale, Left)) {
          int ShiftAmt = MatchShift(Shift, Scale, Left);
          if (0 < ShiftAmt)
            return ShiftAmt;
        }
 
  // no match
  return -1;
}
 
/// Lower VECTOR_SHUFFLE as shift (if possible).
///
/// For example:
///   %2 = shufflevector <4 x i32> %0, <4 x i32> zeroinitializer,
///                      <4 x i32> <i32 4, i32 0, i32 1, i32 2>
/// is lowered to:
///     (VBSLL_V $v0, $v0, 4)
///
///   %2 = shufflevector <4 x i32> %0, <4 x i32> zeroinitializer,
///                      <4 x i32> <i32 4, i32 0, i32 4, i32 2>
/// is lowered to:
///     (VSLLI_D $v0, $v0, 32)
static SDValue lowerVECTOR_SHUFFLEAsShift(const SDLoc &DL, ArrayRef<int> Mask,
                                          MVT VT, SDValue V1, SDValue V2,
                                          SelectionDAG &DAG,
                                          const LoongArchSubtarget &Subtarget,
                                          const APInt &Zeroable) {
  int Size = Mask.size();
  assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
 
  MVT ShiftVT;
  SDValue V = V1;
  unsigned Opcode;
 
  // Try to match shuffle against V1 shift.
  int ShiftAmt = matchShuffleAsShift(ShiftVT, Opcode, VT.getScalarSizeInBits(),
                                     Mask, 0, Zeroable);
 
  // If V1 failed, try to match shuffle against V2 shift.
  if (ShiftAmt < 0) {
    ShiftAmt = matchShuffleAsShift(ShiftVT, Opcode, VT.getScalarSizeInBits(),
                                   Mask, Size, Zeroable);
    V = V2;
  }
 
  if (ShiftAmt < 0)
    return SDValue();
 
  assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
         "Illegal integer vector type");
  V = DAG.getBitcast(ShiftVT, V);
  V = DAG.getNode(Opcode, DL, ShiftVT, V,
                  DAG.getConstant(ShiftAmt, DL, Subtarget.getGRLenVT()));
  return DAG.getBitcast(VT, V);
}
 
/// Determine whether a range fits a regular pattern of values.
/// This function accounts for the possibility of jumping over the End iterator.
template <typename ValType>
static bool
fitsRegularPattern(typename SmallVectorImpl<ValType>::const_iterator Begin,
                   unsigned CheckStride,
                   typename SmallVectorImpl<ValType>::const_iterator End,
                   ValType ExpectedIndex, unsigned ExpectedIndexStride) {
  auto &I = Begin;
 
  while (I != End) {
    if (*I != -1 && *I != ExpectedIndex)
      return false;
    ExpectedIndex += ExpectedIndexStride;
 
    // Incrementing past End is undefined behaviour so we must increment one
    // step at a time and check for End at each step.
    for (unsigned n = 0; n < CheckStride && I != End; ++n, ++I)
      ; // Empty loop body.
  }
  return true;
}
 
/// Compute whether each element of a shuffle is zeroable.
///
/// A "zeroable" vector shuffle element is one which can be lowered to zero.
static void computeZeroableShuffleElements(ArrayRef<int> Mask, SDValue V1,
                                           SDValue V2, APInt &KnownUndef,
                                           APInt &KnownZero) {
  int Size = Mask.size();
  KnownUndef = KnownZero = APInt::getZero(Size);
 
  V1 = peekThroughBitcasts(V1);
  V2 = peekThroughBitcasts(V2);
 
  bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
  bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
 
  int VectorSizeInBits = V1.getValueSizeInBits();
  int ScalarSizeInBits = VectorSizeInBits / Size;
  assert(!(VectorSizeInBits % ScalarSizeInBits) && "Illegal shuffle mask size");
  (void)ScalarSizeInBits;
 
  for (int i = 0; i < Size; ++i) {
    int M = Mask[i];
    if (M < 0) {
      KnownUndef.setBit(i);
      continue;
    }
    if ((M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
      KnownZero.setBit(i);
      continue;
    }
  }
}
 
/// Test whether a shuffle mask is equivalent within each sub-lane.
///
/// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
/// non-trivial to compute in the face of undef lanes. The representation is
/// suitable for use with existing 128-bit shuffles as entries from the second
/// vector have been remapped to [LaneSize, 2*LaneSize).
static bool isRepeatedShuffleMask(unsigned LaneSizeInBits, MVT VT,
                                  ArrayRef<int> Mask,
                                  SmallVectorImpl<int> &RepeatedMask) {
  auto LaneSize = LaneSizeInBits / VT.getScalarSizeInBits();
  RepeatedMask.assign(LaneSize, -1);
  int Size = Mask.size();
  for (int i = 0; i < Size; ++i) {
    assert(Mask[i] == -1 || Mask[i] >= 0);
    if (Mask[i] < 0)
      continue;
    if ((Mask[i] % Size) / LaneSize != i / LaneSize)
      // This entry crosses lanes, so there is no way to model this shuffle.
      return false;
 
    // Ok, handle the in-lane shuffles by detecting if and when they repeat.
    // Adjust second vector indices to start at LaneSize instead of Size.
    int LocalM =
        Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + LaneSize;
    if (RepeatedMask[i % LaneSize] < 0)
      // This is the first non-undef entry in this slot of a 128-bit lane.
      RepeatedMask[i % LaneSize] = LocalM;
    else if (RepeatedMask[i % LaneSize] != LocalM)
      // Found a mismatch with the repeated mask.
      return false;
  }
  return true;
}
 
/// Attempts to match vector shuffle as byte rotation.
static int matchShuffleAsByteRotate(MVT VT, SDValue &V1, SDValue &V2,
                                    ArrayRef<int> Mask) {
 
  SDValue Lo, Hi;
  SmallVector<int, 16> RepeatedMask;
 
  if (!isRepeatedShuffleMask(128, VT, Mask, RepeatedMask))
    return -1;
 
  int NumElts = RepeatedMask.size();
  int Rotation = 0;
  int Scale = 16 / NumElts;
 
  for (int i = 0; i < NumElts; ++i) {
    int M = RepeatedMask[i];
    assert((M == -1 || (0 <= M && M < (2 * NumElts))) &&
           "Unexpected mask index.");
    if (M < 0)
      continue;
 
    // Determine where a rotated vector would have started.
    int StartIdx = i - (M % NumElts);
    if (StartIdx == 0)
      return -1;
 
    // If we found the tail of a vector the rotation must be the missing
    // front. If we found the head of a vector, it must be how much of the
    // head.
    int CandidateRotation = StartIdx < 0 ? -StartIdx : NumElts - StartIdx;
 
    if (Rotation == 0)
      Rotation = CandidateRotation;
    else if (Rotation != CandidateRotation)
      return -1;
 
    // Compute which value this mask is pointing at.
    SDValue MaskV = M < NumElts ? V1 : V2;
 
    // Compute which of the two target values this index should be assigned
    // to. This reflects whether the high elements are remaining or the low
    // elements are remaining.
    SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
 
    // Either set up this value if we've not encountered it before, or check
    // that it remains consistent.
    if (!TargetV)
      TargetV = MaskV;
    else if (TargetV != MaskV)
      return -1;
  }
 
  // Check that we successfully analyzed the mask, and normalize the results.
  assert(Rotation != 0 && "Failed to locate a viable rotation!");
  assert((Lo || Hi) && "Failed to find a rotated input vector!");
  if (!Lo)
    Lo = Hi;
  else if (!Hi)
    Hi = Lo;
 
  V1 = Lo;
  V2 = Hi;
 
  return Rotation * Scale;
}
 
/// Lower VECTOR_SHUFFLE as byte rotate (if possible).
///
/// For example:
///   %shuffle = shufflevector <2 x i64> %a, <2 x i64> %b,
///                            <2 x i32> <i32 3, i32 0>
/// is lowered to:
///      (VBSRL_V $v1, $v1, 8)
///      (VBSLL_V $v0, $v0, 8)
///      (VOR_V $v0, $V0, $v1)
static SDValue
lowerVECTOR_SHUFFLEAsByteRotate(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                                SDValue V1, SDValue V2, SelectionDAG &DAG,
                                const LoongArchSubtarget &Subtarget) {
 
  SDValue Lo = V1, Hi = V2;
  int ByteRotation = matchShuffleAsByteRotate(VT, Lo, Hi, Mask);
  if (ByteRotation <= 0)
    return SDValue();
 
  MVT ByteVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
  Lo = DAG.getBitcast(ByteVT, Lo);
  Hi = DAG.getBitcast(ByteVT, Hi);
 
  int LoByteShift = 16 - ByteRotation;
  int HiByteShift = ByteRotation;
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  SDValue LoShift = DAG.getNode(LoongArchISD::VBSLL, DL, ByteVT, Lo,
                                DAG.getConstant(LoByteShift, DL, GRLenVT));
  SDValue HiShift = DAG.getNode(LoongArchISD::VBSRL, DL, ByteVT, Hi,
                                DAG.getConstant(HiByteShift, DL, GRLenVT));
  return DAG.getBitcast(VT, DAG.getNode(ISD::OR, DL, ByteVT, LoShift, HiShift));
}
 
/// Lower VECTOR_SHUFFLE as ZERO_EXTEND Or ANY_EXTEND (if possible).
///
/// For example:
///   %2 = shufflevector <4 x i32> %0, <4 x i32> zeroinitializer,
///                      <4 x i32> <i32 0, i32 4, i32 1, i32 4>
///   %3 = bitcast <4 x i32> %2 to <2 x i64>
/// is lowered to:
///     (VREPLI $v1, 0)
///     (VILVL $v0, $v1, $v0)
static SDValue lowerVECTOR_SHUFFLEAsZeroOrAnyExtend(const SDLoc &DL,
                                                    ArrayRef<int> Mask, MVT VT,
                                                    SDValue V1, SDValue V2,
                                                    SelectionDAG &DAG,
                                                    const APInt &Zeroable) {
  int Bits = VT.getSizeInBits();
  int EltBits = VT.getScalarSizeInBits();
  int NumElements = VT.getVectorNumElements();
 
  if (Zeroable.isAllOnes())
    return DAG.getConstant(0, DL, VT);
 
  // Define a helper function to check a particular ext-scale and lower to it if
  // valid.
  auto Lower = [&](int Scale) -> SDValue {
    SDValue InputV;
    bool AnyExt = true;
    int Offset = 0;
    for (int i = 0; i < NumElements; i++) {
      int M = Mask[i];
      if (M < 0)
        continue;
      if (i % Scale != 0) {
        // Each of the extended elements need to be zeroable.
        if (!Zeroable[i])
          return SDValue();
 
        AnyExt = false;
        continue;
      }
 
      // Each of the base elements needs to be consecutive indices into the
      // same input vector.
      SDValue V = M < NumElements ? V1 : V2;
      M = M % NumElements;
      if (!InputV) {
        InputV = V;
        Offset = M - (i / Scale);
 
        // These offset can't be handled
        if (Offset % (NumElements / Scale))
          return SDValue();
      } else if (InputV != V)
        return SDValue();
 
      if (M != (Offset + (i / Scale)))
        return SDValue(); // Non-consecutive strided elements.
    }
 
    // If we fail to find an input, we have a zero-shuffle which should always
    // have already been handled.
    if (!InputV)
      return SDValue();
 
    do {
      unsigned VilVLoHi = LoongArchISD::VILVL;
      if (Offset >= (NumElements / 2)) {
        VilVLoHi = LoongArchISD::VILVH;
        Offset -= (NumElements / 2);
      }
 
      MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
      SDValue Ext =
          AnyExt ? DAG.getFreeze(InputV) : DAG.getConstant(0, DL, InputVT);
      InputV = DAG.getBitcast(InputVT, InputV);
      InputV = DAG.getNode(VilVLoHi, DL, InputVT, Ext, InputV);
      Scale /= 2;
      EltBits *= 2;
      NumElements /= 2;
    } while (Scale > 1);
    return DAG.getBitcast(VT, InputV);
  };
 
  // Each iteration, try extending the elements half as much, but into twice as
  // many elements.
  for (int NumExtElements = Bits / 64; NumExtElements < NumElements;
       NumExtElements *= 2) {
    if (SDValue V = Lower(NumElements / NumExtElements))
      return V;
  }
  return SDValue();
}
 
/// Lower VECTOR_SHUFFLE into VREPLVEI (if possible).
///
/// VREPLVEI performs vector broadcast based on an element specified by an
/// integer immediate, with its mask being similar to:
///   <x, x, x, ...>
/// where x is any valid index.
///
/// When undef's appear in the mask they are treated as if they were whatever
/// value is necessary in order to fit the above form.
static SDValue
lowerVECTOR_SHUFFLE_VREPLVEI(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                             SDValue V1, SelectionDAG &DAG,
                             const LoongArchSubtarget &Subtarget) {
  int SplatIndex = -1;
  for (const auto &M : Mask) {
    if (M != -1) {
      SplatIndex = M;
      break;
    }
  }
 
  if (SplatIndex == -1)
    return DAG.getUNDEF(VT);
 
  assert(SplatIndex < (int)Mask.size() && "Out of bounds mask index");
  if (fitsRegularPattern<int>(Mask.begin(), 1, Mask.end(), SplatIndex, 0)) {
    return DAG.getNode(LoongArchISD::VREPLVEI, DL, VT, V1,
                       DAG.getConstant(SplatIndex, DL, Subtarget.getGRLenVT()));
  }
 
  return SDValue();
}
 
/// Lower VECTOR_SHUFFLE into VSHUF4I (if possible).
///
/// VSHUF4I splits the vector into blocks of four elements, then shuffles these
/// elements according to a <4 x i2> constant (encoded as an integer immediate).
///
/// It is therefore possible to lower into VSHUF4I when the mask takes the form:
///   <a, b, c, d, a+4, b+4, c+4, d+4, a+8, b+8, c+8, d+8, ...>
/// When undef's appear they are treated as if they were whatever value is
/// necessary in order to fit the above forms.
///
/// For example:
///   %2 = shufflevector <8 x i16> %0, <8 x i16> undef,
///                      <8 x i32> <i32 3, i32 2, i32 1, i32 0,
///                                 i32 7, i32 6, i32 5, i32 4>
/// is lowered to:
///   (VSHUF4I_H $v0, $v1, 27)
/// where the 27 comes from:
///   3 + (2 << 2) + (1 << 4) + (0 << 6)
static SDValue
lowerVECTOR_SHUFFLE_VSHUF4I(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                            SDValue V1, SDValue V2, SelectionDAG &DAG,
                            const LoongArchSubtarget &Subtarget) {
 
  unsigned SubVecSize = 4;
  if (VT == MVT::v2f64 || VT == MVT::v2i64)
    SubVecSize = 2;
 
  int SubMask[4] = {-1, -1, -1, -1};
  for (unsigned i = 0; i < SubVecSize; ++i) {
    for (unsigned j = i; j < Mask.size(); j += SubVecSize) {
      int M = Mask[j];
 
      // Convert from vector index to 4-element subvector index
      // If an index refers to an element outside of the subvector then give up
      if (M != -1) {
        M -= 4 * (j / SubVecSize);
        if (M < 0 || M >= 4)
          return SDValue();
      }
 
      // If the mask has an undef, replace it with the current index.
      // Note that it might still be undef if the current index is also undef
      if (SubMask[i] == -1)
        SubMask[i] = M;
      // Check that non-undef values are the same as in the mask. If they
      // aren't then give up
      else if (M != -1 && M != SubMask[i])
        return SDValue();
    }
  }
 
  // Calculate the immediate. Replace any remaining undefs with zero
  int Imm = 0;
  for (int i = SubVecSize - 1; i >= 0; --i) {
    int M = SubMask[i];
 
    if (M == -1)
      M = 0;
 
    Imm <<= 2;
    Imm |= M & 0x3;
  }
 
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  // Return vshuf4i.d
  if (VT == MVT::v2f64 || VT == MVT::v2i64)
    return DAG.getNode(LoongArchISD::VSHUF4I_D, DL, VT, V1, V2,
                       DAG.getConstant(Imm, DL, GRLenVT));
 
  return DAG.getNode(LoongArchISD::VSHUF4I, DL, VT, V1,
                     DAG.getConstant(Imm, DL, GRLenVT));
}
 
/// Lower VECTOR_SHUFFLE whose result is the reversed source vector.
///
/// It is possible to do optimization for VECTOR_SHUFFLE performing vector
/// reverse whose mask likes:
///   <7, 6, 5, 4, 3, 2, 1, 0>
///
/// When undef's appear in the mask they are treated as if they were whatever
/// value is necessary in order to fit the above forms.
static SDValue
lowerVECTOR_SHUFFLE_IsReverse(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                              SDValue V1, SelectionDAG &DAG,
                              const LoongArchSubtarget &Subtarget) {
  // Only vectors with i8/i16 elements which cannot match other patterns
  // directly needs to do this.
  if (VT != MVT::v16i8 && VT != MVT::v8i16 && VT != MVT::v32i8 &&
      VT != MVT::v16i16)
    return SDValue();
 
  if (!ShuffleVectorInst::isReverseMask(Mask, Mask.size()))
    return SDValue();
 
  int WidenNumElts = VT.getVectorNumElements() / 4;
  SmallVector<int, 16> WidenMask(WidenNumElts, -1);
  for (int i = 0; i < WidenNumElts; ++i)
    WidenMask[i] = WidenNumElts - 1 - i;
 
  MVT WidenVT = MVT::getVectorVT(
      VT.getVectorElementType() == MVT::i8 ? MVT::i32 : MVT::i64, WidenNumElts);
  SDValue NewV1 = DAG.getBitcast(WidenVT, V1);
  SDValue WidenRev = DAG.getVectorShuffle(WidenVT, DL, NewV1,
                                          DAG.getUNDEF(WidenVT), WidenMask);
 
  return DAG.getNode(LoongArchISD::VSHUF4I, DL, VT,
                     DAG.getBitcast(VT, WidenRev),
                     DAG.getConstant(27, DL, Subtarget.getGRLenVT()));
}
 
/// Lower VECTOR_SHUFFLE into VPACKEV (if possible).
///
/// VPACKEV interleaves the even elements from each vector.
///
/// It is possible to lower into VPACKEV when the mask consists of two of the
/// following forms interleaved:
///   <0, 2, 4, ...>
///   <n, n+2, n+4, ...>
/// where n is the number of elements in the vector.
/// For example:
///   <0, 0, 2, 2, 4, 4, ...>
///   <0, n, 2, n+2, 4, n+4, ...>
///
/// When undef's appear in the mask they are treated as if they were whatever
/// value is necessary in order to fit the above forms.
static SDValue lowerVECTOR_SHUFFLE_VPACKEV(const SDLoc &DL, ArrayRef<int> Mask,
                                           MVT VT, SDValue V1, SDValue V2,
                                           SelectionDAG &DAG) {
 
  const auto &Begin = Mask.begin();
  const auto &End = Mask.end();
  SDValue OriV1 = V1, OriV2 = V2;
 
  if (fitsRegularPattern<int>(Begin, 2, End, 0, 2))
    V1 = OriV1;
  else if (fitsRegularPattern<int>(Begin, 2, End, Mask.size(), 2))
    V1 = OriV2;
  else
    return SDValue();
 
  if (fitsRegularPattern<int>(Begin + 1, 2, End, 0, 2))
    V2 = OriV1;
  else if (fitsRegularPattern<int>(Begin + 1, 2, End, Mask.size(), 2))
    V2 = OriV2;
  else
    return SDValue();
 
  return DAG.getNode(LoongArchISD::VPACKEV, DL, VT, V2, V1);
}
 
/// Lower VECTOR_SHUFFLE into VPACKOD (if possible).
///
/// VPACKOD interleaves the odd elements from each vector.
///
/// It is possible to lower into VPACKOD when the mask consists of two of the
/// following forms interleaved:
///   <1, 3, 5, ...>
///   <n+1, n+3, n+5, ...>
/// where n is the number of elements in the vector.
/// For example:
///   <1, 1, 3, 3, 5, 5, ...>
///   <1, n+1, 3, n+3, 5, n+5, ...>
///
/// When undef's appear in the mask they are treated as if they were whatever
/// value is necessary in order to fit the above forms.
static SDValue lowerVECTOR_SHUFFLE_VPACKOD(const SDLoc &DL, ArrayRef<int> Mask,
                                           MVT VT, SDValue V1, SDValue V2,
                                           SelectionDAG &DAG) {
 
  const auto &Begin = Mask.begin();
  const auto &End = Mask.end();
  SDValue OriV1 = V1, OriV2 = V2;
 
  if (fitsRegularPattern<int>(Begin, 2, End, 1, 2))
    V1 = OriV1;
  else if (fitsRegularPattern<int>(Begin, 2, End, Mask.size() + 1, 2))
    V1 = OriV2;
  else
    return SDValue();
 
  if (fitsRegularPattern<int>(Begin + 1, 2, End, 1, 2))
    V2 = OriV1;
  else if (fitsRegularPattern<int>(Begin + 1, 2, End, Mask.size() + 1, 2))
    V2 = OriV2;
  else
    return SDValue();
 
  return DAG.getNode(LoongArchISD::VPACKOD, DL, VT, V2, V1);
}
 
/// Lower VECTOR_SHUFFLE into VILVH (if possible).
///
/// VILVH interleaves consecutive elements from the left (highest-indexed) half
/// of each vector.
///
/// It is possible to lower into VILVH when the mask consists of two of the
/// following forms interleaved:
///   <x, x+1, x+2, ...>
///   <n+x, n+x+1, n+x+2, ...>
/// where n is the number of elements in the vector and x is half n.
/// For example:
///   <x, x, x+1, x+1, x+2, x+2, ...>
///   <x, n+x, x+1, n+x+1, x+2, n+x+2, ...>
///
/// When undef's appear in the mask they are treated as if they were whatever
/// value is necessary in order to fit the above forms.
static SDValue lowerVECTOR_SHUFFLE_VILVH(const SDLoc &DL, ArrayRef<int> Mask,
                                         MVT VT, SDValue V1, SDValue V2,
                                         SelectionDAG &DAG) {
 
  const auto &Begin = Mask.begin();
  const auto &End = Mask.end();
  unsigned HalfSize = Mask.size() / 2;
  SDValue OriV1 = V1, OriV2 = V2;
 
  if (fitsRegularPattern<int>(Begin, 2, End, HalfSize, 1))
    V1 = OriV1;
  else if (fitsRegularPattern<int>(Begin, 2, End, Mask.size() + HalfSize, 1))
    V1 = OriV2;
  else
    return SDValue();
 
  if (fitsRegularPattern<int>(Begin + 1, 2, End, HalfSize, 1))
    V2 = OriV1;
  else if (fitsRegularPattern<int>(Begin + 1, 2, End, Mask.size() + HalfSize,
                                   1))
    V2 = OriV2;
  else
    return SDValue();
 
  return DAG.getNode(LoongArchISD::VILVH, DL, VT, V2, V1);
}
 
/// Lower VECTOR_SHUFFLE into VILVL (if possible).
///
/// VILVL interleaves consecutive elements from the right (lowest-indexed) half
/// of each vector.
///
/// It is possible to lower into VILVL when the mask consists of two of the
/// following forms interleaved:
///   <0, 1, 2, ...>
///   <n, n+1, n+2, ...>
/// where n is the number of elements in the vector.
/// For example:
///   <0, 0, 1, 1, 2, 2, ...>
///   <0, n, 1, n+1, 2, n+2, ...>
///
/// When undef's appear in the mask they are treated as if they were whatever
/// value is necessary in order to fit the above forms.
static SDValue lowerVECTOR_SHUFFLE_VILVL(const SDLoc &DL, ArrayRef<int> Mask,
                                         MVT VT, SDValue V1, SDValue V2,
                                         SelectionDAG &DAG) {
 
  const auto &Begin = Mask.begin();
  const auto &End = Mask.end();
  SDValue OriV1 = V1, OriV2 = V2;
 
  if (fitsRegularPattern<int>(Begin, 2, End, 0, 1))
    V1 = OriV1;
  else if (fitsRegularPattern<int>(Begin, 2, End, Mask.size(), 1))
    V1 = OriV2;
  else
    return SDValue();
 
  if (fitsRegularPattern<int>(Begin + 1, 2, End, 0, 1))
    V2 = OriV1;
  else if (fitsRegularPattern<int>(Begin + 1, 2, End, Mask.size(), 1))
    V2 = OriV2;
  else
    return SDValue();
 
  return DAG.getNode(LoongArchISD::VILVL, DL, VT, V2, V1);
}
 
/// Lower VECTOR_SHUFFLE into VPICKEV (if possible).
///
/// VPICKEV copies the even elements of each vector into the result vector.
///
/// It is possible to lower into VPICKEV when the mask consists of two of the
/// following forms concatenated:
///   <0, 2, 4, ...>
///   <n, n+2, n+4, ...>
/// where n is the number of elements in the vector.
/// For example:
///   <0, 2, 4, ..., 0, 2, 4, ...>
///   <0, 2, 4, ..., n, n+2, n+4, ...>
///
/// When undef's appear in the mask they are treated as if they were whatever
/// value is necessary in order to fit the above forms.
static SDValue lowerVECTOR_SHUFFLE_VPICKEV(const SDLoc &DL, ArrayRef<int> Mask,
                                           MVT VT, SDValue V1, SDValue V2,
                                           SelectionDAG &DAG) {
 
  const auto &Begin = Mask.begin();
  const auto &Mid = Mask.begin() + Mask.size() / 2;
  const auto &End = Mask.end();
  SDValue OriV1 = V1, OriV2 = V2;
 
  if (fitsRegularPattern<int>(Begin, 1, Mid, 0, 2))
    V1 = OriV1;
  else if (fitsRegularPattern<int>(Begin, 1, Mid, Mask.size(), 2))
    V1 = OriV2;
  else
    return SDValue();
 
  if (fitsRegularPattern<int>(Mid, 1, End, 0, 2))
    V2 = OriV1;
  else if (fitsRegularPattern<int>(Mid, 1, End, Mask.size(), 2))
    V2 = OriV2;
 
  else
    return SDValue();
 
  return DAG.getNode(LoongArchISD::VPICKEV, DL, VT, V2, V1);
}
 
/// Lower VECTOR_SHUFFLE into VPICKOD (if possible).
///
/// VPICKOD copies the odd elements of each vector into the result vector.
///
/// It is possible to lower into VPICKOD when the mask consists of two of the
/// following forms concatenated:
///   <1, 3, 5, ...>
///   <n+1, n+3, n+5, ...>
/// where n is the number of elements in the vector.
/// For example:
///   <1, 3, 5, ..., 1, 3, 5, ...>
///   <1, 3, 5, ..., n+1, n+3, n+5, ...>
///
/// When undef's appear in the mask they are treated as if they were whatever
/// value is necessary in order to fit the above forms.
static SDValue lowerVECTOR_SHUFFLE_VPICKOD(const SDLoc &DL, ArrayRef<int> Mask,
                                           MVT VT, SDValue V1, SDValue V2,
                                           SelectionDAG &DAG) {
 
  const auto &Begin = Mask.begin();
  const auto &Mid = Mask.begin() + Mask.size() / 2;
  const auto &End = Mask.end();
  SDValue OriV1 = V1, OriV2 = V2;
 
  if (fitsRegularPattern<int>(Begin, 1, Mid, 1, 2))
    V1 = OriV1;
  else if (fitsRegularPattern<int>(Begin, 1, Mid, Mask.size() + 1, 2))
    V1 = OriV2;
  else
    return SDValue();
 
  if (fitsRegularPattern<int>(Mid, 1, End, 1, 2))
    V2 = OriV1;
  else if (fitsRegularPattern<int>(Mid, 1, End, Mask.size() + 1, 2))
    V2 = OriV2;
  else
    return SDValue();
 
  return DAG.getNode(LoongArchISD::VPICKOD, DL, VT, V2, V1);
}
 
/// Lower VECTOR_SHUFFLE into VEXTRINS (if possible).
///
/// VEXTRINS copies one element of a vector into any place of the result
/// vector and makes no change to the rest elements of the result vector.
///
/// It is possible to lower into VEXTRINS when the mask takes the form:
///   <0, 1, 2, ..., n+i, ..., n-1>  or  <n, n+1, n+2, ..., i, ..., 2n-1> or
///   <0, 1, 2, ..., i, ..., n-1>  or  <n, n+1, n+2, ..., n+i, ..., 2n-1>
/// where n is the number of elements in the vector and i is in [0, n).
/// For example:
///   <0, 1, 2, 3, 4, 5, 6, 8> , <2, 9, 10, 11, 12, 13, 14, 15> ,
///   <0, 1, 2, 6, 4, 5, 6, 7> , <8, 9, 10, 11, 12, 9, 14, 15>
///
/// When undef's appear in the mask they are treated as if they were whatever
/// value is necessary in order to fit the above forms.
static SDValue
lowerVECTOR_SHUFFLE_VEXTRINS(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                             SDValue V1, SDValue V2, SelectionDAG &DAG,
                             const LoongArchSubtarget &Subtarget) {
  unsigned NumElts = VT.getVectorNumElements();
  MVT EltVT = VT.getVectorElementType();
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  if (Mask.size() != NumElts)
    return SDValue();
 
  auto tryLowerToExtrAndIns = [&](unsigned Base) -> SDValue {
    int DiffCount = 0;
    int DiffPos = -1;
    for (unsigned i = 0; i < NumElts; ++i) {
      if (Mask[i] == -1)
        continue;
      if (Mask[i] != int(Base + i)) {
        ++DiffCount;
        DiffPos = int(i);
        if (DiffCount > 1)
          return SDValue();
      }
    }
 
    // Need exactly one differing element to lower into VEXTRINS.
    if (DiffCount != 1)
      return SDValue();
 
    // DiffMask must be in [0, 2N).
    int DiffMask = Mask[DiffPos];
    if (DiffMask < 0 || DiffMask >= int(2 * NumElts))
      return SDValue();
 
    // Determine source vector and source index.
    SDValue SrcVec;
    unsigned SrcIdx;
    if (unsigned(DiffMask) < NumElts) {
      SrcVec = V1;
      SrcIdx = unsigned(DiffMask);
    } else {
      SrcVec = V2;
      SrcIdx = unsigned(DiffMask) - NumElts;
    }
 
    // Replace with EXTRACT_VECTOR_ELT + INSERT_VECTOR_ELT, it will match the
    // patterns of VEXTRINS in tablegen.
    SDValue Extracted = DAG.getNode(
        ISD::EXTRACT_VECTOR_ELT, DL, EltVT.isFloatingPoint() ? EltVT : GRLenVT,
        SrcVec, DAG.getConstant(SrcIdx, DL, GRLenVT));
    SDValue Result =
        DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, (Base == 0) ? V1 : V2,
                    Extracted, DAG.getConstant(DiffPos, DL, GRLenVT));
 
    return Result;
  };
 
  // Try [0, n-1) insertion then [n, 2n-1) insertion.
  if (SDValue Result = tryLowerToExtrAndIns(0))
    return Result;
  return tryLowerToExtrAndIns(NumElts);
}
 
// Check the Mask and then build SrcVec and MaskImm infos which will
// be used to build LoongArchISD nodes for VPERMI_W or XVPERMI_W.
// On success, return true. Otherwise, return false.
static bool buildVPERMIInfo(ArrayRef<int> Mask, SDValue V1, SDValue V2,
                            SmallVectorImpl<SDValue> &SrcVec,
                            unsigned &MaskImm) {
  unsigned MaskSize = Mask.size();
 
  auto isValid = [&](int M, int Off) {
    return (M == -1) || (M >= Off && M < Off + 4);
  };
 
  auto buildImm = [&](int MLo, int MHi, unsigned Off, unsigned I) {
    auto immPart = [&](int M, unsigned Off) {
      return (M == -1 ? 0 : (M - Off)) & 0x3;
    };
    MaskImm |= immPart(MLo, Off) << (I * 2);
    MaskImm |= immPart(MHi, Off) << ((I + 1) * 2);
  };
 
  for (unsigned i = 0; i < 4; i += 2) {
    int MLo = Mask[i];
    int MHi = Mask[i + 1];
 
    if (MaskSize == 8) { // Only v8i32/v8f32 need this check.
      auto isValid2 = [&](int &M, int M2) {
        // If high half index is undef, it's always valid.
        if (M2 == -1)
          return true;
        if (M == -1) {
          // If low half index is undef, use index from high half,
          // remapped to low half.
          if ((M2 % MaskSize) < 4)
            return false;
          M = M2 - 4;
          return true;
        }
        // Index in low half must be same as index in high half.
        return M2 == M + 4;
      };
      if (!isValid2(MLo, Mask[i + 4]) || !isValid2(MHi, Mask[i + 5]))
        return false;
    }
 
    if (isValid(MLo, 0) && isValid(MHi, 0)) {
      SrcVec.push_back(V1);
      buildImm(MLo, MHi, 0, i);
    } else if (isValid(MLo, MaskSize) && isValid(MHi, MaskSize)) {
      SrcVec.push_back(V2);
      buildImm(MLo, MHi, MaskSize, i);
    } else {
      return false;
    }
  }
 
  return true;
}
 
/// Lower VECTOR_SHUFFLE into VPERMI (if possible).
///
/// VPERMI selects two elements from each of the two vectors based on the
/// mask and places them in the corresponding positions of the result vector
/// in order. Only v4i32 and v4f32 types are allowed.
///
/// It is possible to lower into VPERMI when the mask consists of two of the
/// following forms concatenated:
///   <i, j, u, v>
///   <u, v, i, j>
/// where i,j are in [0,4) and u,v are in [4, 8).
/// For example:
///   <2, 3, 4, 5>
///   <5, 7, 0, 2>
///
/// When undef's appear in the mask they are treated as if they were whatever
/// value is necessary in order to fit the above forms.
static SDValue lowerVECTOR_SHUFFLE_VPERMI(const SDLoc &DL, ArrayRef<int> Mask,
                                          MVT VT, SDValue V1, SDValue V2,
                                          SelectionDAG &DAG,
                                          const LoongArchSubtarget &Subtarget) {
  if ((VT != MVT::v4i32 && VT != MVT::v4f32) ||
      Mask.size() != VT.getVectorNumElements())
    return SDValue();
 
  SmallVector<SDValue, 2> SrcVec;
  unsigned MaskImm = 0;
  if (!buildVPERMIInfo(Mask, V1, V2, SrcVec, MaskImm))
    return SDValue();
 
  return DAG.getNode(LoongArchISD::VPERMI, DL, VT, SrcVec[1], SrcVec[0],
                     DAG.getConstant(MaskImm, DL, Subtarget.getGRLenVT()));
}
 
/// Lower VECTOR_SHUFFLE into VSHUF.
///
/// This mostly consists of converting the shuffle mask into a BUILD_VECTOR and
/// adding it as an operand to the resulting VSHUF.
static SDValue lowerVECTOR_SHUFFLE_VSHUF(const SDLoc &DL, ArrayRef<int> Mask,
                                         MVT VT, SDValue V1, SDValue V2,
                                         SelectionDAG &DAG,
                                         const LoongArchSubtarget &Subtarget) {
 
  SmallVector<SDValue, 16> Ops;
  for (auto M : Mask)
    Ops.push_back(DAG.getSignedConstant(M, DL, Subtarget.getGRLenVT()));
 
  EVT MaskVecTy = VT.changeVectorElementTypeToInteger();
  SDValue MaskVec = DAG.getBuildVector(MaskVecTy, DL, Ops);
 
  // VECTOR_SHUFFLE concatenates the vectors in an vectorwise fashion.
  // <0b00, 0b01> + <0b10, 0b11> -> <0b00, 0b01, 0b10, 0b11>
  // VSHF concatenates the vectors in a bitwise fashion:
  // <0b00, 0b01> + <0b10, 0b11> ->
  // 0b0100       + 0b1110       -> 0b01001110
  //                                <0b10, 0b11, 0b00, 0b01>
  // We must therefore swap the operands to get the correct result.
  return DAG.getNode(LoongArchISD::VSHUF, DL, VT, MaskVec, V2, V1);
}
 
/// Dispatching routine to lower various 128-bit LoongArch vector shuffles.
///
/// This routine breaks down the specific type of 128-bit shuffle and
/// dispatches to the lowering routines accordingly.
static SDValue lower128BitShuffle(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                                  SDValue V1, SDValue V2, SelectionDAG &DAG,
                                  const LoongArchSubtarget &Subtarget) {
  assert((VT.SimpleTy == MVT::v16i8 || VT.SimpleTy == MVT::v8i16 ||
          VT.SimpleTy == MVT::v4i32 || VT.SimpleTy == MVT::v2i64 ||
          VT.SimpleTy == MVT::v4f32 || VT.SimpleTy == MVT::v2f64) &&
         "Vector type is unsupported for lsx!");
  assert(V1.getSimpleValueType() == V2.getSimpleValueType() &&
         "Two operands have different types!");
  assert(VT.getVectorNumElements() == Mask.size() &&
         "Unexpected mask size for shuffle!");
  assert(Mask.size() % 2 == 0 && "Expected even mask size.");
 
  APInt KnownUndef, KnownZero;
  computeZeroableShuffleElements(Mask, V1, V2, KnownUndef, KnownZero);
  APInt Zeroable = KnownUndef | KnownZero;
 
  SDValue Result;
  // TODO: Add more comparison patterns.
  if (V2.isUndef()) {
    if ((Result =
             lowerVECTOR_SHUFFLE_VREPLVEI(DL, Mask, VT, V1, DAG, Subtarget)))
      return Result;
    if ((Result =
             lowerVECTOR_SHUFFLE_VSHUF4I(DL, Mask, VT, V1, V2, DAG, Subtarget)))
      return Result;
    if ((Result =
             lowerVECTOR_SHUFFLE_IsReverse(DL, Mask, VT, V1, DAG, Subtarget)))
      return Result;
 
    // TODO: This comment may be enabled in the future to better match the
    // pattern for instruction selection.
    /* V2 = V1; */
  }
 
  // It is recommended not to change the pattern comparison order for better
  // performance.
  if ((Result = lowerVECTOR_SHUFFLE_VPACKEV(DL, Mask, VT, V1, V2, DAG)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLE_VPACKOD(DL, Mask, VT, V1, V2, DAG)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLE_VILVH(DL, Mask, VT, V1, V2, DAG)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLE_VILVL(DL, Mask, VT, V1, V2, DAG)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLE_VPICKEV(DL, Mask, VT, V1, V2, DAG)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLE_VPICKOD(DL, Mask, VT, V1, V2, DAG)))
    return Result;
  if ((VT.SimpleTy == MVT::v2i64 || VT.SimpleTy == MVT::v2f64) &&
      (Result =
           lowerVECTOR_SHUFFLE_VSHUF4I(DL, Mask, VT, V1, V2, DAG, Subtarget)))
    return Result;
  if ((Result =
           lowerVECTOR_SHUFFLE_VEXTRINS(DL, Mask, VT, V1, V2, DAG, Subtarget)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLEAsShift(DL, Mask, VT, V1, V2, DAG, Subtarget,
                                           Zeroable)))
    return Result;
  if ((Result =
           lowerVECTOR_SHUFFLE_VPERMI(DL, Mask, VT, V1, V2, DAG, Subtarget)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLEAsZeroOrAnyExtend(DL, Mask, VT, V1, V2, DAG,
                                                     Zeroable)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLEAsByteRotate(DL, Mask, VT, V1, V2, DAG,
                                                Subtarget)))
    return Result;
  if (SDValue NewShuffle = widenShuffleMask(DL, Mask, VT, V1, V2, DAG))
    return NewShuffle;
  if ((Result =
           lowerVECTOR_SHUFFLE_VSHUF(DL, Mask, VT, V1, V2, DAG, Subtarget)))
    return Result;
  return SDValue();
}
 
/// Lower VECTOR_SHUFFLE into XVREPLVEI (if possible).
///
/// It is a XVREPLVEI when the mask is:
///   <x, x, x, ..., x+n, x+n, x+n, ...>
/// where the number of x is equal to n and n is half the length of vector.
///
/// When undef's appear in the mask they are treated as if they were whatever
/// value is necessary in order to fit the above form.
static SDValue
lowerVECTOR_SHUFFLE_XVREPLVEI(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                              SDValue V1, SelectionDAG &DAG,
                              const LoongArchSubtarget &Subtarget) {
  int SplatIndex = -1;
  for (const auto &M : Mask) {
    if (M != -1) {
      SplatIndex = M;
      break;
    }
  }
 
  if (SplatIndex == -1)
    return DAG.getUNDEF(VT);
 
  const auto &Begin = Mask.begin();
  const auto &End = Mask.end();
  int HalfSize = Mask.size() / 2;
 
  if (SplatIndex >= HalfSize)
    return SDValue();
 
  assert(SplatIndex < (int)Mask.size() && "Out of bounds mask index");
  if (fitsRegularPattern<int>(Begin, 1, End - HalfSize, SplatIndex, 0) &&
      fitsRegularPattern<int>(Begin + HalfSize, 1, End, SplatIndex + HalfSize,
                              0)) {
    return DAG.getNode(LoongArchISD::VREPLVEI, DL, VT, V1,
                       DAG.getConstant(SplatIndex, DL, Subtarget.getGRLenVT()));
  }
 
  return SDValue();
}
 
/// Lower VECTOR_SHUFFLE into XVSHUF4I (if possible).
static SDValue
lowerVECTOR_SHUFFLE_XVSHUF4I(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                             SDValue V1, SDValue V2, SelectionDAG &DAG,
                             const LoongArchSubtarget &Subtarget) {
  // XVSHUF4I_D must be handled separately because it is different from other
  // types of [X]VSHUF4I instructions.
  if (Mask.size() == 4) {
    unsigned MaskImm = 0;
    for (int i = 1; i >= 0; --i) {
      int MLo = Mask[i];
      int MHi = Mask[i + 2];
      if (!(MLo == -1 || (MLo >= 0 && MLo <= 1) || (MLo >= 4 && MLo <= 5)) ||
          !(MHi == -1 || (MHi >= 2 && MHi <= 3) || (MHi >= 6 && MHi <= 7)))
        return SDValue();
      if (MHi != -1 && MLo != -1 && MHi != MLo + 2)
        return SDValue();
 
      MaskImm <<= 2;
      if (MLo != -1)
        MaskImm |= ((MLo <= 1) ? MLo : (MLo - 2)) & 0x3;
      else if (MHi != -1)
        MaskImm |= ((MHi <= 3) ? (MHi - 2) : (MHi - 4)) & 0x3;
    }
 
    return DAG.getNode(LoongArchISD::VSHUF4I_D, DL, VT, V1, V2,
                       DAG.getConstant(MaskImm, DL, Subtarget.getGRLenVT()));
  }
 
  return lowerVECTOR_SHUFFLE_VSHUF4I(DL, Mask, VT, V1, V2, DAG, Subtarget);
}
 
/// Lower VECTOR_SHUFFLE into XVPERMI (if possible).
static SDValue
lowerVECTOR_SHUFFLE_XVPERMI(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                            SDValue V1, SDValue V2, SelectionDAG &DAG,
                            const LoongArchSubtarget &Subtarget) {
  MVT GRLenVT = Subtarget.getGRLenVT();
  unsigned MaskSize = Mask.size();
  if (MaskSize != VT.getVectorNumElements())
    return SDValue();
 
  // Consider XVPERMI_W.
  if (VT == MVT::v8i32 || VT == MVT::v8f32) {
    SmallVector<SDValue, 2> SrcVec;
    unsigned MaskImm = 0;
    if (!buildVPERMIInfo(Mask, V1, V2, SrcVec, MaskImm))
      return SDValue();
 
    return DAG.getNode(LoongArchISD::VPERMI, DL, VT, SrcVec[1], SrcVec[0],
                       DAG.getConstant(MaskImm, DL, GRLenVT));
  }
 
  // Consider XVPERMI_D.
  if (VT == MVT::v4i64 || VT == MVT::v4f64) {
    unsigned MaskImm = 0;
    for (unsigned i = 0; i < MaskSize; ++i) {
      if (Mask[i] == -1)
        continue;
      if (Mask[i] >= (int)MaskSize)
        return SDValue();
      MaskImm |= Mask[i] << (i * 2);
    }
 
    return DAG.getNode(LoongArchISD::XVPERMI, DL, VT, V1,
                       DAG.getConstant(MaskImm, DL, GRLenVT));
  }
 
  return SDValue();
}
 
/// Lower VECTOR_SHUFFLE into XVPERM (if possible).
static SDValue lowerVECTOR_SHUFFLE_XVPERM(const SDLoc &DL, ArrayRef<int> Mask,
                                          MVT VT, SDValue V1, SelectionDAG &DAG,
                                          const LoongArchSubtarget &Subtarget) {
  // LoongArch LASX only have XVPERM_W.
  if (Mask.size() != 8 || (VT != MVT::v8i32 && VT != MVT::v8f32))
    return SDValue();
 
  unsigned NumElts = VT.getVectorNumElements();
  unsigned HalfSize = NumElts / 2;
  bool FrontLo = true, FrontHi = true;
  bool BackLo = true, BackHi = true;
 
  auto inRange = [](int val, int low, int high) {
    return (val == -1) || (val >= low && val < high);
  };
 
  for (unsigned i = 0; i < HalfSize; ++i) {
    int Fronti = Mask[i];
    int Backi = Mask[i + HalfSize];
 
    FrontLo &= inRange(Fronti, 0, HalfSize);
    FrontHi &= inRange(Fronti, HalfSize, NumElts);
    BackLo &= inRange(Backi, 0, HalfSize);
    BackHi &= inRange(Backi, HalfSize, NumElts);
  }
 
  // If both the lower and upper 128-bit parts access only one half of the
  // vector (either lower or upper), avoid using xvperm.w. The latency of
  // xvperm.w(3) is higher than using xvshuf(1) and xvori(1).
  if ((FrontLo || FrontHi) && (BackLo || BackHi))
    return SDValue();
 
  SmallVector<SDValue, 8> Masks;
  MVT GRLenVT = Subtarget.getGRLenVT();
  for (unsigned i = 0; i < NumElts; ++i)
    Masks.push_back(Mask[i] == -1 ? DAG.getUNDEF(GRLenVT)
                                  : DAG.getConstant(Mask[i], DL, GRLenVT));
  SDValue MaskVec = DAG.getBuildVector(MVT::v8i32, DL, Masks);
 
  return DAG.getNode(LoongArchISD::XVPERM, DL, VT, V1, MaskVec);
}
 
/// Lower VECTOR_SHUFFLE into XVPACKEV (if possible).
static SDValue lowerVECTOR_SHUFFLE_XVPACKEV(const SDLoc &DL, ArrayRef<int> Mask,
                                            MVT VT, SDValue V1, SDValue V2,
                                            SelectionDAG &DAG) {
  return lowerVECTOR_SHUFFLE_VPACKEV(DL, Mask, VT, V1, V2, DAG);
}
 
/// Lower VECTOR_SHUFFLE into XVPACKOD (if possible).
static SDValue lowerVECTOR_SHUFFLE_XVPACKOD(const SDLoc &DL, ArrayRef<int> Mask,
                                            MVT VT, SDValue V1, SDValue V2,
                                            SelectionDAG &DAG) {
  return lowerVECTOR_SHUFFLE_VPACKOD(DL, Mask, VT, V1, V2, DAG);
}
 
/// Lower VECTOR_SHUFFLE into XVILVH (if possible).
static SDValue lowerVECTOR_SHUFFLE_XVILVH(const SDLoc &DL, ArrayRef<int> Mask,
                                          MVT VT, SDValue V1, SDValue V2,
                                          SelectionDAG &DAG) {
 
  const auto &Begin = Mask.begin();
  const auto &End = Mask.end();
  unsigned HalfSize = Mask.size() / 2;
  unsigned LeftSize = HalfSize / 2;
  SDValue OriV1 = V1, OriV2 = V2;
 
  if (fitsRegularPattern<int>(Begin, 2, End - HalfSize, HalfSize - LeftSize,
                              1) &&
      fitsRegularPattern<int>(Begin + HalfSize, 2, End, HalfSize + LeftSize, 1))
    V1 = OriV1;
  else if (fitsRegularPattern<int>(Begin, 2, End - HalfSize,
                                   Mask.size() + HalfSize - LeftSize, 1) &&
           fitsRegularPattern<int>(Begin + HalfSize, 2, End,
                                   Mask.size() + HalfSize + LeftSize, 1))
    V1 = OriV2;
  else
    return SDValue();
 
  if (fitsRegularPattern<int>(Begin + 1, 2, End - HalfSize, HalfSize - LeftSize,
                              1) &&
      fitsRegularPattern<int>(Begin + 1 + HalfSize, 2, End, HalfSize + LeftSize,
                              1))
    V2 = OriV1;
  else if (fitsRegularPattern<int>(Begin + 1, 2, End - HalfSize,
                                   Mask.size() + HalfSize - LeftSize, 1) &&
           fitsRegularPattern<int>(Begin + 1 + HalfSize, 2, End,
                                   Mask.size() + HalfSize + LeftSize, 1))
    V2 = OriV2;
  else
    return SDValue();
 
  return DAG.getNode(LoongArchISD::VILVH, DL, VT, V2, V1);
}
 
/// Lower VECTOR_SHUFFLE into XVILVL (if possible).
static SDValue lowerVECTOR_SHUFFLE_XVILVL(const SDLoc &DL, ArrayRef<int> Mask,
                                          MVT VT, SDValue V1, SDValue V2,
                                          SelectionDAG &DAG) {
 
  const auto &Begin = Mask.begin();
  const auto &End = Mask.end();
  unsigned HalfSize = Mask.size() / 2;
  SDValue OriV1 = V1, OriV2 = V2;
 
  if (fitsRegularPattern<int>(Begin, 2, End - HalfSize, 0, 1) &&
      fitsRegularPattern<int>(Begin + HalfSize, 2, End, HalfSize, 1))
    V1 = OriV1;
  else if (fitsRegularPattern<int>(Begin, 2, End - HalfSize, Mask.size(), 1) &&
           fitsRegularPattern<int>(Begin + HalfSize, 2, End,
                                   Mask.size() + HalfSize, 1))
    V1 = OriV2;
  else
    return SDValue();
 
  if (fitsRegularPattern<int>(Begin + 1, 2, End - HalfSize, 0, 1) &&
      fitsRegularPattern<int>(Begin + 1 + HalfSize, 2, End, HalfSize, 1))
    V2 = OriV1;
  else if (fitsRegularPattern<int>(Begin + 1, 2, End - HalfSize, Mask.size(),
                                   1) &&
           fitsRegularPattern<int>(Begin + 1 + HalfSize, 2, End,
                                   Mask.size() + HalfSize, 1))
    V2 = OriV2;
  else
    return SDValue();
 
  return DAG.getNode(LoongArchISD::VILVL, DL, VT, V2, V1);
}
 
/// Lower VECTOR_SHUFFLE into XVPICKEV (if possible).
static SDValue lowerVECTOR_SHUFFLE_XVPICKEV(const SDLoc &DL, ArrayRef<int> Mask,
                                            MVT VT, SDValue V1, SDValue V2,
                                            SelectionDAG &DAG) {
 
  const auto &Begin = Mask.begin();
  const auto &LeftMid = Mask.begin() + Mask.size() / 4;
  const auto &Mid = Mask.begin() + Mask.size() / 2;
  const auto &RightMid = Mask.end() - Mask.size() / 4;
  const auto &End = Mask.end();
  unsigned HalfSize = Mask.size() / 2;
  SDValue OriV1 = V1, OriV2 = V2;
 
  if (fitsRegularPattern<int>(Begin, 1, LeftMid, 0, 2) &&
      fitsRegularPattern<int>(Mid, 1, RightMid, HalfSize, 2))
    V1 = OriV1;
  else if (fitsRegularPattern<int>(Begin, 1, LeftMid, Mask.size(), 2) &&
           fitsRegularPattern<int>(Mid, 1, RightMid, Mask.size() + HalfSize, 2))
    V1 = OriV2;
  else
    return SDValue();
 
  if (fitsRegularPattern<int>(LeftMid, 1, Mid, 0, 2) &&
      fitsRegularPattern<int>(RightMid, 1, End, HalfSize, 2))
    V2 = OriV1;
  else if (fitsRegularPattern<int>(LeftMid, 1, Mid, Mask.size(), 2) &&
           fitsRegularPattern<int>(RightMid, 1, End, Mask.size() + HalfSize, 2))
    V2 = OriV2;
 
  else
    return SDValue();
 
  return DAG.getNode(LoongArchISD::VPICKEV, DL, VT, V2, V1);
}
 
/// Lower VECTOR_SHUFFLE into XVPICKOD (if possible).
static SDValue lowerVECTOR_SHUFFLE_XVPICKOD(const SDLoc &DL, ArrayRef<int> Mask,
                                            MVT VT, SDValue V1, SDValue V2,
                                            SelectionDAG &DAG) {
 
  const auto &Begin = Mask.begin();
  const auto &LeftMid = Mask.begin() + Mask.size() / 4;
  const auto &Mid = Mask.begin() + Mask.size() / 2;
  const auto &RightMid = Mask.end() - Mask.size() / 4;
  const auto &End = Mask.end();
  unsigned HalfSize = Mask.size() / 2;
  SDValue OriV1 = V1, OriV2 = V2;
 
  if (fitsRegularPattern<int>(Begin, 1, LeftMid, 1, 2) &&
      fitsRegularPattern<int>(Mid, 1, RightMid, HalfSize + 1, 2))
    V1 = OriV1;
  else if (fitsRegularPattern<int>(Begin, 1, LeftMid, Mask.size() + 1, 2) &&
           fitsRegularPattern<int>(Mid, 1, RightMid, Mask.size() + HalfSize + 1,
                                   2))
    V1 = OriV2;
  else
    return SDValue();
 
  if (fitsRegularPattern<int>(LeftMid, 1, Mid, 1, 2) &&
      fitsRegularPattern<int>(RightMid, 1, End, HalfSize + 1, 2))
    V2 = OriV1;
  else if (fitsRegularPattern<int>(LeftMid, 1, Mid, Mask.size() + 1, 2) &&
           fitsRegularPattern<int>(RightMid, 1, End, Mask.size() + HalfSize + 1,
                                   2))
    V2 = OriV2;
  else
    return SDValue();
 
  return DAG.getNode(LoongArchISD::VPICKOD, DL, VT, V2, V1);
}
 
/// Lower VECTOR_SHUFFLE into XVEXTRINS (if possible).
static SDValue
lowerVECTOR_SHUFFLE_XVEXTRINS(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                              SDValue V1, SDValue V2, SelectionDAG &DAG,
                              const LoongArchSubtarget &Subtarget) {
  int NumElts = VT.getVectorNumElements();
  int HalfSize = NumElts / 2;
  MVT EltVT = VT.getVectorElementType();
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  if ((int)Mask.size() != NumElts)
    return SDValue();
 
  auto tryLowerToExtrAndIns = [&](int Base) -> SDValue {
    SmallVector<int> DiffPos;
    for (int i = 0; i < NumElts; ++i) {
      if (Mask[i] == -1)
        continue;
      if (Mask[i] != Base + i) {
        DiffPos.push_back(i);
        if (DiffPos.size() > 2)
          return SDValue();
      }
    }
 
    // Need exactly two differing element to lower into XVEXTRINS.
    // If only one differing element, the element at a distance of
    // HalfSize from it must be undef.
    if (DiffPos.size() == 1) {
      if (DiffPos[0] < HalfSize && Mask[DiffPos[0] + HalfSize] == -1)
        DiffPos.push_back(DiffPos[0] + HalfSize);
      else if (DiffPos[0] >= HalfSize && Mask[DiffPos[0] - HalfSize] == -1)
        DiffPos.insert(DiffPos.begin(), DiffPos[0] - HalfSize);
      else
        return SDValue();
    }
    if (DiffPos.size() != 2 || DiffPos[1] != DiffPos[0] + HalfSize)
      return SDValue();
 
    // DiffMask must be in its low or high part.
    int DiffMaskLo = Mask[DiffPos[0]];
    int DiffMaskHi = Mask[DiffPos[1]];
    DiffMaskLo = DiffMaskLo == -1 ? DiffMaskHi - HalfSize : DiffMaskLo;
    DiffMaskHi = DiffMaskHi == -1 ? DiffMaskLo + HalfSize : DiffMaskHi;
    if (!(DiffMaskLo >= 0 && DiffMaskLo < HalfSize) &&
        !(DiffMaskLo >= NumElts && DiffMaskLo < NumElts + HalfSize))
      return SDValue();
    if (!(DiffMaskHi >= HalfSize && DiffMaskHi < NumElts) &&
        !(DiffMaskHi >= NumElts + HalfSize && DiffMaskHi < 2 * NumElts))
      return SDValue();
    if (DiffMaskHi != DiffMaskLo + HalfSize)
      return SDValue();
 
    // Determine source vector and source index.
    SDValue SrcVec = (DiffMaskLo < HalfSize) ? V1 : V2;
    int SrcIdxLo =
        (DiffMaskLo < HalfSize) ? DiffMaskLo : (DiffMaskLo - NumElts);
    bool IsEltFP = EltVT.isFloatingPoint();
 
    // Replace with 2*EXTRACT_VECTOR_ELT + 2*INSERT_VECTOR_ELT, it will match
    // the patterns of XVEXTRINS in tablegen.
    SDValue BaseVec = (Base == 0) ? V1 : V2;
    SDValue EltLo =
        DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, IsEltFP ? EltVT : GRLenVT,
                    SrcVec, DAG.getConstant(SrcIdxLo, DL, GRLenVT));
    SDValue InsLo = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, BaseVec, EltLo,
                                DAG.getConstant(DiffPos[0], DL, GRLenVT));
    SDValue EltHi =
        DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, IsEltFP ? EltVT : GRLenVT,
                    SrcVec, DAG.getConstant(SrcIdxLo + HalfSize, DL, GRLenVT));
    SDValue Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, InsLo, EltHi,
                                 DAG.getConstant(DiffPos[1], DL, GRLenVT));
 
    return Result;
  };
 
  // Try [0, n-1) insertion then [n, 2n-1) insertion.
  if (SDValue Result = tryLowerToExtrAndIns(0))
    return Result;
  return tryLowerToExtrAndIns(NumElts);
}
 
/// Lower VECTOR_SHUFFLE into XVINSVE0 (if possible).
static SDValue
lowerVECTOR_SHUFFLE_XVINSVE0(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                             SDValue V1, SDValue V2, SelectionDAG &DAG,
                             const LoongArchSubtarget &Subtarget) {
  // LoongArch LASX only supports xvinsve0.{w/d}.
  if (VT != MVT::v8i32 && VT != MVT::v8f32 && VT != MVT::v4i64 &&
      VT != MVT::v4f64)
    return SDValue();
 
  MVT GRLenVT = Subtarget.getGRLenVT();
  int MaskSize = Mask.size();
  assert(MaskSize == (int)VT.getVectorNumElements() && "Unexpected mask size");
 
  // Check if exactly one element of the Mask is replaced by 'Replaced', while
  // all other elements are either 'Base + i' or undef (-1). On success, return
  // the index of the replaced element. Otherwise, just return -1.
  auto checkReplaceOne = [&](int Base, int Replaced) -> int {
    int Idx = -1;
    for (int i = 0; i < MaskSize; ++i) {
      if (Mask[i] == Base + i || Mask[i] == -1)
        continue;
      if (Mask[i] != Replaced)
        return -1;
      if (Idx == -1)
        Idx = i;
      else
        return -1;
    }
    return Idx;
  };
 
  // Case 1: the lowest element of V2 replaces one element in V1.
  int Idx = checkReplaceOne(0, MaskSize);
  if (Idx != -1)
    return DAG.getNode(LoongArchISD::XVINSVE0, DL, VT, V1, V2,
                       DAG.getConstant(Idx, DL, GRLenVT));
 
  // Case 2: the lowest element of V1 replaces one element in V2.
  Idx = checkReplaceOne(MaskSize, 0);
  if (Idx != -1)
    return DAG.getNode(LoongArchISD::XVINSVE0, DL, VT, V2, V1,
                       DAG.getConstant(Idx, DL, GRLenVT));
 
  return SDValue();
}
 
/// Lower VECTOR_SHUFFLE into XVSHUF (if possible).
static SDValue lowerVECTOR_SHUFFLE_XVSHUF(const SDLoc &DL, ArrayRef<int> Mask,
                                          MVT VT, SDValue V1, SDValue V2,
                                          SelectionDAG &DAG) {
 
  int MaskSize = Mask.size();
  int HalfSize = Mask.size() / 2;
  const auto &Begin = Mask.begin();
  const auto &Mid = Mask.begin() + HalfSize;
  const auto &End = Mask.end();
 
  // VECTOR_SHUFFLE concatenates the vectors:
  //  <0, 1, 2, 3, 4, 5, 6, 7> + <8, 9, 10, 11, 12, 13, 14, 15>
  //  shuffling ->
  //  <0, 1, 2, 3, 8, 9, 10, 11> <4, 5, 6, 7, 12, 13, 14, 15>
  //
  // XVSHUF concatenates the vectors:
  //  <a0, a1, a2, a3, b0, b1, b2, b3> + <a4, a5, a6, a7, b4, b5, b6, b7>
  //  shuffling ->
  //  <a0, a1, a2, a3, a4, a5, a6, a7> + <b0, b1, b2, b3, b4, b5, b6, b7>
  SmallVector<SDValue, 8> MaskAlloc;
  for (auto it = Begin; it < Mid; it++) {
    if (*it < 0) // UNDEF
      MaskAlloc.push_back(DAG.getTargetConstant(0, DL, MVT::i64));
    else if ((*it >= 0 && *it < HalfSize) ||
             (*it >= MaskSize && *it < MaskSize + HalfSize)) {
      int M = *it < HalfSize ? *it : *it - HalfSize;
      MaskAlloc.push_back(DAG.getTargetConstant(M, DL, MVT::i64));
    } else
      return SDValue();
  }
  assert((int)MaskAlloc.size() == HalfSize && "xvshuf convert failed!");
 
  for (auto it = Mid; it < End; it++) {
    if (*it < 0) // UNDEF
      MaskAlloc.push_back(DAG.getTargetConstant(0, DL, MVT::i64));
    else if ((*it >= HalfSize && *it < MaskSize) ||
             (*it >= MaskSize + HalfSize && *it < MaskSize * 2)) {
      int M = *it < MaskSize ? *it - HalfSize : *it - MaskSize;
      MaskAlloc.push_back(DAG.getTargetConstant(M, DL, MVT::i64));
    } else
      return SDValue();
  }
  assert((int)MaskAlloc.size() == MaskSize && "xvshuf convert failed!");
 
  EVT MaskVecTy = VT.changeVectorElementTypeToInteger();
  SDValue MaskVec = DAG.getBuildVector(MaskVecTy, DL, MaskAlloc);
  return DAG.getNode(LoongArchISD::VSHUF, DL, VT, MaskVec, V2, V1);
}
 
/// Shuffle vectors by lane to generate more optimized instructions.
/// 256-bit shuffles are always considered as 2-lane 128-bit shuffles.
///
/// Therefore, except for the following four cases, other cases are regarded
/// as cross-lane shuffles, where optimization is relatively limited.
///
/// - Shuffle high, low lanes of two inputs vector
///   <0, 1, 2, 3> + <4, 5, 6, 7> --- <0, 5, 3, 6>
/// - Shuffle low, high lanes of two inputs vector
///   <0, 1, 2, 3> + <4, 5, 6, 7> --- <3, 6, 0, 5>
/// - Shuffle low, low lanes of two inputs vector
///   <0, 1, 2, 3> + <4, 5, 6, 7> --- <3, 6, 3, 6>
/// - Shuffle high, high lanes of two inputs vector
///   <0, 1, 2, 3> + <4, 5, 6, 7> --- <0, 5, 0, 5>
///
/// The first case is the closest to LoongArch instructions and the other
/// cases need to be converted to it for processing.
///
/// This function will return true for the last three cases above and will
/// modify V1, V2 and Mask. Otherwise, return false for the first case and
/// cross-lane shuffle cases.
static bool canonicalizeShuffleVectorByLane(
    const SDLoc &DL, MutableArrayRef<int> Mask, MVT VT, SDValue &V1,
    SDValue &V2, SelectionDAG &DAG, const LoongArchSubtarget &Subtarget) {
 
  enum HalfMaskType { HighLaneTy, LowLaneTy, None };
 
  int MaskSize = Mask.size();
  int HalfSize = Mask.size() / 2;
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  HalfMaskType preMask = None, postMask = None;
 
  if (std::all_of(Mask.begin(), Mask.begin() + HalfSize, [&](int M) {
        return M < 0 || (M >= 0 && M < HalfSize) ||
               (M >= MaskSize && M < MaskSize + HalfSize);
      }))
    preMask = HighLaneTy;
  else if (std::all_of(Mask.begin(), Mask.begin() + HalfSize, [&](int M) {
             return M < 0 || (M >= HalfSize && M < MaskSize) ||
                    (M >= MaskSize + HalfSize && M < MaskSize * 2);
           }))
    preMask = LowLaneTy;
 
  if (std::all_of(Mask.begin() + HalfSize, Mask.end(), [&](int M) {
        return M < 0 || (M >= HalfSize && M < MaskSize) ||
               (M >= MaskSize + HalfSize && M < MaskSize * 2);
      }))
    postMask = LowLaneTy;
  else if (std::all_of(Mask.begin() + HalfSize, Mask.end(), [&](int M) {
             return M < 0 || (M >= 0 && M < HalfSize) ||
                    (M >= MaskSize && M < MaskSize + HalfSize);
           }))
    postMask = HighLaneTy;
 
  // The pre-half of mask is high lane type, and the post-half of mask
  // is low lane type, which is closest to the LoongArch instructions.
  //
  // Note: In the LoongArch architecture, the high lane of mask corresponds
  // to the lower 128-bit of vector register, and the low lane of mask
  // corresponds the higher 128-bit of vector register.
  if (preMask == HighLaneTy && postMask == LowLaneTy) {
    return false;
  }
  if (preMask == LowLaneTy && postMask == HighLaneTy) {
    V1 = DAG.getBitcast(MVT::v4i64, V1);
    V1 = DAG.getNode(LoongArchISD::XVPERMI, DL, MVT::v4i64, V1,
                     DAG.getConstant(0b01001110, DL, GRLenVT));
    V1 = DAG.getBitcast(VT, V1);
 
    if (!V2.isUndef()) {
      V2 = DAG.getBitcast(MVT::v4i64, V2);
      V2 = DAG.getNode(LoongArchISD::XVPERMI, DL, MVT::v4i64, V2,
                       DAG.getConstant(0b01001110, DL, GRLenVT));
      V2 = DAG.getBitcast(VT, V2);
    }
 
    for (auto it = Mask.begin(); it < Mask.begin() + HalfSize; it++) {
      *it = *it < 0 ? *it : *it - HalfSize;
    }
    for (auto it = Mask.begin() + HalfSize; it < Mask.end(); it++) {
      *it = *it < 0 ? *it : *it + HalfSize;
    }
  } else if (preMask == LowLaneTy && postMask == LowLaneTy) {
    V1 = DAG.getBitcast(MVT::v4i64, V1);
    V1 = DAG.getNode(LoongArchISD::XVPERMI, DL, MVT::v4i64, V1,
                     DAG.getConstant(0b11101110, DL, GRLenVT));
    V1 = DAG.getBitcast(VT, V1);
 
    if (!V2.isUndef()) {
      V2 = DAG.getBitcast(MVT::v4i64, V2);
      V2 = DAG.getNode(LoongArchISD::XVPERMI, DL, MVT::v4i64, V2,
                       DAG.getConstant(0b11101110, DL, GRLenVT));
      V2 = DAG.getBitcast(VT, V2);
    }
 
    for (auto it = Mask.begin(); it < Mask.begin() + HalfSize; it++) {
      *it = *it < 0 ? *it : *it - HalfSize;
    }
  } else if (preMask == HighLaneTy && postMask == HighLaneTy) {
    V1 = DAG.getBitcast(MVT::v4i64, V1);
    V1 = DAG.getNode(LoongArchISD::XVPERMI, DL, MVT::v4i64, V1,
                     DAG.getConstant(0b01000100, DL, GRLenVT));
    V1 = DAG.getBitcast(VT, V1);
 
    if (!V2.isUndef()) {
      V2 = DAG.getBitcast(MVT::v4i64, V2);
      V2 = DAG.getNode(LoongArchISD::XVPERMI, DL, MVT::v4i64, V2,
                       DAG.getConstant(0b01000100, DL, GRLenVT));
      V2 = DAG.getBitcast(VT, V2);
    }
 
    for (auto it = Mask.begin() + HalfSize; it < Mask.end(); it++) {
      *it = *it < 0 ? *it : *it + HalfSize;
    }
  } else { // cross-lane
    return false;
  }
 
  return true;
}
 
/// Lower VECTOR_SHUFFLE as lane permute and then shuffle (if possible).
/// Only for 256-bit vector.
///
/// For example:
/// %2 = shufflevector <4 x i64> %0, <4 x i64> posion,
///                    <4 x i64> <i32 0, i32 3, i32 2, i32 0>
/// is lowerded to:
///     (XVPERMI $xr2, $xr0, 78)
///     (XVSHUF  $xr1, $xr2, $xr0)
///     (XVORI   $xr0, $xr1, 0)
static SDValue lowerVECTOR_SHUFFLEAsLanePermuteAndShuffle(const SDLoc &DL,
                                                          ArrayRef<int> Mask,
                                                          MVT VT, SDValue V1,
                                                          SDValue V2,
                                                          SelectionDAG &DAG) {
  assert(VT.is256BitVector() && "Only for 256-bit vector shuffles!");
  int Size = Mask.size();
  int LaneSize = Size / 2;
 
  bool LaneCrossing[2] = {false, false};
  for (int i = 0; i < Size; ++i)
    if (Mask[i] >= 0 && ((Mask[i] % Size) / LaneSize) != (i / LaneSize))
      LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
 
  // Ensure that all lanes ared involved.
  if (!LaneCrossing[0] && !LaneCrossing[1])
    return SDValue();
 
  SmallVector<int> InLaneMask;
  InLaneMask.assign(Mask.begin(), Mask.end());
  for (int i = 0; i < Size; ++i) {
    int &M = InLaneMask[i];
    if (M < 0)
      continue;
    if (((M % Size) / LaneSize) != (i / LaneSize))
      M = (M % LaneSize) + ((i / LaneSize) * LaneSize) + Size;
  }
 
  SDValue Flipped = DAG.getBitcast(MVT::v4i64, V1);
  Flipped = DAG.getVectorShuffle(MVT::v4i64, DL, Flipped,
                                 DAG.getUNDEF(MVT::v4i64), {2, 3, 0, 1});
  Flipped = DAG.getBitcast(VT, Flipped);
  return DAG.getVectorShuffle(VT, DL, V1, Flipped, InLaneMask);
}
 
/// Dispatching routine to lower various 256-bit LoongArch vector shuffles.
///
/// This routine breaks down the specific type of 256-bit shuffle and
/// dispatches to the lowering routines accordingly.
static SDValue lower256BitShuffle(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                                  SDValue V1, SDValue V2, SelectionDAG &DAG,
                                  const LoongArchSubtarget &Subtarget) {
  assert((VT.SimpleTy == MVT::v32i8 || VT.SimpleTy == MVT::v16i16 ||
          VT.SimpleTy == MVT::v8i32 || VT.SimpleTy == MVT::v4i64 ||
          VT.SimpleTy == MVT::v8f32 || VT.SimpleTy == MVT::v4f64) &&
         "Vector type is unsupported for lasx!");
  assert(V1.getSimpleValueType() == V2.getSimpleValueType() &&
         "Two operands have different types!");
  assert(VT.getVectorNumElements() == Mask.size() &&
         "Unexpected mask size for shuffle!");
  assert(Mask.size() % 2 == 0 && "Expected even mask size.");
  assert(Mask.size() >= 4 && "Mask size is less than 4.");
 
  APInt KnownUndef, KnownZero;
  computeZeroableShuffleElements(Mask, V1, V2, KnownUndef, KnownZero);
  APInt Zeroable = KnownUndef | KnownZero;
 
  SDValue Result;
  // TODO: Add more comparison patterns.
  if (V2.isUndef()) {
    if ((Result =
             lowerVECTOR_SHUFFLE_XVREPLVEI(DL, Mask, VT, V1, DAG, Subtarget)))
      return Result;
    if ((Result = lowerVECTOR_SHUFFLE_XVSHUF4I(DL, Mask, VT, V1, V2, DAG,
                                               Subtarget)))
      return Result;
    // Try to widen vectors to gain more optimization opportunities.
    if (SDValue NewShuffle = widenShuffleMask(DL, Mask, VT, V1, V2, DAG))
      return NewShuffle;
    if ((Result =
             lowerVECTOR_SHUFFLE_XVPERMI(DL, Mask, VT, V1, V2, DAG, Subtarget)))
      return Result;
    if ((Result = lowerVECTOR_SHUFFLE_XVPERM(DL, Mask, VT, V1, DAG, Subtarget)))
      return Result;
    if ((Result =
             lowerVECTOR_SHUFFLE_IsReverse(DL, Mask, VT, V1, DAG, Subtarget)))
      return Result;
 
    // TODO: This comment may be enabled in the future to better match the
    // pattern for instruction selection.
    /* V2 = V1; */
  }
 
  // It is recommended not to change the pattern comparison order for better
  // performance.
  if ((Result = lowerVECTOR_SHUFFLE_XVPACKEV(DL, Mask, VT, V1, V2, DAG)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLE_XVPACKOD(DL, Mask, VT, V1, V2, DAG)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLE_XVILVH(DL, Mask, VT, V1, V2, DAG)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLE_XVILVL(DL, Mask, VT, V1, V2, DAG)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLE_XVPICKEV(DL, Mask, VT, V1, V2, DAG)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLE_XVPICKOD(DL, Mask, VT, V1, V2, DAG)))
    return Result;
  if ((VT.SimpleTy == MVT::v4i64 || VT.SimpleTy == MVT::v4f64) &&
      (Result =
           lowerVECTOR_SHUFFLE_XVSHUF4I(DL, Mask, VT, V1, V2, DAG, Subtarget)))
    return Result;
  if ((Result =
           lowerVECTOR_SHUFFLE_XVEXTRINS(DL, Mask, VT, V1, V2, DAG, Subtarget)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLEAsShift(DL, Mask, VT, V1, V2, DAG, Subtarget,
                                           Zeroable)))
    return Result;
  if ((Result =
           lowerVECTOR_SHUFFLE_XVPERMI(DL, Mask, VT, V1, V2, DAG, Subtarget)))
    return Result;
  if ((Result =
           lowerVECTOR_SHUFFLE_XVINSVE0(DL, Mask, VT, V1, V2, DAG, Subtarget)))
    return Result;
  if ((Result = lowerVECTOR_SHUFFLEAsByteRotate(DL, Mask, VT, V1, V2, DAG,
                                                Subtarget)))
    return Result;
 
  // canonicalize non cross-lane shuffle vector
  SmallVector<int> NewMask(Mask);
  if (canonicalizeShuffleVectorByLane(DL, NewMask, VT, V1, V2, DAG, Subtarget))
    return lower256BitShuffle(DL, NewMask, VT, V1, V2, DAG, Subtarget);
 
  // FIXME: Handling the remaining cases earlier can degrade performance
  // in some situations. Further analysis is required to enable more
  // effective optimizations.
  if (V2.isUndef()) {
    if ((Result = lowerVECTOR_SHUFFLEAsLanePermuteAndShuffle(DL, NewMask, VT,
                                                             V1, V2, DAG)))
      return Result;
  }
 
  if (SDValue NewShuffle = widenShuffleMask(DL, NewMask, VT, V1, V2, DAG))
    return NewShuffle;
  if ((Result = lowerVECTOR_SHUFFLE_XVSHUF(DL, NewMask, VT, V1, V2, DAG)))
    return Result;
 
  return SDValue();
}
 
SDValue LoongArchTargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
                                                     SelectionDAG &DAG) const {
  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
  ArrayRef<int> OrigMask = SVOp->getMask();
  SDValue V1 = Op.getOperand(0);
  SDValue V2 = Op.getOperand(1);
  MVT VT = Op.getSimpleValueType();
  int NumElements = VT.getVectorNumElements();
  SDLoc DL(Op);
 
  bool V1IsUndef = V1.isUndef();
  bool V2IsUndef = V2.isUndef();
  if (V1IsUndef && V2IsUndef)
    return DAG.getUNDEF(VT);
 
  // When we create a shuffle node we put the UNDEF node to second operand,
  // but in some cases the first operand may be transformed to UNDEF.
  // In this case we should just commute the node.
  if (V1IsUndef)
    return DAG.getCommutedVectorShuffle(*SVOp);
 
  // Check for non-undef masks pointing at an undef vector and make the masks
  // undef as well. This makes it easier to match the shuffle based solely on
  // the mask.
  if (V2IsUndef &&
      any_of(OrigMask, [NumElements](int M) { return M >= NumElements; })) {
    SmallVector<int, 8> NewMask(OrigMask);
    for (int &M : NewMask)
      if (M >= NumElements)
        M = -1;
    return DAG.getVectorShuffle(VT, DL, V1, V2, NewMask);
  }
 
  // Check for illegal shuffle mask element index values.
  int MaskUpperLimit = OrigMask.size() * (V2IsUndef ? 1 : 2);
  (void)MaskUpperLimit;
  assert(llvm::all_of(OrigMask,
                      [&](int M) { return -1 <= M && M < MaskUpperLimit; }) &&
         "Out of bounds shuffle index");
 
  // For each vector width, delegate to a specialized lowering routine.
  if (VT.is128BitVector())
    return lower128BitShuffle(DL, OrigMask, VT, V1, V2, DAG, Subtarget);
 
  if (VT.is256BitVector())
    return lower256BitShuffle(DL, OrigMask, VT, V1, V2, DAG, Subtarget);
 
  return SDValue();
}
 
SDValue LoongArchTargetLowering::lowerFP_TO_FP16(SDValue Op,
                                                 SelectionDAG &DAG) const {
  // Custom lower to ensure the libcall return is passed in an FPR on hard
  // float ABIs.
  SDLoc DL(Op);
  MakeLibCallOptions CallOptions;
  SDValue Op0 = Op.getOperand(0);
  SDValue Chain = SDValue();
  RTLIB::Libcall LC = RTLIB::getFPROUND(Op0.getValueType(), MVT::f16);
  SDValue Res;
  std::tie(Res, Chain) =
      makeLibCall(DAG, LC, MVT::f32, Op0, CallOptions, DL, Chain);
  if (Subtarget.is64Bit())
    return DAG.getNode(LoongArchISD::MOVFR2GR_S_LA64, DL, MVT::i64, Res);
  return DAG.getBitcast(MVT::i32, Res);
}
 
SDValue LoongArchTargetLowering::lowerFP16_TO_FP(SDValue Op,
                                                 SelectionDAG &DAG) const {
  // Custom lower to ensure the libcall argument is passed in an FPR on hard
  // float ABIs.
  SDLoc DL(Op);
  MakeLibCallOptions CallOptions;
  SDValue Op0 = Op.getOperand(0);
  SDValue Chain = SDValue();
  SDValue Arg = Subtarget.is64Bit() ? DAG.getNode(LoongArchISD::MOVGR2FR_W_LA64,
                                                  DL, MVT::f32, Op0)
                                    : DAG.getBitcast(MVT::f32, Op0);
  SDValue Res;
  std::tie(Res, Chain) = makeLibCall(DAG, RTLIB::FPEXT_F16_F32, MVT::f32, Arg,
                                     CallOptions, DL, Chain);
  return Res;
}
 
SDValue LoongArchTargetLowering::lowerFP_TO_BF16(SDValue Op,
                                                 SelectionDAG &DAG) const {
  assert(Subtarget.hasBasicF() && "Unexpected custom legalization");
  SDLoc DL(Op);
  MakeLibCallOptions CallOptions;
  RTLIB::Libcall LC =
      RTLIB::getFPROUND(Op.getOperand(0).getValueType(), MVT::bf16);
  SDValue Res =
      makeLibCall(DAG, LC, MVT::f32, Op.getOperand(0), CallOptions, DL).first;
  if (Subtarget.is64Bit())
    return DAG.getNode(LoongArchISD::MOVFR2GR_S_LA64, DL, MVT::i64, Res);
  return DAG.getBitcast(MVT::i32, Res);
}
 
SDValue LoongArchTargetLowering::lowerBF16_TO_FP(SDValue Op,
                                                 SelectionDAG &DAG) const {
  assert(Subtarget.hasBasicF() && "Unexpected custom legalization");
  MVT VT = Op.getSimpleValueType();
  SDLoc DL(Op);
  Op = DAG.getNode(
      ISD::SHL, DL, Op.getOperand(0).getValueType(), Op.getOperand(0),
      DAG.getShiftAmountConstant(16, Op.getOperand(0).getValueType(), DL));
  SDValue Res = Subtarget.is64Bit() ? DAG.getNode(LoongArchISD::MOVGR2FR_W_LA64,
                                                  DL, MVT::f32, Op)
                                    : DAG.getBitcast(MVT::f32, Op);
  if (VT != MVT::f32)
    return DAG.getNode(ISD::FP_EXTEND, DL, VT, Res);
  return Res;
}
 
// Lower BUILD_VECTOR as broadcast load (if possible).
// For example:
//   %a = load i8, ptr %ptr
//   %b = build_vector %a, %a, %a, %a
// is lowered to :
//   (VLDREPL_B $a0, 0)
static SDValue lowerBUILD_VECTORAsBroadCastLoad(BuildVectorSDNode *BVOp,
                                                const SDLoc &DL,
                                                SelectionDAG &DAG) {
  MVT VT = BVOp->getSimpleValueType(0);
  int NumOps = BVOp->getNumOperands();
 
  assert((VT.is128BitVector() || VT.is256BitVector()) &&
         "Unsupported vector type for broadcast.");
 
  SDValue IdentitySrc;
  bool IsIdeneity = true;
 
  for (int i = 0; i != NumOps; i++) {
    SDValue Op = BVOp->getOperand(i);
    if (Op.getOpcode() != ISD::LOAD || (IdentitySrc && Op != IdentitySrc)) {
      IsIdeneity = false;
      break;
    }
    IdentitySrc = BVOp->getOperand(0);
  }
 
  // make sure that this load is valid and only has one user.
  if (!IsIdeneity || !IdentitySrc || !BVOp->isOnlyUserOf(IdentitySrc.getNode()))
    return SDValue();
 
  auto *LN = cast<LoadSDNode>(IdentitySrc);
  auto ExtType = LN->getExtensionType();
 
  if ((ExtType == ISD::EXTLOAD || ExtType == ISD::NON_EXTLOAD) &&
      VT.getScalarSizeInBits() == LN->getMemoryVT().getScalarSizeInBits()) {
    // Indexed loads and stores are not supported on LoongArch.
    assert(LN->isUnindexed() && "Unexpected indexed load.");
 
    SDVTList Tys = DAG.getVTList(VT, MVT::Other);
    // The offset operand of unindexed load is always undefined, so there is
    // no need to pass it to VLDREPL.
    SDValue Ops[] = {LN->getChain(), LN->getBasePtr()};
    SDValue BCast = DAG.getNode(LoongArchISD::VLDREPL, DL, Tys, Ops);
    DAG.ReplaceAllUsesOfValueWith(SDValue(LN, 1), BCast.getValue(1));
    return BCast;
  }
  return SDValue();
}
 
// Sequentially insert elements from Ops into Vector, from low to high indices.
// Note: Ops can have fewer elements than Vector.
static void fillVector(ArrayRef<SDValue> Ops, SelectionDAG &DAG, SDLoc DL,
                       const LoongArchSubtarget &Subtarget, SDValue &Vector,
                       EVT ResTy) {
  assert(Ops.size() <= ResTy.getVectorNumElements());
 
  SDValue Op0 = Ops[0];
  if (!Op0.isUndef())
    Vector = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ResTy, Op0);
  for (unsigned i = 1; i < Ops.size(); ++i) {
    SDValue Opi = Ops[i];
    if (Opi.isUndef())
      continue;
    Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Vector, Opi,
                         DAG.getConstant(i, DL, Subtarget.getGRLenVT()));
  }
}
 
// Build a ResTy subvector from Node, taking NumElts elements starting at index
// 'first'.
static SDValue fillSubVectorFromBuildVector(BuildVectorSDNode *Node,
                                            SelectionDAG &DAG, SDLoc DL,
                                            const LoongArchSubtarget &Subtarget,
                                            EVT ResTy, unsigned first) {
  unsigned NumElts = ResTy.getVectorNumElements();
 
  assert(first + NumElts <= Node->getSimpleValueType(0).getVectorNumElements());
 
  SmallVector<SDValue, 16> Ops(Node->op_begin() + first,
                               Node->op_begin() + first + NumElts);
  SDValue Vector = DAG.getUNDEF(ResTy);
  fillVector(Ops, DAG, DL, Subtarget, Vector, ResTy);
  return Vector;
}
 
SDValue LoongArchTargetLowering::lowerBUILD_VECTOR(SDValue Op,
                                                   SelectionDAG &DAG) const {
  BuildVectorSDNode *Node = cast<BuildVectorSDNode>(Op);
  MVT VT = Node->getSimpleValueType(0);
  EVT ResTy = Op->getValueType(0);
  unsigned NumElts = ResTy.getVectorNumElements();
  SDLoc DL(Op);
  APInt SplatValue, SplatUndef;
  unsigned SplatBitSize;
  bool HasAnyUndefs;
  bool IsConstant = false;
  bool UseSameConstant = true;
  SDValue ConstantValue;
  bool Is128Vec = ResTy.is128BitVector();
  bool Is256Vec = ResTy.is256BitVector();
 
  if ((!Subtarget.hasExtLSX() || !Is128Vec) &&
      (!Subtarget.hasExtLASX() || !Is256Vec))
    return SDValue();
 
  if (SDValue Result = lowerBUILD_VECTORAsBroadCastLoad(Node, DL, DAG))
    return Result;
 
  if (Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs,
                            /*MinSplatBits=*/8) &&
      SplatBitSize <= 64) {
    // We can only cope with 8, 16, 32, or 64-bit elements.
    if (SplatBitSize != 8 && SplatBitSize != 16 && SplatBitSize != 32 &&
        SplatBitSize != 64)
      return SDValue();
 
    if (SplatBitSize == 64 && !Subtarget.is64Bit()) {
      // We can only handle 64-bit elements that are within
      // the signed 10-bit range or match vldi patterns on 32-bit targets.
      // See the BUILD_VECTOR case in LoongArchDAGToDAGISel::Select().
      if (!SplatValue.isSignedIntN(10) &&
          !isImmVLDILegalForMode1(SplatValue, SplatBitSize).first)
        return SDValue();
      if ((Is128Vec && ResTy == MVT::v4i32) ||
          (Is256Vec && ResTy == MVT::v8i32))
        return Op;
    }
 
    EVT ViaVecTy;
 
    switch (SplatBitSize) {
    default:
      return SDValue();
    case 8:
      ViaVecTy = Is128Vec ? MVT::v16i8 : MVT::v32i8;
      break;
    case 16:
      ViaVecTy = Is128Vec ? MVT::v8i16 : MVT::v16i16;
      break;
    case 32:
      ViaVecTy = Is128Vec ? MVT::v4i32 : MVT::v8i32;
      break;
    case 64:
      ViaVecTy = Is128Vec ? MVT::v2i64 : MVT::v4i64;
      break;
    }
 
    // SelectionDAG::getConstant will promote SplatValue appropriately.
    SDValue Result = DAG.getConstant(SplatValue, DL, ViaVecTy);
 
    // Bitcast to the type we originally wanted.
    if (ViaVecTy != ResTy)
      Result = DAG.getNode(ISD::BITCAST, SDLoc(Node), ResTy, Result);
 
    return Result;
  }
 
  if (DAG.isSplatValue(Op, /*AllowUndefs=*/false))
    return Op;
 
  for (unsigned i = 0; i < NumElts; ++i) {
    SDValue Opi = Node->getOperand(i);
    if (isIntOrFPConstant(Opi)) {
      IsConstant = true;
      if (!ConstantValue.getNode())
        ConstantValue = Opi;
      else if (ConstantValue != Opi)
        UseSameConstant = false;
    }
  }
 
  // If the type of BUILD_VECTOR is v2f64, custom legalizing it has no benefits.
  if (IsConstant && UseSameConstant && ResTy != MVT::v2f64) {
    SDValue Result = DAG.getSplatBuildVector(ResTy, DL, ConstantValue);
    for (unsigned i = 0; i < NumElts; ++i) {
      SDValue Opi = Node->getOperand(i);
      if (!isIntOrFPConstant(Opi))
        Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Result, Opi,
                             DAG.getConstant(i, DL, Subtarget.getGRLenVT()));
    }
    return Result;
  }
 
  if (!IsConstant) {
    // If the BUILD_VECTOR has a repeated pattern, use INSERT_VECTOR_ELT to fill
    // the sub-sequence of the vector and then broadcast the sub-sequence.
    //
    // TODO: If the BUILD_VECTOR contains undef elements, consider falling
    // back to use INSERT_VECTOR_ELT to materialize the vector, because it
    // generates worse code in some cases. This could be further optimized
    // with more consideration.
    SmallVector<SDValue> Sequence;
    BitVector UndefElements;
    if (Node->getRepeatedSequence(Sequence, &UndefElements) &&
        UndefElements.count() == 0) {
      // Using LSX instructions to fill the sub-sequence of 256-bits vector,
      // because the high part can be simply treated as undef.
      SDValue Vector = DAG.getUNDEF(ResTy);
      EVT FillTy = Is256Vec
                       ? ResTy.getHalfNumVectorElementsVT(*DAG.getContext())
                       : ResTy;
      SDValue FillVec =
          Is256Vec ? DAG.getExtractSubvector(DL, FillTy, Vector, 0) : Vector;
 
      fillVector(Sequence, DAG, DL, Subtarget, FillVec, FillTy);
 
      unsigned SeqLen = Sequence.size();
      unsigned SplatLen = NumElts / SeqLen;
      MVT SplatEltTy = MVT::getIntegerVT(VT.getScalarSizeInBits() * SeqLen);
      MVT SplatTy = MVT::getVectorVT(SplatEltTy, SplatLen);
 
      // If size of the sub-sequence is half of a 256-bits vector, bitcast the
      // vector to v4i64 type in order to match the pattern of XVREPLVE0Q.
      if (SplatEltTy == MVT::i128)
        SplatTy = MVT::v4i64;
 
      SDValue SplatVec;
      SDValue SrcVec = DAG.getBitcast(
          SplatTy,
          Is256Vec ? DAG.getInsertSubvector(DL, Vector, FillVec, 0) : FillVec);
      if (Is256Vec) {
        SplatVec =
            DAG.getNode((SplatEltTy == MVT::i128) ? LoongArchISD::XVREPLVE0Q
                                                  : LoongArchISD::XVREPLVE0,
                        DL, SplatTy, SrcVec);
      } else {
        SplatVec = DAG.getNode(LoongArchISD::VREPLVEI, DL, SplatTy, SrcVec,
                               DAG.getConstant(0, DL, Subtarget.getGRLenVT()));
      }
 
      return DAG.getBitcast(ResTy, SplatVec);
    }
 
    // Use INSERT_VECTOR_ELT operations rather than expand to stores, because
    // using memory operations is much lower.
    //
    // For 256-bit vectors, normally split into two halves and concatenate.
    // Special case: for v8i32/v8f32/v4i64/v4f64, if the upper half has only
    // one non-undef element, skip spliting to avoid a worse result.
    if (ResTy == MVT::v8i32 || ResTy == MVT::v8f32 || ResTy == MVT::v4i64 ||
        ResTy == MVT::v4f64) {
      unsigned NonUndefCount = 0;
      for (unsigned i = NumElts / 2; i < NumElts; ++i) {
        if (!Node->getOperand(i).isUndef()) {
          ++NonUndefCount;
          if (NonUndefCount > 1)
            break;
        }
      }
      if (NonUndefCount == 1)
        return fillSubVectorFromBuildVector(Node, DAG, DL, Subtarget, ResTy, 0);
    }
 
    EVT VecTy =
        Is256Vec ? ResTy.getHalfNumVectorElementsVT(*DAG.getContext()) : ResTy;
    SDValue Vector =
        fillSubVectorFromBuildVector(Node, DAG, DL, Subtarget, VecTy, 0);
 
    if (Is128Vec)
      return Vector;
 
    SDValue VectorHi = fillSubVectorFromBuildVector(Node, DAG, DL, Subtarget,
                                                    VecTy, NumElts / 2);
 
    return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResTy, Vector, VectorHi);
  }
 
  return SDValue();
}
 
SDValue LoongArchTargetLowering::lowerCONCAT_VECTORS(SDValue Op,
                                                     SelectionDAG &DAG) const {
  SDLoc DL(Op);
  MVT ResVT = Op.getSimpleValueType();
  assert(ResVT.is256BitVector() && Op.getNumOperands() == 2);
 
  if (Op.getOperand(0).getOpcode() == ISD::TRUNCATE &&
      Op.getOperand(1).getOpcode() == ISD::TRUNCATE)
    return Op;
 
  unsigned NumOperands = Op.getNumOperands();
  unsigned NumFreezeUndef = 0;
  unsigned NumZero = 0;
  unsigned NumNonZero = 0;
  unsigned NonZeros = 0;
  SmallSet<SDValue, 4> Undefs;
  for (unsigned i = 0; i != NumOperands; ++i) {
    SDValue SubVec = Op.getOperand(i);
    if (SubVec.isUndef())
      continue;
    if (ISD::isFreezeUndef(SubVec.getNode())) {
      // If the freeze(undef) has multiple uses then we must fold to zero.
      if (SubVec.hasOneUse()) {
        ++NumFreezeUndef;
      } else {
        ++NumZero;
        Undefs.insert(SubVec);
      }
    } else if (ISD::isBuildVectorAllZeros(SubVec.getNode()))
      ++NumZero;
    else {
      assert(i < sizeof(NonZeros) * CHAR_BIT); // Ensure the shift is in range.
      NonZeros |= 1 << i;
      ++NumNonZero;
    }
  }
 
  // If we have more than 2 non-zeros, build each half separately.
  if (NumNonZero > 2) {
    MVT HalfVT = ResVT.getHalfNumVectorElementsVT();
    ArrayRef<SDUse> Ops = Op->ops();
    SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, DL, HalfVT,
                             Ops.slice(0, NumOperands / 2));
    SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, DL, HalfVT,
                             Ops.slice(NumOperands / 2));
    return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
  }
 
  // Otherwise, build it up through insert_subvectors.
  SDValue Vec = NumZero ? DAG.getConstant(0, DL, ResVT)
                        : (NumFreezeUndef ? DAG.getFreeze(DAG.getUNDEF(ResVT))
                                          : DAG.getUNDEF(ResVT));
 
  // Replace Undef operands with ZeroVector.
  for (SDValue U : Undefs)
    DAG.ReplaceAllUsesWith(U, DAG.getConstant(0, DL, U.getSimpleValueType()));
 
  MVT SubVT = Op.getOperand(0).getSimpleValueType();
  unsigned NumSubElems = SubVT.getVectorNumElements();
  for (unsigned i = 0; i != NumOperands; ++i) {
    if ((NonZeros & (1 << i)) == 0)
      continue;
 
    Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ResVT, Vec, Op.getOperand(i),
                      DAG.getVectorIdxConstant(i * NumSubElems, DL));
  }
 
  return Vec;
}
 
SDValue
LoongArchTargetLowering::lowerEXTRACT_VECTOR_ELT(SDValue Op,
                                                 SelectionDAG &DAG) const {
  MVT EltVT = Op.getSimpleValueType();
  SDValue Vec = Op->getOperand(0);
  EVT VecTy = Vec->getValueType(0);
  SDValue Idx = Op->getOperand(1);
  SDLoc DL(Op);
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  assert(VecTy.is256BitVector() && "Unexpected EXTRACT_VECTOR_ELT vector type");
 
  if (isa<ConstantSDNode>(Idx))
    return Op;
 
  switch (VecTy.getSimpleVT().SimpleTy) {
  default:
    llvm_unreachable("Unexpected type");
  case MVT::v32i8:
  case MVT::v16i16:
  case MVT::v4i64:
  case MVT::v4f64: {
    // Extract the high half subvector and place it to the low half of a new
    // vector. It doesn't matter what the high half of the new vector is.
    EVT HalfTy = VecTy.getHalfNumVectorElementsVT(*DAG.getContext());
    SDValue VecHi =
        DAG.getExtractSubvector(DL, HalfTy, Vec, HalfTy.getVectorNumElements());
    SDValue TmpVec =
        DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VecTy, DAG.getUNDEF(VecTy),
                    VecHi, DAG.getConstant(0, DL, GRLenVT));
 
    // Shuffle the origin Vec and the TmpVec using MaskVec, the lowest element
    // of MaskVec is Idx, the rest do not matter. ResVec[0] will hold the
    // desired element.
    SDValue IdxCp =
        Subtarget.is64Bit()
            ? DAG.getNode(LoongArchISD::MOVGR2FR_W_LA64, DL, MVT::f32, Idx)
            : DAG.getBitcast(MVT::f32, Idx);
    SDValue IdxVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v8f32, IdxCp);
    SDValue MaskVec =
        DAG.getBitcast((VecTy == MVT::v4f64) ? MVT::v4i64 : VecTy, IdxVec);
    SDValue ResVec =
        DAG.getNode(LoongArchISD::VSHUF, DL, VecTy, MaskVec, TmpVec, Vec);
 
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ResVec,
                       DAG.getConstant(0, DL, GRLenVT));
  }
  case MVT::v8i32:
  case MVT::v8f32: {
    SDValue SplatIdx = DAG.getSplatBuildVector(MVT::v8i32, DL, Idx);
    SDValue SplatValue =
        DAG.getNode(LoongArchISD::XVPERM, DL, VecTy, Vec, SplatIdx);
 
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, SplatValue,
                       DAG.getConstant(0, DL, GRLenVT));
  }
  }
}
 
SDValue
LoongArchTargetLowering::lowerINSERT_VECTOR_ELT(SDValue Op,
                                                SelectionDAG &DAG) const {
  MVT VT = Op.getSimpleValueType();
  MVT EltVT = VT.getVectorElementType();
  unsigned NumElts = VT.getVectorNumElements();
  unsigned EltSizeInBits = EltVT.getScalarSizeInBits();
  SDLoc DL(Op);
  SDValue Op0 = Op.getOperand(0);
  SDValue Op1 = Op.getOperand(1);
  SDValue Op2 = Op.getOperand(2);
 
  if (isa<ConstantSDNode>(Op2))
    return Op;
 
  MVT IdxTy = MVT::getIntegerVT(EltSizeInBits);
  MVT IdxVTy = MVT::getVectorVT(IdxTy, NumElts);
 
  if (!isTypeLegal(VT) || !isTypeLegal(IdxVTy))
    return SDValue();
 
  SDValue SplatElt = DAG.getSplatBuildVector(VT, DL, Op1);
  SmallVector<SDValue, 32> RawIndices;
  SDValue SplatIdx;
  SDValue Indices;
 
  if (!Subtarget.is64Bit() && IdxTy == MVT::i64) {
    MVT PairVTy = MVT::getVectorVT(MVT::i32, NumElts * 2);
    for (unsigned i = 0; i < NumElts; ++i) {
      RawIndices.push_back(Op2);
      RawIndices.push_back(DAG.getConstant(0, DL, MVT::i32));
    }
    SplatIdx = DAG.getBuildVector(PairVTy, DL, RawIndices);
    SplatIdx = DAG.getBitcast(IdxVTy, SplatIdx);
 
    RawIndices.clear();
    for (unsigned i = 0; i < NumElts; ++i) {
      RawIndices.push_back(DAG.getConstant(i, DL, MVT::i32));
      RawIndices.push_back(DAG.getConstant(0, DL, MVT::i32));
    }
    Indices = DAG.getBuildVector(PairVTy, DL, RawIndices);
    Indices = DAG.getBitcast(IdxVTy, Indices);
  } else {
    SplatIdx = DAG.getSplatBuildVector(IdxVTy, DL, Op2);
 
    for (unsigned i = 0; i < NumElts; ++i)
      RawIndices.push_back(DAG.getConstant(i, DL, Subtarget.getGRLenVT()));
    Indices = DAG.getBuildVector(IdxVTy, DL, RawIndices);
  }
 
  // insert vec, elt, idx
  // =>
  // select (splatidx == {0,1,2...}) ? splatelt : vec
  SDValue SelectCC =
      DAG.getSetCC(DL, IdxVTy, SplatIdx, Indices, ISD::CondCode::SETEQ);
  return DAG.getNode(ISD::VSELECT, DL, VT, SelectCC, SplatElt, Op0);
}
 
SDValue LoongArchTargetLowering::lowerATOMIC_FENCE(SDValue Op,
                                                   SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SyncScope::ID FenceSSID =
      static_cast<SyncScope::ID>(Op.getConstantOperandVal(2));
 
  // singlethread fences only synchronize with signal handlers on the same
  // thread and thus only need to preserve instruction order, not actually
  // enforce memory ordering.
  if (FenceSSID == SyncScope::SingleThread)
    // MEMBARRIER is a compiler barrier; it codegens to a no-op.
    return DAG.getNode(ISD::MEMBARRIER, DL, MVT::Other, Op.getOperand(0));
 
  return Op;
}
 
static SDValue convertRMEncoding(SelectionDAG &DAG, const SDLoc &DL,
                                 MVT GRLenVT, SDValue RMValue) {
  // LLVM rounding mode encoding differs from LoongArch FCSR encoding:
  //   LLVM: 0=RTZ, 1=RNE, 2=RUP, 3=RDN
  //   FCSR: 0=RNE, 1=RZ,  2=RP,  3=RN
  //
  // The conversion swaps encodings 0 and 1 while preserving 2 and 3.
  // Since the transformation is self-inverse, it applies in both directions:
  //   LLVM RM <-> LoongArch FCSR RM
  //
  // Transformation: RM ^ (~(RM >> 1) & 1)
  SDValue ShiftRight1 = DAG.getNode(ISD::SRL, DL, GRLenVT, RMValue,
                                    DAG.getConstant(1, DL, GRLenVT));
 
  SDValue SwapMask = DAG.getNode(ISD::AND, DL, GRLenVT,
                                 DAG.getNode(ISD::XOR, DL, GRLenVT, ShiftRight1,
                                             DAG.getConstant(1, DL, GRLenVT)),
                                 DAG.getConstant(1, DL, GRLenVT));
 
  return DAG.getNode(ISD::XOR, DL, GRLenVT, RMValue, SwapMask);
}
 
SDValue LoongArchTargetLowering::lowerSET_ROUNDING(SDValue Op,
                                                   SelectionDAG &DAG) const {
  MVT GRLenVT = Subtarget.getGRLenVT();
  SDLoc DL(Op);
  SDValue Chain = Op.getOperand(0);
  SDValue RMValue = Op.getOperand(1);
 
  if (auto *CVal = dyn_cast<ConstantSDNode>(RMValue)) {
    uint64_t RM = CVal->getZExtValue();
    if (RM > 3) {
      MachineFunction &MF = DAG.getMachineFunction();
      LLVMContext &C = MF.getFunction().getContext();
      C.diagnose(DiagnosticInfoUnsupported(
          MF.getFunction(),
          "rounding mode is not supported by LoongArch hardware",
          DiagnosticLocation(DL.getDebugLoc()), DS_Error));
      return Chain;
    }
  }
 
  RMValue = DAG.getNode(ISD::ANY_EXTEND, DL, GRLenVT, RMValue);
  RMValue = convertRMEncoding(DAG, DL, GRLenVT, RMValue);
 
  // The RM field in FCSR is at bits [9:8]. Shift the rounding mode value
  // into position before writing via WRFCSR.
  RMValue = DAG.getNode(ISD::SHL, DL, GRLenVT, RMValue,
                        DAG.getConstant(8, DL, GRLenVT));
 
  // FCSR3 is an alias of the RM field; writing it avoids clobbering
  // unrelated fields in FCSR0.
  SDValue FCSRNo = DAG.getTargetConstant(3, DL, GRLenVT);
  MachineSDNode *RN = DAG.getMachineNode(LoongArch::WRFCSR, DL, MVT::Other,
                                         FCSRNo, RMValue, Chain);
  return SDValue(RN, 0);
}
 
SDValue LoongArchTargetLowering::lowerGET_ROUNDING(SDValue Op,
                                                   SelectionDAG &DAG) const {
  MVT GRLenVT = Subtarget.getGRLenVT();
  SDLoc DL(Op);
  SDValue Chain = Op->getOperand(0);
 
  // FCSR3 is an alias of the RM field.
  SDValue FCSRNo = DAG.getTargetConstant(3, DL, GRLenVT);
  MachineSDNode *FCSR = DAG.getMachineNode(LoongArch::RDFCSR, DL, GRLenVT,
                                           MVT::Other, FCSRNo, Chain);
  SDValue RMValue = SDValue(FCSR, 0);
  Chain = SDValue(FCSR, 1);
 
  // The RM field in FCSR is at bits [9:8].
  RMValue = DAG.getNode(ISD::SRL, DL, GRLenVT, RMValue,
                        DAG.getConstant(8, DL, GRLenVT));
  RMValue = convertRMEncoding(DAG, DL, GRLenVT, RMValue);
 
  SDValue RetVal = DAG.getZExtOrTrunc(RMValue, DL, Op.getValueType());
  return DAG.getMergeValues({RetVal, Chain}, DL);
}
 
SDValue LoongArchTargetLowering::lowerWRITE_REGISTER(SDValue Op,
                                                     SelectionDAG &DAG) const {
 
  if (Subtarget.is64Bit() && Op.getOperand(2).getValueType() == MVT::i32) {
    DAG.getContext()->emitError(
        "On LA64, only 64-bit registers can be written.");
    return Op.getOperand(0);
  }
 
  if (!Subtarget.is64Bit() && Op.getOperand(2).getValueType() == MVT::i64) {
    DAG.getContext()->emitError(
        "On LA32, only 32-bit registers can be written.");
    return Op.getOperand(0);
  }
 
  return Op;
}
 
SDValue LoongArchTargetLowering::lowerFRAMEADDR(SDValue Op,
                                                SelectionDAG &DAG) const {
  if (!isa<ConstantSDNode>(Op.getOperand(0))) {
    DAG.getContext()->emitError("argument to '__builtin_frame_address' must "
                                "be a constant integer");
    return SDValue();
  }
 
  MachineFunction &MF = DAG.getMachineFunction();
  MF.getFrameInfo().setFrameAddressIsTaken(true);
  Register FrameReg = Subtarget.getRegisterInfo()->getFrameRegister(MF);
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, FrameReg, VT);
  unsigned Depth = Op.getConstantOperandVal(0);
  int GRLenInBytes = Subtarget.getGRLen() / 8;
 
  while (Depth--) {
    int Offset = -(GRLenInBytes * 2);
    SDValue Ptr = DAG.getNode(ISD::ADD, DL, VT, FrameAddr,
                              DAG.getSignedConstant(Offset, DL, VT));
    FrameAddr =
        DAG.getLoad(VT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo());
  }
  return FrameAddr;
}
 
SDValue LoongArchTargetLowering::lowerRETURNADDR(SDValue Op,
                                                 SelectionDAG &DAG) const {
  // Currently only support lowering return address for current frame.
  if (Op.getConstantOperandVal(0) != 0) {
    DAG.getContext()->emitError(
        "return address can only be determined for the current frame");
    return SDValue();
  }
 
  MachineFunction &MF = DAG.getMachineFunction();
  MF.getFrameInfo().setReturnAddressIsTaken(true);
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  // Return the value of the return address register, marking it an implicit
  // live-in.
  Register Reg = MF.addLiveIn(Subtarget.getRegisterInfo()->getRARegister(),
                              getRegClassFor(GRLenVT));
  return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, GRLenVT);
}
 
SDValue LoongArchTargetLowering::lowerEH_DWARF_CFA(SDValue Op,
                                                   SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  auto Size = Subtarget.getGRLen() / 8;
  auto FI = MF.getFrameInfo().CreateFixedObject(Size, 0, false);
  return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
}
 
SDValue LoongArchTargetLowering::lowerVASTART(SDValue Op,
                                              SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  auto *FuncInfo = MF.getInfo<LoongArchMachineFunctionInfo>();
 
  SDLoc DL(Op);
  SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
                                 getPointerTy(MF.getDataLayout()));
 
  // vastart just stores the address of the VarArgsFrameIndex slot into the
  // memory location argument.
  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
  return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
                      MachinePointerInfo(SV));
}
 
SDValue LoongArchTargetLowering::lowerUINT_TO_FP(SDValue Op,
                                                 SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SDValue Op0 = Op.getOperand(0);
  EVT VT = Op.getValueType();
  EVT Op0VT = Op0.getValueType();
 
  if (VT.isVector()) {
    if (VT.getScalarSizeInBits() != Op0VT.getScalarSizeInBits())
      return SDValue();
    return Op;
  }
 
  if ((DAG.SignBitIsZero(Op0) || Op->getFlags().hasNonNeg()) &&
      !isOperationLegal(ISD::UINT_TO_FP, Op0VT) &&
      isOperationLegal(ISD::SINT_TO_FP, Op0VT))
    return DAG.getNode(ISD::SINT_TO_FP, DL, VT, Op0);
 
  // We can't do uint64 -> double -> float because of double-rounding issue.
  if (Subtarget.hasExtLSX() && Op0VT == MVT::i64 && VT == MVT::f64) {
    Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2i64, Op0);
    SDValue Conv = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::v2f64, Op0);
    Conv = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f64, Conv,
                       DAG.getIntPtrConstant(0, DL));
    return Conv;
  }
 
  if (!Subtarget.is64Bit() || !Subtarget.hasBasicF() || Subtarget.hasBasicD())
    return SDValue();
 
  assert(Subtarget.is64Bit() && Subtarget.hasBasicF() &&
         !Subtarget.hasBasicD() && "unexpected target features");
 
  if (Op0->getOpcode() == ISD::AND) {
    auto *C = dyn_cast<ConstantSDNode>(Op0.getOperand(1));
    if (C && C->getZExtValue() < UINT64_C(0xFFFFFFFF))
      return Op;
  }
 
  if (Op0->getOpcode() == LoongArchISD::BSTRPICK &&
      Op0.getConstantOperandVal(1) < UINT64_C(0X1F) &&
      Op0.getConstantOperandVal(2) == UINT64_C(0))
    return Op;
 
  if (Op0.getOpcode() == ISD::AssertZext &&
      dyn_cast<VTSDNode>(Op0.getOperand(1))->getVT().bitsLT(MVT::i32))
    return Op;
 
  EVT OpVT = Op0.getValueType();
  EVT RetVT = Op.getValueType();
  RTLIB::Libcall LC = RTLIB::getUINTTOFP(OpVT, RetVT);
  MakeLibCallOptions CallOptions;
  CallOptions.setTypeListBeforeSoften(OpVT, RetVT);
  SDValue Chain = SDValue();
  SDValue Result;
  std::tie(Result, Chain) =
      makeLibCall(DAG, LC, Op.getValueType(), Op0, CallOptions, DL, Chain);
  return Result;
}
 
SDValue LoongArchTargetLowering::lowerSINT_TO_FP(SDValue Op,
                                                 SelectionDAG &DAG) const {
  assert(Subtarget.is64Bit() && Subtarget.hasBasicF() &&
         !Subtarget.hasBasicD() && "unexpected target features");
 
  SDLoc DL(Op);
  SDValue Op0 = Op.getOperand(0);
 
  if ((Op0.getOpcode() == ISD::AssertSext ||
       Op0.getOpcode() == ISD::SIGN_EXTEND_INREG) &&
      dyn_cast<VTSDNode>(Op0.getOperand(1))->getVT().bitsLE(MVT::i32))
    return Op;
 
  EVT OpVT = Op0.getValueType();
  EVT RetVT = Op.getValueType();
  RTLIB::Libcall LC = RTLIB::getSINTTOFP(OpVT, RetVT);
  MakeLibCallOptions CallOptions;
  CallOptions.setTypeListBeforeSoften(OpVT, RetVT);
  SDValue Chain = SDValue();
  SDValue Result;
  std::tie(Result, Chain) =
      makeLibCall(DAG, LC, Op.getValueType(), Op0, CallOptions, DL, Chain);
  return Result;
}
 
SDValue LoongArchTargetLowering::lowerBITCAST(SDValue Op,
                                              SelectionDAG &DAG) const {
 
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  SDValue Op0 = Op.getOperand(0);
  EVT Op0VT = Op0.getValueType();
 
  if (Op.getValueType() == MVT::f32 && Op0VT == MVT::i32 &&
      Subtarget.is64Bit() && Subtarget.hasBasicF()) {
    SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op0);
    return DAG.getNode(LoongArchISD::MOVGR2FR_W_LA64, DL, MVT::f32, NewOp0);
  }
  if (VT == MVT::f64 && Op0VT == MVT::i64 && !Subtarget.is64Bit()) {
    SDValue Lo, Hi;
    std::tie(Lo, Hi) = DAG.SplitScalar(Op0, DL, MVT::i32, MVT::i32);
    return DAG.getNode(LoongArchISD::BUILD_PAIR_F64, DL, MVT::f64, Lo, Hi);
  }
  return Op;
}
 
SDValue LoongArchTargetLowering::lowerFP_TO_SINT(SDValue Op,
                                                 SelectionDAG &DAG) const {
 
  SDLoc DL(Op);
  SDValue Op0 = Op.getOperand(0);
 
  if (Op0.getValueType() == MVT::f16)
    Op0 = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Op0);
 
  if (Op.getValueSizeInBits() > 32 && Subtarget.hasBasicF() &&
      !Subtarget.hasBasicD()) {
    SDValue Dst = DAG.getNode(LoongArchISD::FTINT, DL, MVT::f32, Op0);
    return DAG.getNode(LoongArchISD::MOVFR2GR_S_LA64, DL, MVT::i64, Dst);
  }
 
  EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits());
  SDValue Trunc = DAG.getNode(LoongArchISD::FTINT, DL, FPTy, Op0);
  return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Trunc);
}
 
SDValue LoongArchTargetLowering::lowerFP_TO_UINT(SDValue Op,
                                                 SelectionDAG &DAG) const {
  if (!Subtarget.hasExtLSX())
    return SDValue();
 
  SDLoc DL(Op);
  SDValue Src = Op.getOperand(0);
  EVT VT = Op.getValueType();
  EVT SrcVT = Src.getValueType();
 
  if (VT != MVT::i64)
    return SDValue();
 
  if (SrcVT != MVT::f32 && SrcVT != MVT::f64)
    return SDValue();
 
  if (SrcVT == MVT::f32)
    Src = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f64, Src);
  Src = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, Src);
  SDValue Conv = DAG.getNode(ISD::FP_TO_UINT, DL, MVT::v2i64, Src);
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Conv,
                     DAG.getIntPtrConstant(0, DL));
}
 
static SDValue getTargetNode(GlobalAddressSDNode *N, SDLoc DL, EVT Ty,
                             SelectionDAG &DAG, unsigned Flags) {
  return DAG.getTargetGlobalAddress(N->getGlobal(), DL, Ty, 0, Flags);
}
 
static SDValue getTargetNode(BlockAddressSDNode *N, SDLoc DL, EVT Ty,
                             SelectionDAG &DAG, unsigned Flags) {
  return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, N->getOffset(),
                                   Flags);
}
 
static SDValue getTargetNode(ConstantPoolSDNode *N, SDLoc DL, EVT Ty,
                             SelectionDAG &DAG, unsigned Flags) {
  return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlign(),
                                   N->getOffset(), Flags);
}
 
static SDValue getTargetNode(JumpTableSDNode *N, SDLoc DL, EVT Ty,
                             SelectionDAG &DAG, unsigned Flags) {
  return DAG.getTargetJumpTable(N->getIndex(), Ty, Flags);
}
 
template <class NodeTy>
SDValue LoongArchTargetLowering::getAddr(NodeTy *N, SelectionDAG &DAG,
                                         CodeModel::Model M,
                                         bool IsLocal) const {
  SDLoc DL(N);
  EVT Ty = getPointerTy(DAG.getDataLayout());
  SDValue Addr = getTargetNode(N, DL, Ty, DAG, 0);
  SDValue Load;
 
  switch (M) {
  default:
    report_fatal_error("Unsupported code model");
 
  case CodeModel::Large: {
    assert(Subtarget.is64Bit() && "Large code model requires LA64");
 
    // This is not actually used, but is necessary for successfully matching
    // the PseudoLA_*_LARGE nodes.
    SDValue Tmp = DAG.getConstant(0, DL, Ty);
    if (IsLocal) {
      // This generates the pattern (PseudoLA_PCREL_LARGE tmp sym), that
      // eventually becomes the desired 5-insn code sequence.
      Load = SDValue(DAG.getMachineNode(LoongArch::PseudoLA_PCREL_LARGE, DL, Ty,
                                        Tmp, Addr),
                     0);
    } else {
      // This generates the pattern (PseudoLA_GOT_LARGE tmp sym), that
      // eventually becomes the desired 5-insn code sequence.
      Load = SDValue(
          DAG.getMachineNode(LoongArch::PseudoLA_GOT_LARGE, DL, Ty, Tmp, Addr),
          0);
    }
    break;
  }
 
  case CodeModel::Small:
  case CodeModel::Medium:
    if (IsLocal) {
      // This generates the pattern (PseudoLA_PCREL sym), which
      //
      // for la32r expands to:
      //   (addi.w (pcaddu12i %pcadd_hi20(sym)) %pcadd_lo12(.Lpcadd_hi)).
      //
      // for la32s and la64 expands to:
      //   (addi.w/d (pcalau12i %pc_hi20(sym)) %pc_lo12(sym)).
      Load = SDValue(
          DAG.getMachineNode(LoongArch::PseudoLA_PCREL, DL, Ty, Addr), 0);
    } else {
      // This generates the pattern (PseudoLA_GOT sym), which
      //
      // for la32r expands to:
      //   (ld.w (pcaddu12i %got_pcadd_hi20(sym)) %pcadd_lo12(.Lpcadd_hi)).
      //
      // for la32s and la64 expands to:
      //   (ld.w/d (pcalau12i %got_pc_hi20(sym)) %got_pc_lo12(sym)).
      Load =
          SDValue(DAG.getMachineNode(LoongArch::PseudoLA_GOT, DL, Ty, Addr), 0);
    }
  }
 
  if (!IsLocal) {
    // Mark the load instruction as invariant to enable hoisting in MachineLICM.
    MachineFunction &MF = DAG.getMachineFunction();
    MachineMemOperand *MemOp = MF.getMachineMemOperand(
        MachinePointerInfo::getGOT(MF),
        MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
            MachineMemOperand::MOInvariant,
        LLT(Ty.getSimpleVT()), Align(Ty.getFixedSizeInBits() / 8));
    DAG.setNodeMemRefs(cast<MachineSDNode>(Load.getNode()), {MemOp});
  }
 
  return Load;
}
 
SDValue LoongArchTargetLowering::lowerBlockAddress(SDValue Op,
                                                   SelectionDAG &DAG) const {
  return getAddr(cast<BlockAddressSDNode>(Op), DAG,
                 DAG.getTarget().getCodeModel());
}
 
SDValue LoongArchTargetLowering::lowerJumpTable(SDValue Op,
                                                SelectionDAG &DAG) const {
  return getAddr(cast<JumpTableSDNode>(Op), DAG,
                 DAG.getTarget().getCodeModel());
}
 
SDValue LoongArchTargetLowering::lowerConstantPool(SDValue Op,
                                                   SelectionDAG &DAG) const {
  return getAddr(cast<ConstantPoolSDNode>(Op), DAG,
                 DAG.getTarget().getCodeModel());
}
 
SDValue LoongArchTargetLowering::lowerGlobalAddress(SDValue Op,
                                                    SelectionDAG &DAG) const {
  GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
  assert(N->getOffset() == 0 && "unexpected offset in global node");
  auto CM = DAG.getTarget().getCodeModel();
  const GlobalValue *GV = N->getGlobal();
 
  if (GV->isDSOLocal() && isa<GlobalVariable>(GV)) {
    if (auto GCM = dyn_cast<GlobalVariable>(GV)->getCodeModel())
      CM = *GCM;
  }
 
  return getAddr(N, DAG, CM, GV->isDSOLocal());
}
 
SDValue LoongArchTargetLowering::getStaticTLSAddr(GlobalAddressSDNode *N,
                                                  SelectionDAG &DAG,
                                                  unsigned Opc, bool UseGOT,
                                                  bool Large) const {
  SDLoc DL(N);
  EVT Ty = getPointerTy(DAG.getDataLayout());
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  // This is not actually used, but is necessary for successfully matching the
  // PseudoLA_*_LARGE nodes.
  SDValue Tmp = DAG.getConstant(0, DL, Ty);
  SDValue Addr = DAG.getTargetGlobalAddress(N->getGlobal(), DL, Ty, 0, 0);
 
  // Only IE needs an extra argument for large code model.
  SDValue Offset = Opc == LoongArch::PseudoLA_TLS_IE_LARGE
                       ? SDValue(DAG.getMachineNode(Opc, DL, Ty, Tmp, Addr), 0)
                       : SDValue(DAG.getMachineNode(Opc, DL, Ty, Addr), 0);
 
  // If it is LE for normal/medium code model, the add tp operation will occur
  // during the pseudo-instruction expansion.
  if (Opc == LoongArch::PseudoLA_TLS_LE && !Large)
    return Offset;
 
  if (UseGOT) {
    // Mark the load instruction as invariant to enable hoisting in MachineLICM.
    MachineFunction &MF = DAG.getMachineFunction();
    MachineMemOperand *MemOp = MF.getMachineMemOperand(
        MachinePointerInfo::getGOT(MF),
        MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
            MachineMemOperand::MOInvariant,
        LLT(Ty.getSimpleVT()), Align(Ty.getFixedSizeInBits() / 8));
    DAG.setNodeMemRefs(cast<MachineSDNode>(Offset.getNode()), {MemOp});
  }
 
  // Add the thread pointer.
  return DAG.getNode(ISD::ADD, DL, Ty, Offset,
                     DAG.getRegister(LoongArch::R2, GRLenVT));
}
 
SDValue LoongArchTargetLowering::getDynamicTLSAddr(GlobalAddressSDNode *N,
                                                   SelectionDAG &DAG,
                                                   unsigned Opc,
                                                   bool Large) const {
  SDLoc DL(N);
  EVT Ty = getPointerTy(DAG.getDataLayout());
  IntegerType *CallTy = Type::getIntNTy(*DAG.getContext(), Ty.getSizeInBits());
 
  // This is not actually used, but is necessary for successfully matching the
  // PseudoLA_*_LARGE nodes.
  SDValue Tmp = DAG.getConstant(0, DL, Ty);
 
  // Use a PC-relative addressing mode to access the dynamic GOT address.
  SDValue Addr = DAG.getTargetGlobalAddress(N->getGlobal(), DL, Ty, 0, 0);
  SDValue Load = Large ? SDValue(DAG.getMachineNode(Opc, DL, Ty, Tmp, Addr), 0)
                       : SDValue(DAG.getMachineNode(Opc, DL, Ty, Addr), 0);
 
  // Prepare argument list to generate call.
  ArgListTy Args;
  Args.emplace_back(Load, CallTy);
 
  // Setup call to __tls_get_addr.
  TargetLowering::CallLoweringInfo CLI(DAG);
  CLI.setDebugLoc(DL)
      .setChain(DAG.getEntryNode())
      .setLibCallee(CallingConv::C, CallTy,
                    DAG.getExternalSymbol("__tls_get_addr", Ty),
                    std::move(Args));
 
  return LowerCallTo(CLI).first;
}
 
SDValue LoongArchTargetLowering::getTLSDescAddr(GlobalAddressSDNode *N,
                                                SelectionDAG &DAG, unsigned Opc,
                                                bool Large) const {
  SDLoc DL(N);
  EVT Ty = getPointerTy(DAG.getDataLayout());
  const GlobalValue *GV = N->getGlobal();
 
  // This is not actually used, but is necessary for successfully matching the
  // PseudoLA_*_LARGE nodes.
  SDValue Tmp = DAG.getConstant(0, DL, Ty);
 
  // Use a PC-relative addressing mode to access the global dynamic GOT address.
  // This generates the pattern (PseudoLA_TLS_DESC_PC{,LARGE} sym).
  SDValue Addr = DAG.getTargetGlobalAddress(GV, DL, Ty, 0, 0);
  return Large ? SDValue(DAG.getMachineNode(Opc, DL, Ty, Tmp, Addr), 0)
               : SDValue(DAG.getMachineNode(Opc, DL, Ty, Addr), 0);
}
 
SDValue
LoongArchTargetLowering::lowerGlobalTLSAddress(SDValue Op,
                                               SelectionDAG &DAG) const {
  if (DAG.getMachineFunction().getFunction().getCallingConv() ==
      CallingConv::GHC)
    report_fatal_error("In GHC calling convention TLS is not supported");
 
  bool Large = DAG.getTarget().getCodeModel() == CodeModel::Large;
  assert((!Large || Subtarget.is64Bit()) && "Large code model requires LA64");
 
  GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
  assert(N->getOffset() == 0 && "unexpected offset in global node");
 
  if (DAG.getTarget().useEmulatedTLS())
    reportFatalUsageError("the emulated TLS is prohibited");
 
  bool IsDesc = DAG.getTarget().useTLSDESC();
 
  switch (getTargetMachine().getTLSModel(N->getGlobal())) {
  case TLSModel::GeneralDynamic:
    // In this model, application code calls the dynamic linker function
    // __tls_get_addr to locate TLS offsets into the dynamic thread vector at
    // runtime.
    if (!IsDesc)
      return getDynamicTLSAddr(N, DAG,
                               Large ? LoongArch::PseudoLA_TLS_GD_LARGE
                                     : LoongArch::PseudoLA_TLS_GD,
                               Large);
    break;
  case TLSModel::LocalDynamic:
    // Same as GeneralDynamic, except for assembly modifiers and relocation
    // records.
    if (!IsDesc)
      return getDynamicTLSAddr(N, DAG,
                               Large ? LoongArch::PseudoLA_TLS_LD_LARGE
                                     : LoongArch::PseudoLA_TLS_LD,
                               Large);
    break;
  case TLSModel::InitialExec:
    // This model uses the GOT to resolve TLS offsets.
    return getStaticTLSAddr(N, DAG,
                            Large ? LoongArch::PseudoLA_TLS_IE_LARGE
                                  : LoongArch::PseudoLA_TLS_IE,
                            /*UseGOT=*/true, Large);
  case TLSModel::LocalExec:
    // This model is used when static linking as the TLS offsets are resolved
    // during program linking.
    //
    // This node doesn't need an extra argument for the large code model.
    return getStaticTLSAddr(N, DAG, LoongArch::PseudoLA_TLS_LE,
                            /*UseGOT=*/false, Large);
  }
 
  return getTLSDescAddr(N, DAG,
                        Large ? LoongArch::PseudoLA_TLS_DESC_LARGE
                              : LoongArch::PseudoLA_TLS_DESC,
                        Large);
}
 
template <unsigned N>
static SDValue checkIntrinsicImmArg(SDValue Op, unsigned ImmOp,
                                    SelectionDAG &DAG, bool IsSigned = false) {
  auto *CImm = cast<ConstantSDNode>(Op->getOperand(ImmOp));
  // Check the ImmArg.
  if ((IsSigned && !isInt<N>(CImm->getSExtValue())) ||
      (!IsSigned && !isUInt<N>(CImm->getZExtValue()))) {
    DAG.getContext()->emitError(Op->getOperationName(0) +
                                ": argument out of range.");
    return DAG.getNode(ISD::UNDEF, SDLoc(Op), Op.getValueType());
  }
  return SDValue();
}
 
SDValue
LoongArchTargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
                                                 SelectionDAG &DAG) const {
  switch (Op.getConstantOperandVal(0)) {
  default:
    return SDValue(); // Don't custom lower most intrinsics.
  case Intrinsic::thread_pointer: {
    EVT PtrVT = getPointerTy(DAG.getDataLayout());
    return DAG.getRegister(LoongArch::R2, PtrVT);
  }
  case Intrinsic::loongarch_lsx_vpickve2gr_d:
  case Intrinsic::loongarch_lsx_vpickve2gr_du:
  case Intrinsic::loongarch_lsx_vreplvei_d:
  case Intrinsic::loongarch_lasx_xvrepl128vei_d:
    return checkIntrinsicImmArg<1>(Op, 2, DAG);
  case Intrinsic::loongarch_lsx_vreplvei_w:
  case Intrinsic::loongarch_lasx_xvrepl128vei_w:
  case Intrinsic::loongarch_lasx_xvpickve2gr_d:
  case Intrinsic::loongarch_lasx_xvpickve2gr_du:
  case Intrinsic::loongarch_lasx_xvpickve_d:
  case Intrinsic::loongarch_lasx_xvpickve_d_f:
    return checkIntrinsicImmArg<2>(Op, 2, DAG);
  case Intrinsic::loongarch_lasx_xvinsve0_d:
    return checkIntrinsicImmArg<2>(Op, 3, DAG);
  case Intrinsic::loongarch_lsx_vsat_b:
  case Intrinsic::loongarch_lsx_vsat_bu:
  case Intrinsic::loongarch_lsx_vrotri_b:
  case Intrinsic::loongarch_lsx_vsllwil_h_b:
  case Intrinsic::loongarch_lsx_vsllwil_hu_bu:
  case Intrinsic::loongarch_lsx_vsrlri_b:
  case Intrinsic::loongarch_lsx_vsrari_b:
  case Intrinsic::loongarch_lsx_vreplvei_h:
  case Intrinsic::loongarch_lasx_xvsat_b:
  case Intrinsic::loongarch_lasx_xvsat_bu:
  case Intrinsic::loongarch_lasx_xvrotri_b:
  case Intrinsic::loongarch_lasx_xvsllwil_h_b:
  case Intrinsic::loongarch_lasx_xvsllwil_hu_bu:
  case Intrinsic::loongarch_lasx_xvsrlri_b:
  case Intrinsic::loongarch_lasx_xvsrari_b:
  case Intrinsic::loongarch_lasx_xvrepl128vei_h:
  case Intrinsic::loongarch_lasx_xvpickve_w:
  case Intrinsic::loongarch_lasx_xvpickve_w_f:
    return checkIntrinsicImmArg<3>(Op, 2, DAG);
  case Intrinsic::loongarch_lasx_xvinsve0_w:
    return checkIntrinsicImmArg<3>(Op, 3, DAG);
  case Intrinsic::loongarch_lsx_vsat_h:
  case Intrinsic::loongarch_lsx_vsat_hu:
  case Intrinsic::loongarch_lsx_vrotri_h:
  case Intrinsic::loongarch_lsx_vsllwil_w_h:
  case Intrinsic::loongarch_lsx_vsllwil_wu_hu:
  case Intrinsic::loongarch_lsx_vsrlri_h:
  case Intrinsic::loongarch_lsx_vsrari_h:
  case Intrinsic::loongarch_lsx_vreplvei_b:
  case Intrinsic::loongarch_lasx_xvsat_h:
  case Intrinsic::loongarch_lasx_xvsat_hu:
  case Intrinsic::loongarch_lasx_xvrotri_h:
  case Intrinsic::loongarch_lasx_xvsllwil_w_h:
  case Intrinsic::loongarch_lasx_xvsllwil_wu_hu:
  case Intrinsic::loongarch_lasx_xvsrlri_h:
  case Intrinsic::loongarch_lasx_xvsrari_h:
  case Intrinsic::loongarch_lasx_xvrepl128vei_b:
    return checkIntrinsicImmArg<4>(Op, 2, DAG);
  case Intrinsic::loongarch_lsx_vsrlni_b_h:
  case Intrinsic::loongarch_lsx_vsrani_b_h:
  case Intrinsic::loongarch_lsx_vsrlrni_b_h:
  case Intrinsic::loongarch_lsx_vsrarni_b_h:
  case Intrinsic::loongarch_lsx_vssrlni_b_h:
  case Intrinsic::loongarch_lsx_vssrani_b_h:
  case Intrinsic::loongarch_lsx_vssrlni_bu_h:
  case Intrinsic::loongarch_lsx_vssrani_bu_h:
  case Intrinsic::loongarch_lsx_vssrlrni_b_h:
  case Intrinsic::loongarch_lsx_vssrarni_b_h:
  case Intrinsic::loongarch_lsx_vssrlrni_bu_h:
  case Intrinsic::loongarch_lsx_vssrarni_bu_h:
  case Intrinsic::loongarch_lasx_xvsrlni_b_h:
  case Intrinsic::loongarch_lasx_xvsrani_b_h:
  case Intrinsic::loongarch_lasx_xvsrlrni_b_h:
  case Intrinsic::loongarch_lasx_xvsrarni_b_h:
  case Intrinsic::loongarch_lasx_xvssrlni_b_h:
  case Intrinsic::loongarch_lasx_xvssrani_b_h:
  case Intrinsic::loongarch_lasx_xvssrlni_bu_h:
  case Intrinsic::loongarch_lasx_xvssrani_bu_h:
  case Intrinsic::loongarch_lasx_xvssrlrni_b_h:
  case Intrinsic::loongarch_lasx_xvssrarni_b_h:
  case Intrinsic::loongarch_lasx_xvssrlrni_bu_h:
  case Intrinsic::loongarch_lasx_xvssrarni_bu_h:
    return checkIntrinsicImmArg<4>(Op, 3, DAG);
  case Intrinsic::loongarch_lsx_vsat_w:
  case Intrinsic::loongarch_lsx_vsat_wu:
  case Intrinsic::loongarch_lsx_vrotri_w:
  case Intrinsic::loongarch_lsx_vsllwil_d_w:
  case Intrinsic::loongarch_lsx_vsllwil_du_wu:
  case Intrinsic::loongarch_lsx_vsrlri_w:
  case Intrinsic::loongarch_lsx_vsrari_w:
  case Intrinsic::loongarch_lsx_vslei_bu:
  case Intrinsic::loongarch_lsx_vslei_hu:
  case Intrinsic::loongarch_lsx_vslei_wu:
  case Intrinsic::loongarch_lsx_vslei_du:
  case Intrinsic::loongarch_lsx_vslti_bu:
  case Intrinsic::loongarch_lsx_vslti_hu:
  case Intrinsic::loongarch_lsx_vslti_wu:
  case Intrinsic::loongarch_lsx_vslti_du:
  case Intrinsic::loongarch_lsx_vbsll_v:
  case Intrinsic::loongarch_lsx_vbsrl_v:
  case Intrinsic::loongarch_lasx_xvsat_w:
  case Intrinsic::loongarch_lasx_xvsat_wu:
  case Intrinsic::loongarch_lasx_xvrotri_w:
  case Intrinsic::loongarch_lasx_xvsllwil_d_w:
  case Intrinsic::loongarch_lasx_xvsllwil_du_wu:
  case Intrinsic::loongarch_lasx_xvsrlri_w:
  case Intrinsic::loongarch_lasx_xvsrari_w:
  case Intrinsic::loongarch_lasx_xvslei_bu:
  case Intrinsic::loongarch_lasx_xvslei_hu:
  case Intrinsic::loongarch_lasx_xvslei_wu:
  case Intrinsic::loongarch_lasx_xvslei_du:
  case Intrinsic::loongarch_lasx_xvslti_bu:
  case Intrinsic::loongarch_lasx_xvslti_hu:
  case Intrinsic::loongarch_lasx_xvslti_wu:
  case Intrinsic::loongarch_lasx_xvslti_du:
  case Intrinsic::loongarch_lasx_xvbsll_v:
  case Intrinsic::loongarch_lasx_xvbsrl_v:
    return checkIntrinsicImmArg<5>(Op, 2, DAG);
  case Intrinsic::loongarch_lsx_vseqi_b:
  case Intrinsic::loongarch_lsx_vseqi_h:
  case Intrinsic::loongarch_lsx_vseqi_w:
  case Intrinsic::loongarch_lsx_vseqi_d:
  case Intrinsic::loongarch_lsx_vslei_b:
  case Intrinsic::loongarch_lsx_vslei_h:
  case Intrinsic::loongarch_lsx_vslei_w:
  case Intrinsic::loongarch_lsx_vslei_d:
  case Intrinsic::loongarch_lsx_vslti_b:
  case Intrinsic::loongarch_lsx_vslti_h:
  case Intrinsic::loongarch_lsx_vslti_w:
  case Intrinsic::loongarch_lsx_vslti_d:
  case Intrinsic::loongarch_lasx_xvseqi_b:
  case Intrinsic::loongarch_lasx_xvseqi_h:
  case Intrinsic::loongarch_lasx_xvseqi_w:
  case Intrinsic::loongarch_lasx_xvseqi_d:
  case Intrinsic::loongarch_lasx_xvslei_b:
  case Intrinsic::loongarch_lasx_xvslei_h:
  case Intrinsic::loongarch_lasx_xvslei_w:
  case Intrinsic::loongarch_lasx_xvslei_d:
  case Intrinsic::loongarch_lasx_xvslti_b:
  case Intrinsic::loongarch_lasx_xvslti_h:
  case Intrinsic::loongarch_lasx_xvslti_w:
  case Intrinsic::loongarch_lasx_xvslti_d:
    return checkIntrinsicImmArg<5>(Op, 2, DAG, /*IsSigned=*/true);
  case Intrinsic::loongarch_lsx_vsrlni_h_w:
  case Intrinsic::loongarch_lsx_vsrani_h_w:
  case Intrinsic::loongarch_lsx_vsrlrni_h_w:
  case Intrinsic::loongarch_lsx_vsrarni_h_w:
  case Intrinsic::loongarch_lsx_vssrlni_h_w:
  case Intrinsic::loongarch_lsx_vssrani_h_w:
  case Intrinsic::loongarch_lsx_vssrlni_hu_w:
  case Intrinsic::loongarch_lsx_vssrani_hu_w:
  case Intrinsic::loongarch_lsx_vssrlrni_h_w:
  case Intrinsic::loongarch_lsx_vssrarni_h_w:
  case Intrinsic::loongarch_lsx_vssrlrni_hu_w:
  case Intrinsic::loongarch_lsx_vssrarni_hu_w:
  case Intrinsic::loongarch_lsx_vfrstpi_b:
  case Intrinsic::loongarch_lsx_vfrstpi_h:
  case Intrinsic::loongarch_lasx_xvsrlni_h_w:
  case Intrinsic::loongarch_lasx_xvsrani_h_w:
  case Intrinsic::loongarch_lasx_xvsrlrni_h_w:
  case Intrinsic::loongarch_lasx_xvsrarni_h_w:
  case Intrinsic::loongarch_lasx_xvssrlni_h_w:
  case Intrinsic::loongarch_lasx_xvssrani_h_w:
  case Intrinsic::loongarch_lasx_xvssrlni_hu_w:
  case Intrinsic::loongarch_lasx_xvssrani_hu_w:
  case Intrinsic::loongarch_lasx_xvssrlrni_h_w:
  case Intrinsic::loongarch_lasx_xvssrarni_h_w:
  case Intrinsic::loongarch_lasx_xvssrlrni_hu_w:
  case Intrinsic::loongarch_lasx_xvssrarni_hu_w:
  case Intrinsic::loongarch_lasx_xvfrstpi_b:
  case Intrinsic::loongarch_lasx_xvfrstpi_h:
    return checkIntrinsicImmArg<5>(Op, 3, DAG);
  case Intrinsic::loongarch_lsx_vsat_d:
  case Intrinsic::loongarch_lsx_vsat_du:
  case Intrinsic::loongarch_lsx_vrotri_d:
  case Intrinsic::loongarch_lsx_vsrlri_d:
  case Intrinsic::loongarch_lsx_vsrari_d:
  case Intrinsic::loongarch_lasx_xvsat_d:
  case Intrinsic::loongarch_lasx_xvsat_du:
  case Intrinsic::loongarch_lasx_xvrotri_d:
  case Intrinsic::loongarch_lasx_xvsrlri_d:
  case Intrinsic::loongarch_lasx_xvsrari_d:
    return checkIntrinsicImmArg<6>(Op, 2, DAG);
  case Intrinsic::loongarch_lsx_vsrlni_w_d:
  case Intrinsic::loongarch_lsx_vsrani_w_d:
  case Intrinsic::loongarch_lsx_vsrlrni_w_d:
  case Intrinsic::loongarch_lsx_vsrarni_w_d:
  case Intrinsic::loongarch_lsx_vssrlni_w_d:
  case Intrinsic::loongarch_lsx_vssrani_w_d:
  case Intrinsic::loongarch_lsx_vssrlni_wu_d:
  case Intrinsic::loongarch_lsx_vssrani_wu_d:
  case Intrinsic::loongarch_lsx_vssrlrni_w_d:
  case Intrinsic::loongarch_lsx_vssrarni_w_d:
  case Intrinsic::loongarch_lsx_vssrlrni_wu_d:
  case Intrinsic::loongarch_lsx_vssrarni_wu_d:
  case Intrinsic::loongarch_lasx_xvsrlni_w_d:
  case Intrinsic::loongarch_lasx_xvsrani_w_d:
  case Intrinsic::loongarch_lasx_xvsrlrni_w_d:
  case Intrinsic::loongarch_lasx_xvsrarni_w_d:
  case Intrinsic::loongarch_lasx_xvssrlni_w_d:
  case Intrinsic::loongarch_lasx_xvssrani_w_d:
  case Intrinsic::loongarch_lasx_xvssrlni_wu_d:
  case Intrinsic::loongarch_lasx_xvssrani_wu_d:
  case Intrinsic::loongarch_lasx_xvssrlrni_w_d:
  case Intrinsic::loongarch_lasx_xvssrarni_w_d:
  case Intrinsic::loongarch_lasx_xvssrlrni_wu_d:
  case Intrinsic::loongarch_lasx_xvssrarni_wu_d:
    return checkIntrinsicImmArg<6>(Op, 3, DAG);
  case Intrinsic::loongarch_lsx_vsrlni_d_q:
  case Intrinsic::loongarch_lsx_vsrani_d_q:
  case Intrinsic::loongarch_lsx_vsrlrni_d_q:
  case Intrinsic::loongarch_lsx_vsrarni_d_q:
  case Intrinsic::loongarch_lsx_vssrlni_d_q:
  case Intrinsic::loongarch_lsx_vssrani_d_q:
  case Intrinsic::loongarch_lsx_vssrlni_du_q:
  case Intrinsic::loongarch_lsx_vssrani_du_q:
  case Intrinsic::loongarch_lsx_vssrlrni_d_q:
  case Intrinsic::loongarch_lsx_vssrarni_d_q:
  case Intrinsic::loongarch_lsx_vssrlrni_du_q:
  case Intrinsic::loongarch_lsx_vssrarni_du_q:
  case Intrinsic::loongarch_lasx_xvsrlni_d_q:
  case Intrinsic::loongarch_lasx_xvsrani_d_q:
  case Intrinsic::loongarch_lasx_xvsrlrni_d_q:
  case Intrinsic::loongarch_lasx_xvsrarni_d_q:
  case Intrinsic::loongarch_lasx_xvssrlni_d_q:
  case Intrinsic::loongarch_lasx_xvssrani_d_q:
  case Intrinsic::loongarch_lasx_xvssrlni_du_q:
  case Intrinsic::loongarch_lasx_xvssrani_du_q:
  case Intrinsic::loongarch_lasx_xvssrlrni_d_q:
  case Intrinsic::loongarch_lasx_xvssrarni_d_q:
  case Intrinsic::loongarch_lasx_xvssrlrni_du_q:
  case Intrinsic::loongarch_lasx_xvssrarni_du_q:
    return checkIntrinsicImmArg<7>(Op, 3, DAG);
  case Intrinsic::loongarch_lsx_vnori_b:
  case Intrinsic::loongarch_lsx_vshuf4i_b:
  case Intrinsic::loongarch_lsx_vshuf4i_h:
  case Intrinsic::loongarch_lsx_vshuf4i_w:
  case Intrinsic::loongarch_lasx_xvnori_b:
  case Intrinsic::loongarch_lasx_xvshuf4i_b:
  case Intrinsic::loongarch_lasx_xvshuf4i_h:
  case Intrinsic::loongarch_lasx_xvshuf4i_w:
  case Intrinsic::loongarch_lasx_xvpermi_d:
    return checkIntrinsicImmArg<8>(Op, 2, DAG);
  case Intrinsic::loongarch_lsx_vshuf4i_d:
  case Intrinsic::loongarch_lsx_vpermi_w:
  case Intrinsic::loongarch_lsx_vbitseli_b:
  case Intrinsic::loongarch_lsx_vextrins_b:
  case Intrinsic::loongarch_lsx_vextrins_h:
  case Intrinsic::loongarch_lsx_vextrins_w:
  case Intrinsic::loongarch_lsx_vextrins_d:
  case Intrinsic::loongarch_lasx_xvshuf4i_d:
  case Intrinsic::loongarch_lasx_xvpermi_w:
  case Intrinsic::loongarch_lasx_xvpermi_q:
  case Intrinsic::loongarch_lasx_xvbitseli_b:
  case Intrinsic::loongarch_lasx_xvextrins_b:
  case Intrinsic::loongarch_lasx_xvextrins_h:
  case Intrinsic::loongarch_lasx_xvextrins_w:
  case Intrinsic::loongarch_lasx_xvextrins_d:
    return checkIntrinsicImmArg<8>(Op, 3, DAG);
  case Intrinsic::loongarch_lsx_vrepli_b:
  case Intrinsic::loongarch_lsx_vrepli_h:
  case Intrinsic::loongarch_lsx_vrepli_w:
  case Intrinsic::loongarch_lsx_vrepli_d:
  case Intrinsic::loongarch_lasx_xvrepli_b:
  case Intrinsic::loongarch_lasx_xvrepli_h:
  case Intrinsic::loongarch_lasx_xvrepli_w:
  case Intrinsic::loongarch_lasx_xvrepli_d:
    return checkIntrinsicImmArg<10>(Op, 1, DAG, /*IsSigned=*/true);
  case Intrinsic::loongarch_lsx_vldi:
  case Intrinsic::loongarch_lasx_xvldi:
    return checkIntrinsicImmArg<13>(Op, 1, DAG, /*IsSigned=*/true);
  }
}
 
// Helper function that emits error message for intrinsics with chain and return
// merge values of a UNDEF and the chain.
static SDValue emitIntrinsicWithChainErrorMessage(SDValue Op,
                                                  StringRef ErrorMsg,
                                                  SelectionDAG &DAG) {
  DAG.getContext()->emitError(Op->getOperationName(0) + ": " + ErrorMsg + ".");
  return DAG.getMergeValues({DAG.getUNDEF(Op.getValueType()), Op.getOperand(0)},
                            SDLoc(Op));
}
 
SDValue
LoongArchTargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op,
                                                SelectionDAG &DAG) const {
  SDLoc DL(Op);
  MVT GRLenVT = Subtarget.getGRLenVT();
  EVT VT = Op.getValueType();
  SDValue Chain = Op.getOperand(0);
  const StringRef ErrorMsgOOR = "argument out of range";
  const StringRef ErrorMsgReqLA64 = "requires loongarch64";
  const StringRef ErrorMsgReqF = "requires basic 'f' target feature";
 
  switch (Op.getConstantOperandVal(1)) {
  default:
    return Op;
  case Intrinsic::loongarch_crc_w_b_w:
  case Intrinsic::loongarch_crc_w_h_w:
  case Intrinsic::loongarch_crc_w_w_w:
  case Intrinsic::loongarch_crc_w_d_w:
  case Intrinsic::loongarch_crcc_w_b_w:
  case Intrinsic::loongarch_crcc_w_h_w:
  case Intrinsic::loongarch_crcc_w_w_w:
  case Intrinsic::loongarch_crcc_w_d_w:
    return emitIntrinsicWithChainErrorMessage(Op, ErrorMsgReqLA64, DAG);
  case Intrinsic::loongarch_csrrd_w:
  case Intrinsic::loongarch_csrrd_d: {
    unsigned Imm = Op.getConstantOperandVal(2);
    return !isUInt<14>(Imm)
               ? emitIntrinsicWithChainErrorMessage(Op, ErrorMsgOOR, DAG)
               : DAG.getNode(LoongArchISD::CSRRD, DL, {GRLenVT, MVT::Other},
                             {Chain, DAG.getConstant(Imm, DL, GRLenVT)});
  }
  case Intrinsic::loongarch_csrwr_w:
  case Intrinsic::loongarch_csrwr_d: {
    unsigned Imm = Op.getConstantOperandVal(3);
    return !isUInt<14>(Imm)
               ? emitIntrinsicWithChainErrorMessage(Op, ErrorMsgOOR, DAG)
               : DAG.getNode(LoongArchISD::CSRWR, DL, {GRLenVT, MVT::Other},
                             {Chain, Op.getOperand(2),
                              DAG.getConstant(Imm, DL, GRLenVT)});
  }
  case Intrinsic::loongarch_csrxchg_w:
  case Intrinsic::loongarch_csrxchg_d: {
    unsigned Imm = Op.getConstantOperandVal(4);
    return !isUInt<14>(Imm)
               ? emitIntrinsicWithChainErrorMessage(Op, ErrorMsgOOR, DAG)
               : DAG.getNode(LoongArchISD::CSRXCHG, DL, {GRLenVT, MVT::Other},
                             {Chain, Op.getOperand(2), Op.getOperand(3),
                              DAG.getConstant(Imm, DL, GRLenVT)});
  }
  case Intrinsic::loongarch_iocsrrd_d: {
    return DAG.getNode(
        LoongArchISD::IOCSRRD_D, DL, {GRLenVT, MVT::Other},
        {Chain, DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op.getOperand(2))});
  }
#define IOCSRRD_CASE(NAME, NODE)                                               \
  case Intrinsic::loongarch_##NAME: {                                          \
    return DAG.getNode(LoongArchISD::NODE, DL, {GRLenVT, MVT::Other},          \
                       {Chain, Op.getOperand(2)});                             \
  }
    IOCSRRD_CASE(iocsrrd_b, IOCSRRD_B);
    IOCSRRD_CASE(iocsrrd_h, IOCSRRD_H);
    IOCSRRD_CASE(iocsrrd_w, IOCSRRD_W);
#undef IOCSRRD_CASE
  case Intrinsic::loongarch_cpucfg: {
    return DAG.getNode(LoongArchISD::CPUCFG, DL, {GRLenVT, MVT::Other},
                       {Chain, Op.getOperand(2)});
  }
  case Intrinsic::loongarch_lddir_d: {
    unsigned Imm = Op.getConstantOperandVal(3);
    return !isUInt<8>(Imm)
               ? emitIntrinsicWithChainErrorMessage(Op, ErrorMsgOOR, DAG)
               : Op;
  }
  case Intrinsic::loongarch_movfcsr2gr: {
    if (!Subtarget.hasBasicF())
      return emitIntrinsicWithChainErrorMessage(Op, ErrorMsgReqF, DAG);
    unsigned Imm = Op.getConstantOperandVal(2);
    return !isUInt<2>(Imm)
               ? emitIntrinsicWithChainErrorMessage(Op, ErrorMsgOOR, DAG)
               : DAG.getNode(LoongArchISD::MOVFCSR2GR, DL, {VT, MVT::Other},
                             {Chain, DAG.getConstant(Imm, DL, GRLenVT)});
  }
  case Intrinsic::loongarch_lsx_vld:
  case Intrinsic::loongarch_lsx_vldrepl_b:
  case Intrinsic::loongarch_lasx_xvld:
  case Intrinsic::loongarch_lasx_xvldrepl_b:
    return !isInt<12>(cast<ConstantSDNode>(Op.getOperand(3))->getSExtValue())
               ? emitIntrinsicWithChainErrorMessage(Op, ErrorMsgOOR, DAG)
               : SDValue();
  case Intrinsic::loongarch_lsx_vldrepl_h:
  case Intrinsic::loongarch_lasx_xvldrepl_h:
    return !isShiftedInt<11, 1>(
               cast<ConstantSDNode>(Op.getOperand(3))->getSExtValue())
               ? emitIntrinsicWithChainErrorMessage(
                     Op, "argument out of range or not a multiple of 2", DAG)
               : SDValue();
  case Intrinsic::loongarch_lsx_vldrepl_w:
  case Intrinsic::loongarch_lasx_xvldrepl_w:
    return !isShiftedInt<10, 2>(
               cast<ConstantSDNode>(Op.getOperand(3))->getSExtValue())
               ? emitIntrinsicWithChainErrorMessage(
                     Op, "argument out of range or not a multiple of 4", DAG)
               : SDValue();
  case Intrinsic::loongarch_lsx_vldrepl_d:
  case Intrinsic::loongarch_lasx_xvldrepl_d:
    return !isShiftedInt<9, 3>(
               cast<ConstantSDNode>(Op.getOperand(3))->getSExtValue())
               ? emitIntrinsicWithChainErrorMessage(
                     Op, "argument out of range or not a multiple of 8", DAG)
               : SDValue();
  }
}
 
// Helper function that emits error message for intrinsics with void return
// value and return the chain.
static SDValue emitIntrinsicErrorMessage(SDValue Op, StringRef ErrorMsg,
                                         SelectionDAG &DAG) {
 
  DAG.getContext()->emitError(Op->getOperationName(0) + ": " + ErrorMsg + ".");
  return Op.getOperand(0);
}
 
SDValue LoongArchTargetLowering::lowerINTRINSIC_VOID(SDValue Op,
                                                     SelectionDAG &DAG) const {
  SDLoc DL(Op);
  MVT GRLenVT = Subtarget.getGRLenVT();
  SDValue Chain = Op.getOperand(0);
  uint64_t IntrinsicEnum = Op.getConstantOperandVal(1);
  SDValue Op2 = Op.getOperand(2);
  const StringRef ErrorMsgOOR = "argument out of range";
  const StringRef ErrorMsgReqLA64 = "requires loongarch64";
  const StringRef ErrorMsgReqLA32 = "requires loongarch32";
  const StringRef ErrorMsgReqF = "requires basic 'f' target feature";
 
  switch (IntrinsicEnum) {
  default:
    // TODO: Add more Intrinsics.
    return SDValue();
  case Intrinsic::loongarch_cacop_d:
  case Intrinsic::loongarch_cacop_w: {
    if (IntrinsicEnum == Intrinsic::loongarch_cacop_d && !Subtarget.is64Bit())
      return emitIntrinsicErrorMessage(Op, ErrorMsgReqLA64, DAG);
    if (IntrinsicEnum == Intrinsic::loongarch_cacop_w && Subtarget.is64Bit())
      return emitIntrinsicErrorMessage(Op, ErrorMsgReqLA32, DAG);
    // call void @llvm.loongarch.cacop.[d/w](uimm5, rj, simm12)
    unsigned Imm1 = Op2->getAsZExtVal();
    int Imm2 = cast<ConstantSDNode>(Op.getOperand(4))->getSExtValue();
    if (!isUInt<5>(Imm1) || !isInt<12>(Imm2))
      return emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG);
    return Op;
  }
  case Intrinsic::loongarch_dbar: {
    unsigned Imm = Op2->getAsZExtVal();
    return !isUInt<15>(Imm)
               ? emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG)
               : DAG.getNode(LoongArchISD::DBAR, DL, MVT::Other, Chain,
                             DAG.getConstant(Imm, DL, GRLenVT));
  }
  case Intrinsic::loongarch_ibar: {
    unsigned Imm = Op2->getAsZExtVal();
    return !isUInt<15>(Imm)
               ? emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG)
               : DAG.getNode(LoongArchISD::IBAR, DL, MVT::Other, Chain,
                             DAG.getConstant(Imm, DL, GRLenVT));
  }
  case Intrinsic::loongarch_break: {
    unsigned Imm = Op2->getAsZExtVal();
    return !isUInt<15>(Imm)
               ? emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG)
               : DAG.getNode(LoongArchISD::BREAK, DL, MVT::Other, Chain,
                             DAG.getConstant(Imm, DL, GRLenVT));
  }
  case Intrinsic::loongarch_movgr2fcsr: {
    if (!Subtarget.hasBasicF())
      return emitIntrinsicErrorMessage(Op, ErrorMsgReqF, DAG);
    unsigned Imm = Op2->getAsZExtVal();
    return !isUInt<2>(Imm)
               ? emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG)
               : DAG.getNode(LoongArchISD::MOVGR2FCSR, DL, MVT::Other, Chain,
                             DAG.getConstant(Imm, DL, GRLenVT),
                             DAG.getNode(ISD::ANY_EXTEND, DL, GRLenVT,
                                         Op.getOperand(3)));
  }
  case Intrinsic::loongarch_syscall: {
    unsigned Imm = Op2->getAsZExtVal();
    return !isUInt<15>(Imm)
               ? emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG)
               : DAG.getNode(LoongArchISD::SYSCALL, DL, MVT::Other, Chain,
                             DAG.getConstant(Imm, DL, GRLenVT));
  }
#define IOCSRWR_CASE(NAME, NODE)                                               \
  case Intrinsic::loongarch_##NAME: {                                          \
    SDValue Op3 = Op.getOperand(3);                                            \
    return Subtarget.is64Bit()                                                 \
               ? DAG.getNode(LoongArchISD::NODE, DL, MVT::Other, Chain,        \
                             DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op2),  \
                             DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op3))  \
               : DAG.getNode(LoongArchISD::NODE, DL, MVT::Other, Chain, Op2,   \
                             Op3);                                             \
  }
    IOCSRWR_CASE(iocsrwr_b, IOCSRWR_B);
    IOCSRWR_CASE(iocsrwr_h, IOCSRWR_H);
    IOCSRWR_CASE(iocsrwr_w, IOCSRWR_W);
#undef IOCSRWR_CASE
  case Intrinsic::loongarch_iocsrwr_d: {
    return !Subtarget.is64Bit()
               ? emitIntrinsicErrorMessage(Op, ErrorMsgReqLA64, DAG)
               : DAG.getNode(LoongArchISD::IOCSRWR_D, DL, MVT::Other, Chain,
                             Op2,
                             DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64,
                                         Op.getOperand(3)));
  }
#define ASRT_LE_GT_CASE(NAME)                                                  \
  case Intrinsic::loongarch_##NAME: {                                          \
    return !Subtarget.is64Bit()                                                \
               ? emitIntrinsicErrorMessage(Op, ErrorMsgReqLA64, DAG)           \
               : Op;                                                           \
  }
    ASRT_LE_GT_CASE(asrtle_d)
    ASRT_LE_GT_CASE(asrtgt_d)
#undef ASRT_LE_GT_CASE
  case Intrinsic::loongarch_ldpte_d: {
    unsigned Imm = Op.getConstantOperandVal(3);
    return !Subtarget.is64Bit()
               ? emitIntrinsicErrorMessage(Op, ErrorMsgReqLA64, DAG)
           : !isUInt<8>(Imm) ? emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG)
                             : Op;
  }
  case Intrinsic::loongarch_lsx_vst:
  case Intrinsic::loongarch_lasx_xvst:
    return !isInt<12>(cast<ConstantSDNode>(Op.getOperand(4))->getSExtValue())
               ? emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG)
               : SDValue();
  case Intrinsic::loongarch_lasx_xvstelm_b:
    return (!isInt<8>(cast<ConstantSDNode>(Op.getOperand(4))->getSExtValue()) ||
            !isUInt<5>(Op.getConstantOperandVal(5)))
               ? emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG)
               : SDValue();
  case Intrinsic::loongarch_lsx_vstelm_b:
    return (!isInt<8>(cast<ConstantSDNode>(Op.getOperand(4))->getSExtValue()) ||
            !isUInt<4>(Op.getConstantOperandVal(5)))
               ? emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG)
               : SDValue();
  case Intrinsic::loongarch_lasx_xvstelm_h:
    return (!isShiftedInt<8, 1>(
                cast<ConstantSDNode>(Op.getOperand(4))->getSExtValue()) ||
            !isUInt<4>(Op.getConstantOperandVal(5)))
               ? emitIntrinsicErrorMessage(
                     Op, "argument out of range or not a multiple of 2", DAG)
               : SDValue();
  case Intrinsic::loongarch_lsx_vstelm_h:
    return (!isShiftedInt<8, 1>(
                cast<ConstantSDNode>(Op.getOperand(4))->getSExtValue()) ||
            !isUInt<3>(Op.getConstantOperandVal(5)))
               ? emitIntrinsicErrorMessage(
                     Op, "argument out of range or not a multiple of 2", DAG)
               : SDValue();
  case Intrinsic::loongarch_lasx_xvstelm_w:
    return (!isShiftedInt<8, 2>(
                cast<ConstantSDNode>(Op.getOperand(4))->getSExtValue()) ||
            !isUInt<3>(Op.getConstantOperandVal(5)))
               ? emitIntrinsicErrorMessage(
                     Op, "argument out of range or not a multiple of 4", DAG)
               : SDValue();
  case Intrinsic::loongarch_lsx_vstelm_w:
    return (!isShiftedInt<8, 2>(
                cast<ConstantSDNode>(Op.getOperand(4))->getSExtValue()) ||
            !isUInt<2>(Op.getConstantOperandVal(5)))
               ? emitIntrinsicErrorMessage(
                     Op, "argument out of range or not a multiple of 4", DAG)
               : SDValue();
  case Intrinsic::loongarch_lasx_xvstelm_d:
    return (!isShiftedInt<8, 3>(
                cast<ConstantSDNode>(Op.getOperand(4))->getSExtValue()) ||
            !isUInt<2>(Op.getConstantOperandVal(5)))
               ? emitIntrinsicErrorMessage(
                     Op, "argument out of range or not a multiple of 8", DAG)
               : SDValue();
  case Intrinsic::loongarch_lsx_vstelm_d:
    return (!isShiftedInt<8, 3>(
                cast<ConstantSDNode>(Op.getOperand(4))->getSExtValue()) ||
            !isUInt<1>(Op.getConstantOperandVal(5)))
               ? emitIntrinsicErrorMessage(
                     Op, "argument out of range or not a multiple of 8", DAG)
               : SDValue();
  }
}
 
SDValue LoongArchTargetLowering::lowerShiftLeftParts(SDValue Op,
                                                     SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SDValue Lo = Op.getOperand(0);
  SDValue Hi = Op.getOperand(1);
  SDValue Shamt = Op.getOperand(2);
  EVT VT = Lo.getValueType();
 
  // if Shamt-GRLen < 0: // Shamt < GRLen
  //   Lo = Lo << Shamt
  //   Hi = (Hi << Shamt) | ((Lo >>u 1) >>u (GRLen-1 ^ Shamt))
  // else:
  //   Lo = 0
  //   Hi = Lo << (Shamt-GRLen)
 
  SDValue Zero = DAG.getConstant(0, DL, VT);
  SDValue One = DAG.getConstant(1, DL, VT);
  SDValue MinusGRLen =
      DAG.getSignedConstant(-(int)Subtarget.getGRLen(), DL, VT);
  SDValue GRLenMinus1 = DAG.getConstant(Subtarget.getGRLen() - 1, DL, VT);
  SDValue ShamtMinusGRLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusGRLen);
  SDValue GRLenMinus1Shamt = DAG.getNode(ISD::XOR, DL, VT, Shamt, GRLenMinus1);
 
  SDValue LoTrue = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt);
  SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo, One);
  SDValue ShiftRightLo =
      DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, GRLenMinus1Shamt);
  SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt);
  SDValue HiTrue = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
  SDValue HiFalse = DAG.getNode(ISD::SHL, DL, VT, Lo, ShamtMinusGRLen);
 
  SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusGRLen, Zero, ISD::SETLT);
 
  Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, Zero);
  Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse);
 
  SDValue Parts[2] = {Lo, Hi};
  return DAG.getMergeValues(Parts, DL);
}
 
SDValue LoongArchTargetLowering::lowerShiftRightParts(SDValue Op,
                                                      SelectionDAG &DAG,
                                                      bool IsSRA) const {
  SDLoc DL(Op);
  SDValue Lo = Op.getOperand(0);
  SDValue Hi = Op.getOperand(1);
  SDValue Shamt = Op.getOperand(2);
  EVT VT = Lo.getValueType();
 
  // SRA expansion:
  //   if Shamt-GRLen < 0: // Shamt < GRLen
  //     Lo = (Lo >>u Shamt) | ((Hi << 1) << (ShAmt ^ GRLen-1))
  //     Hi = Hi >>s Shamt
  //   else:
  //     Lo = Hi >>s (Shamt-GRLen);
  //     Hi = Hi >>s (GRLen-1)
  //
  // SRL expansion:
  //   if Shamt-GRLen < 0: // Shamt < GRLen
  //     Lo = (Lo >>u Shamt) | ((Hi << 1) << (ShAmt ^ GRLen-1))
  //     Hi = Hi >>u Shamt
  //   else:
  //     Lo = Hi >>u (Shamt-GRLen);
  //     Hi = 0;
 
  unsigned ShiftRightOp = IsSRA ? ISD::SRA : ISD::SRL;
 
  SDValue Zero = DAG.getConstant(0, DL, VT);
  SDValue One = DAG.getConstant(1, DL, VT);
  SDValue MinusGRLen =
      DAG.getSignedConstant(-(int)Subtarget.getGRLen(), DL, VT);
  SDValue GRLenMinus1 = DAG.getConstant(Subtarget.getGRLen() - 1, DL, VT);
  SDValue ShamtMinusGRLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusGRLen);
  SDValue GRLenMinus1Shamt = DAG.getNode(ISD::XOR, DL, VT, Shamt, GRLenMinus1);
 
  SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt);
  SDValue ShiftLeftHi1 = DAG.getNode(ISD::SHL, DL, VT, Hi, One);
  SDValue ShiftLeftHi =
      DAG.getNode(ISD::SHL, DL, VT, ShiftLeftHi1, GRLenMinus1Shamt);
  SDValue LoTrue = DAG.getNode(ISD::OR, DL, VT, ShiftRightLo, ShiftLeftHi);
  SDValue HiTrue = DAG.getNode(ShiftRightOp, DL, VT, Hi, Shamt);
  SDValue LoFalse = DAG.getNode(ShiftRightOp, DL, VT, Hi, ShamtMinusGRLen);
  SDValue HiFalse =
      IsSRA ? DAG.getNode(ISD::SRA, DL, VT, Hi, GRLenMinus1) : Zero;
 
  SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusGRLen, Zero, ISD::SETLT);
 
  Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, LoFalse);
  Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse);
 
  SDValue Parts[2] = {Lo, Hi};
  return DAG.getMergeValues(Parts, DL);
}
 
// Returns the opcode of the target-specific SDNode that implements the 32-bit
// form of the given Opcode.
static unsigned getLoongArchWOpcode(unsigned Opcode) {
  switch (Opcode) {
  default:
    llvm_unreachable("Unexpected opcode");
  case ISD::SDIV:
    return LoongArchISD::DIV_W;
  case ISD::UDIV:
    return LoongArchISD::DIV_WU;
  case ISD::SREM:
    return LoongArchISD::MOD_W;
  case ISD::UREM:
    return LoongArchISD::MOD_WU;
  case ISD::SHL:
    return LoongArchISD::SLL_W;
  case ISD::SRA:
    return LoongArchISD::SRA_W;
  case ISD::SRL:
    return LoongArchISD::SRL_W;
  case ISD::ROTL:
  case ISD::ROTR:
    return LoongArchISD::ROTR_W;
  case ISD::CTTZ:
    return LoongArchISD::CTZ_W;
  case ISD::CTLZ:
    return LoongArchISD::CLZ_W;
  }
}
 
// Converts the given i8/i16/i32 operation to a target-specific SelectionDAG
// node. Because i8/i16/i32 isn't a legal type for LA64, these operations would
// otherwise be promoted to i64, making it difficult to select the
// SLL_W/.../*W later one because the fact the operation was originally of
// type i8/i16/i32 is lost.
static SDValue customLegalizeToWOp(SDNode *N, SelectionDAG &DAG, int NumOp,
                                   unsigned ExtOpc = ISD::ANY_EXTEND) {
  SDLoc DL(N);
  unsigned WOpcode = getLoongArchWOpcode(N->getOpcode());
  SDValue NewOp0, NewRes;
 
  switch (NumOp) {
  default:
    llvm_unreachable("Unexpected NumOp");
  case 1: {
    NewOp0 = DAG.getNode(ExtOpc, DL, MVT::i64, N->getOperand(0));
    NewRes = DAG.getNode(WOpcode, DL, MVT::i64, NewOp0);
    break;
  }
  case 2: {
    NewOp0 = DAG.getNode(ExtOpc, DL, MVT::i64, N->getOperand(0));
    SDValue NewOp1 = DAG.getNode(ExtOpc, DL, MVT::i64, N->getOperand(1));
    if (N->getOpcode() == ISD::ROTL) {
      SDValue TmpOp = DAG.getConstant(32, DL, MVT::i64);
      NewOp1 = DAG.getNode(ISD::SUB, DL, MVT::i64, TmpOp, NewOp1);
    }
    NewRes = DAG.getNode(WOpcode, DL, MVT::i64, NewOp0, NewOp1);
    break;
  }
    // TODO:Handle more NumOp.
  }
 
  // ReplaceNodeResults requires we maintain the same type for the return
  // value.
  return DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), NewRes);
}
 
// Converts the given 32-bit operation to a i64 operation with signed extension
// semantic to reduce the signed extension instructions.
static SDValue customLegalizeToWOpWithSExt(SDNode *N, SelectionDAG &DAG) {
  SDLoc DL(N);
  SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0));
  SDValue NewOp1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
  SDValue NewWOp = DAG.getNode(N->getOpcode(), DL, MVT::i64, NewOp0, NewOp1);
  SDValue NewRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, NewWOp,
                               DAG.getValueType(MVT::i32));
  return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, NewRes);
}
 
// Helper function that emits error message for intrinsics with/without chain
// and return a UNDEF or and the chain as the results.
static void emitErrorAndReplaceIntrinsicResults(
    SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG,
    StringRef ErrorMsg, bool WithChain = true) {
  DAG.getContext()->emitError(N->getOperationName(0) + ": " + ErrorMsg + ".");
  Results.push_back(DAG.getUNDEF(N->getValueType(0)));
  if (!WithChain)
    return;
  Results.push_back(N->getOperand(0));
}
 
template <unsigned N>
static void
replaceVPICKVE2GRResults(SDNode *Node, SmallVectorImpl<SDValue> &Results,
                         SelectionDAG &DAG, const LoongArchSubtarget &Subtarget,
                         unsigned ResOp) {
  const StringRef ErrorMsgOOR = "argument out of range";
  unsigned Imm = Node->getConstantOperandVal(2);
  if (!isUInt<N>(Imm)) {
    emitErrorAndReplaceIntrinsicResults(Node, Results, DAG, ErrorMsgOOR,
                                        /*WithChain=*/false);
    return;
  }
  SDLoc DL(Node);
  SDValue Vec = Node->getOperand(1);
 
  SDValue PickElt =
      DAG.getNode(ResOp, DL, Subtarget.getGRLenVT(), Vec,
                  DAG.getConstant(Imm, DL, Subtarget.getGRLenVT()),
                  DAG.getValueType(Vec.getValueType().getVectorElementType()));
  Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, Node->getValueType(0),
                                PickElt.getValue(0)));
}
 
static void replaceVecCondBranchResults(SDNode *N,
                                        SmallVectorImpl<SDValue> &Results,
                                        SelectionDAG &DAG,
                                        const LoongArchSubtarget &Subtarget,
                                        unsigned ResOp) {
  SDLoc DL(N);
  SDValue Vec = N->getOperand(1);
 
  SDValue CB = DAG.getNode(ResOp, DL, Subtarget.getGRLenVT(), Vec);
  Results.push_back(
      DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), CB.getValue(0)));
}
 
static void
replaceINTRINSIC_WO_CHAINResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
                                 SelectionDAG &DAG,
                                 const LoongArchSubtarget &Subtarget) {
  switch (N->getConstantOperandVal(0)) {
  default:
    llvm_unreachable("Unexpected Intrinsic.");
  case Intrinsic::loongarch_lsx_vpickve2gr_b:
    replaceVPICKVE2GRResults<4>(N, Results, DAG, Subtarget,
                                LoongArchISD::VPICK_SEXT_ELT);
    break;
  case Intrinsic::loongarch_lsx_vpickve2gr_h:
  case Intrinsic::loongarch_lasx_xvpickve2gr_w:
    replaceVPICKVE2GRResults<3>(N, Results, DAG, Subtarget,
                                LoongArchISD::VPICK_SEXT_ELT);
    break;
  case Intrinsic::loongarch_lsx_vpickve2gr_w:
    replaceVPICKVE2GRResults<2>(N, Results, DAG, Subtarget,
                                LoongArchISD::VPICK_SEXT_ELT);
    break;
  case Intrinsic::loongarch_lsx_vpickve2gr_bu:
    replaceVPICKVE2GRResults<4>(N, Results, DAG, Subtarget,
                                LoongArchISD::VPICK_ZEXT_ELT);
    break;
  case Intrinsic::loongarch_lsx_vpickve2gr_hu:
  case Intrinsic::loongarch_lasx_xvpickve2gr_wu:
    replaceVPICKVE2GRResults<3>(N, Results, DAG, Subtarget,
                                LoongArchISD::VPICK_ZEXT_ELT);
    break;
  case Intrinsic::loongarch_lsx_vpickve2gr_wu:
    replaceVPICKVE2GRResults<2>(N, Results, DAG, Subtarget,
                                LoongArchISD::VPICK_ZEXT_ELT);
    break;
  case Intrinsic::loongarch_lsx_bz_b:
  case Intrinsic::loongarch_lsx_bz_h:
  case Intrinsic::loongarch_lsx_bz_w:
  case Intrinsic::loongarch_lsx_bz_d:
  case Intrinsic::loongarch_lasx_xbz_b:
  case Intrinsic::loongarch_lasx_xbz_h:
  case Intrinsic::loongarch_lasx_xbz_w:
  case Intrinsic::loongarch_lasx_xbz_d:
    replaceVecCondBranchResults(N, Results, DAG, Subtarget,
                                LoongArchISD::VALL_ZERO);
    break;
  case Intrinsic::loongarch_lsx_bz_v:
  case Intrinsic::loongarch_lasx_xbz_v:
    replaceVecCondBranchResults(N, Results, DAG, Subtarget,
                                LoongArchISD::VANY_ZERO);
    break;
  case Intrinsic::loongarch_lsx_bnz_b:
  case Intrinsic::loongarch_lsx_bnz_h:
  case Intrinsic::loongarch_lsx_bnz_w:
  case Intrinsic::loongarch_lsx_bnz_d:
  case Intrinsic::loongarch_lasx_xbnz_b:
  case Intrinsic::loongarch_lasx_xbnz_h:
  case Intrinsic::loongarch_lasx_xbnz_w:
  case Intrinsic::loongarch_lasx_xbnz_d:
    replaceVecCondBranchResults(N, Results, DAG, Subtarget,
                                LoongArchISD::VALL_NONZERO);
    break;
  case Intrinsic::loongarch_lsx_bnz_v:
  case Intrinsic::loongarch_lasx_xbnz_v:
    replaceVecCondBranchResults(N, Results, DAG, Subtarget,
                                LoongArchISD::VANY_NONZERO);
    break;
  }
}
 
static void replaceCMP_XCHG_128Results(SDNode *N,
                                       SmallVectorImpl<SDValue> &Results,
                                       SelectionDAG &DAG) {
  assert(N->getValueType(0) == MVT::i128 &&
         "AtomicCmpSwap on types less than 128 should be legal");
  MachineMemOperand *MemOp = cast<MemSDNode>(N)->getMemOperand();
 
  unsigned Opcode;
  switch (MemOp->getMergedOrdering()) {
  case AtomicOrdering::Acquire:
  case AtomicOrdering::AcquireRelease:
  case AtomicOrdering::SequentiallyConsistent:
    Opcode = LoongArch::PseudoCmpXchg128Acquire;
    break;
  case AtomicOrdering::Monotonic:
  case AtomicOrdering::Release:
    Opcode = LoongArch::PseudoCmpXchg128;
    break;
  default:
    llvm_unreachable("Unexpected ordering!");
  }
 
  SDLoc DL(N);
  auto CmpVal = DAG.SplitScalar(N->getOperand(2), DL, MVT::i64, MVT::i64);
  auto NewVal = DAG.SplitScalar(N->getOperand(3), DL, MVT::i64, MVT::i64);
  SDValue Ops[] = {N->getOperand(1), CmpVal.first,  CmpVal.second,
                   NewVal.first,     NewVal.second, N->getOperand(0)};
 
  SDNode *CmpSwap = DAG.getMachineNode(
      Opcode, SDLoc(N), DAG.getVTList(MVT::i64, MVT::i64, MVT::i64, MVT::Other),
      Ops);
  DAG.setNodeMemRefs(cast<MachineSDNode>(CmpSwap), {MemOp});
  Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i128,
                                SDValue(CmpSwap, 0), SDValue(CmpSwap, 1)));
  Results.push_back(SDValue(CmpSwap, 3));
}
 
void LoongArchTargetLowering::ReplaceNodeResults(
    SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
  SDLoc DL(N);
  EVT VT = N->getValueType(0);
  switch (N->getOpcode()) {
  default:
    llvm_unreachable("Don't know how to legalize this operation");
  case ISD::ADD:
  case ISD::SUB:
    assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
           "Unexpected custom legalisation");
    Results.push_back(customLegalizeToWOpWithSExt(N, DAG));
    break;
  case ISD::SDIV:
  case ISD::UDIV:
  case ISD::SREM:
  case ISD::UREM:
    assert(VT == MVT::i32 && Subtarget.is64Bit() &&
           "Unexpected custom legalisation");
    Results.push_back(customLegalizeToWOp(N, DAG, 2,
                                          Subtarget.hasDiv32() && VT == MVT::i32
                                              ? ISD::ANY_EXTEND
                                              : ISD::SIGN_EXTEND));
    break;
  case ISD::SHL:
  case ISD::SRA:
  case ISD::SRL:
    assert(VT == MVT::i32 && Subtarget.is64Bit() &&
           "Unexpected custom legalisation");
    if (N->getOperand(1).getOpcode() != ISD::Constant) {
      Results.push_back(customLegalizeToWOp(N, DAG, 2));
      break;
    }
    break;
  case ISD::ROTL:
  case ISD::ROTR:
    assert(VT == MVT::i32 && Subtarget.is64Bit() &&
           "Unexpected custom legalisation");
    Results.push_back(customLegalizeToWOp(N, DAG, 2));
    break;
  case ISD::LOAD: {
    // Use an f64 load and a scalar_to_vector for v2f32 loads. This avoids
    // scalarizing in 32-bit mode. In 64-bit mode this avoids a int->fp
    // cast since type legalization will try to use an i64 load.
    MVT VT = N->getSimpleValueType(0);
    assert(VT == MVT::v2f32 && Subtarget.hasExtLSX() &&
           "Unexpected custom legalisation");
    assert(getTypeAction(*DAG.getContext(), VT) == TypeWidenVector &&
           "Unexpected type action!");
    if (!ISD::isNON_EXTLoad(N))
      return;
    auto *Ld = cast<LoadSDNode>(N);
    SDValue Res = DAG.getLoad(MVT::f64, DL, Ld->getChain(), Ld->getBasePtr(),
                              Ld->getPointerInfo(), Ld->getBaseAlign(),
                              Ld->getMemOperand()->getFlags());
    SDValue Chain = Res.getValue(1);
    MVT VecVT = MVT::getVectorVT(MVT::f64, 2);
    Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Res);
    EVT WideVT = getTypeToTransformTo(*DAG.getContext(), VT);
    Res = DAG.getBitcast(WideVT, Res);
    Results.push_back(Res);
    Results.push_back(Chain);
    break;
  }
  case ISD::FP_TO_SINT: {
    assert(VT == MVT::i32 && Subtarget.is64Bit() &&
           "Unexpected custom legalisation");
    SDValue Src = N->getOperand(0);
    EVT FVT = EVT::getFloatingPointVT(N->getValueSizeInBits(0));
    if (getTypeAction(*DAG.getContext(), Src.getValueType()) !=
        TargetLowering::TypeSoftenFloat) {
      if (!isTypeLegal(Src.getValueType()))
        return;
      if (Src.getValueType() == MVT::f16)
        Src = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Src);
      SDValue Dst = DAG.getNode(LoongArchISD::FTINT, DL, FVT, Src);
      Results.push_back(DAG.getNode(ISD::BITCAST, DL, VT, Dst));
      return;
    }
    // If the FP type needs to be softened, emit a library call using the 'si'
    // version. If we left it to default legalization we'd end up with 'di'.
    RTLIB::Libcall LC;
    LC = RTLIB::getFPTOSINT(Src.getValueType(), VT);
    MakeLibCallOptions CallOptions;
    EVT OpVT = Src.getValueType();
    CallOptions.setTypeListBeforeSoften(OpVT, VT);
    SDValue Chain = SDValue();
    SDValue Result;
    std::tie(Result, Chain) =
        makeLibCall(DAG, LC, VT, Src, CallOptions, DL, Chain);
    Results.push_back(Result);
    break;
  }
  case ISD::BITCAST: {
    SDValue Src = N->getOperand(0);
    EVT SrcVT = Src.getValueType();
    if (VT == MVT::i32 && SrcVT == MVT::f32 && Subtarget.is64Bit() &&
        Subtarget.hasBasicF()) {
      SDValue Dst =
          DAG.getNode(LoongArchISD::MOVFR2GR_S_LA64, DL, MVT::i64, Src);
      Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Dst));
    } else if (VT == MVT::i64 && SrcVT == MVT::f64 && !Subtarget.is64Bit()) {
      SDValue NewReg = DAG.getNode(LoongArchISD::SPLIT_PAIR_F64, DL,
                                   DAG.getVTList(MVT::i32, MVT::i32), Src);
      SDValue RetReg = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
                                   NewReg.getValue(0), NewReg.getValue(1));
      Results.push_back(RetReg);
    }
    break;
  }
  case ISD::FP_TO_UINT: {
    assert(VT == MVT::i32 && Subtarget.is64Bit() &&
           "Unexpected custom legalisation");
    auto &TLI = DAG.getTargetLoweringInfo();
    SDValue Tmp1, Tmp2;
    TLI.expandFP_TO_UINT(N, Tmp1, Tmp2, DAG);
    Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Tmp1));
    break;
  }
  case ISD::FP_ROUND: {
    assert(VT == MVT::v2f32 && Subtarget.hasExtLSX() &&
           "Unexpected custom legalisation");
    // On LSX platforms, rounding from v2f64 to v4f32 (after legalization from
    // v2f32) is scalarized. Add a customized v2f32 widening to convert it into
    // a target-specific LoongArchISD::VFCVT to optimize it.
    SDValue Op0 = N->getOperand(0);
    EVT OpVT = Op0.getValueType();
    if (OpVT == MVT::v2f64) {
      SDValue Undef = DAG.getUNDEF(OpVT);
      SDValue Dst =
          DAG.getNode(LoongArchISD::VFCVT, DL, MVT::v4f32, Undef, Op0);
      Results.push_back(Dst);
    }
    break;
  }
  case ISD::BSWAP: {
    SDValue Src = N->getOperand(0);
    assert((VT == MVT::i16 || VT == MVT::i32) &&
           "Unexpected custom legalization");
    MVT GRLenVT = Subtarget.getGRLenVT();
    SDValue NewSrc = DAG.getNode(ISD::ANY_EXTEND, DL, GRLenVT, Src);
    SDValue Tmp;
    switch (VT.getSizeInBits()) {
    default:
      llvm_unreachable("Unexpected operand width");
    case 16:
      Tmp = DAG.getNode(LoongArchISD::REVB_2H, DL, GRLenVT, NewSrc);
      break;
    case 32:
      // Only LA64 will get to here due to the size mismatch between VT and
      // GRLenVT, LA32 lowering is directly defined in LoongArchInstrInfo.
      Tmp = DAG.getNode(LoongArchISD::REVB_2W, DL, GRLenVT, NewSrc);
      break;
    }
    Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, Tmp));
    break;
  }
  case ISD::BITREVERSE: {
    SDValue Src = N->getOperand(0);
    assert((VT == MVT::i8 || (VT == MVT::i32 && Subtarget.is64Bit())) &&
           "Unexpected custom legalization");
    MVT GRLenVT = Subtarget.getGRLenVT();
    SDValue NewSrc = DAG.getNode(ISD::ANY_EXTEND, DL, GRLenVT, Src);
    SDValue Tmp;
    switch (VT.getSizeInBits()) {
    default:
      llvm_unreachable("Unexpected operand width");
    case 8:
      Tmp = DAG.getNode(LoongArchISD::BITREV_4B, DL, GRLenVT, NewSrc);
      break;
    case 32:
      Tmp = DAG.getNode(LoongArchISD::BITREV_W, DL, GRLenVT, NewSrc);
      break;
    }
    Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, Tmp));
    break;
  }
  case ISD::CTLZ:
  case ISD::CTTZ: {
    assert(VT == MVT::i32 && Subtarget.is64Bit() &&
           "Unexpected custom legalisation");
    Results.push_back(customLegalizeToWOp(N, DAG, 1));
    break;
  }
  case ISD::INTRINSIC_W_CHAIN: {
    SDValue Chain = N->getOperand(0);
    SDValue Op2 = N->getOperand(2);
    MVT GRLenVT = Subtarget.getGRLenVT();
    const StringRef ErrorMsgOOR = "argument out of range";
    const StringRef ErrorMsgReqLA64 = "requires loongarch64";
    const StringRef ErrorMsgReqF = "requires basic 'f' target feature";
 
    switch (N->getConstantOperandVal(1)) {
    default:
      llvm_unreachable("Unexpected Intrinsic.");
    case Intrinsic::loongarch_movfcsr2gr: {
      if (!Subtarget.hasBasicF()) {
        emitErrorAndReplaceIntrinsicResults(N, Results, DAG, ErrorMsgReqF);
        return;
      }
      unsigned Imm = Op2->getAsZExtVal();
      if (!isUInt<2>(Imm)) {
        emitErrorAndReplaceIntrinsicResults(N, Results, DAG, ErrorMsgOOR);
        return;
      }
      SDValue MOVFCSR2GRResults = DAG.getNode(
          LoongArchISD::MOVFCSR2GR, SDLoc(N), {MVT::i64, MVT::Other},
          {Chain, DAG.getConstant(Imm, DL, GRLenVT)});
      Results.push_back(
          DAG.getNode(ISD::TRUNCATE, DL, VT, MOVFCSR2GRResults.getValue(0)));
      Results.push_back(MOVFCSR2GRResults.getValue(1));
      break;
    }
#define CRC_CASE_EXT_BINARYOP(NAME, NODE)                                      \
  case Intrinsic::loongarch_##NAME: {                                          \
    SDValue NODE = DAG.getNode(                                                \
        LoongArchISD::NODE, DL, {MVT::i64, MVT::Other},                        \
        {Chain, DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op2),               \
         DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(3))});       \
    Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, NODE.getValue(0)));   \
    Results.push_back(NODE.getValue(1));                                       \
    break;                                                                     \
  }
      CRC_CASE_EXT_BINARYOP(crc_w_b_w, CRC_W_B_W)
      CRC_CASE_EXT_BINARYOP(crc_w_h_w, CRC_W_H_W)
      CRC_CASE_EXT_BINARYOP(crc_w_w_w, CRC_W_W_W)
      CRC_CASE_EXT_BINARYOP(crcc_w_b_w, CRCC_W_B_W)
      CRC_CASE_EXT_BINARYOP(crcc_w_h_w, CRCC_W_H_W)
      CRC_CASE_EXT_BINARYOP(crcc_w_w_w, CRCC_W_W_W)
#undef CRC_CASE_EXT_BINARYOP
 
#define CRC_CASE_EXT_UNARYOP(NAME, NODE)                                       \
  case Intrinsic::loongarch_##NAME: {                                          \
    SDValue NODE = DAG.getNode(                                                \
        LoongArchISD::NODE, DL, {MVT::i64, MVT::Other},                        \
        {Chain, Op2,                                                           \
         DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(3))});       \
    Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, NODE.getValue(0)));   \
    Results.push_back(NODE.getValue(1));                                       \
    break;                                                                     \
  }
      CRC_CASE_EXT_UNARYOP(crc_w_d_w, CRC_W_D_W)
      CRC_CASE_EXT_UNARYOP(crcc_w_d_w, CRCC_W_D_W)
#undef CRC_CASE_EXT_UNARYOP
#define CSR_CASE(ID)                                                           \
  case Intrinsic::loongarch_##ID: {                                            \
    if (!Subtarget.is64Bit())                                                  \
      emitErrorAndReplaceIntrinsicResults(N, Results, DAG, ErrorMsgReqLA64);   \
    break;                                                                     \
  }
      CSR_CASE(csrrd_d);
      CSR_CASE(csrwr_d);
      CSR_CASE(csrxchg_d);
      CSR_CASE(iocsrrd_d);
#undef CSR_CASE
    case Intrinsic::loongarch_csrrd_w: {
      unsigned Imm = Op2->getAsZExtVal();
      if (!isUInt<14>(Imm)) {
        emitErrorAndReplaceIntrinsicResults(N, Results, DAG, ErrorMsgOOR);
        return;
      }
      SDValue CSRRDResults =
          DAG.getNode(LoongArchISD::CSRRD, DL, {GRLenVT, MVT::Other},
                      {Chain, DAG.getConstant(Imm, DL, GRLenVT)});
      Results.push_back(
          DAG.getNode(ISD::TRUNCATE, DL, VT, CSRRDResults.getValue(0)));
      Results.push_back(CSRRDResults.getValue(1));
      break;
    }
    case Intrinsic::loongarch_csrwr_w: {
      unsigned Imm = N->getConstantOperandVal(3);
      if (!isUInt<14>(Imm)) {
        emitErrorAndReplaceIntrinsicResults(N, Results, DAG, ErrorMsgOOR);
        return;
      }
      SDValue CSRWRResults =
          DAG.getNode(LoongArchISD::CSRWR, DL, {GRLenVT, MVT::Other},
                      {Chain, DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op2),
                       DAG.getConstant(Imm, DL, GRLenVT)});
      Results.push_back(
          DAG.getNode(ISD::TRUNCATE, DL, VT, CSRWRResults.getValue(0)));
      Results.push_back(CSRWRResults.getValue(1));
      break;
    }
    case Intrinsic::loongarch_csrxchg_w: {
      unsigned Imm = N->getConstantOperandVal(4);
      if (!isUInt<14>(Imm)) {
        emitErrorAndReplaceIntrinsicResults(N, Results, DAG, ErrorMsgOOR);
        return;
      }
      SDValue CSRXCHGResults = DAG.getNode(
          LoongArchISD::CSRXCHG, DL, {GRLenVT, MVT::Other},
          {Chain, DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op2),
           DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(3)),
           DAG.getConstant(Imm, DL, GRLenVT)});
      Results.push_back(
          DAG.getNode(ISD::TRUNCATE, DL, VT, CSRXCHGResults.getValue(0)));
      Results.push_back(CSRXCHGResults.getValue(1));
      break;
    }
#define IOCSRRD_CASE(NAME, NODE)                                               \
  case Intrinsic::loongarch_##NAME: {                                          \
    SDValue IOCSRRDResults =                                                   \
        DAG.getNode(LoongArchISD::NODE, DL, {MVT::i64, MVT::Other},            \
                    {Chain, DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op2)}); \
    Results.push_back(                                                         \
        DAG.getNode(ISD::TRUNCATE, DL, VT, IOCSRRDResults.getValue(0)));       \
    Results.push_back(IOCSRRDResults.getValue(1));                             \
    break;                                                                     \
  }
      IOCSRRD_CASE(iocsrrd_b, IOCSRRD_B);
      IOCSRRD_CASE(iocsrrd_h, IOCSRRD_H);
      IOCSRRD_CASE(iocsrrd_w, IOCSRRD_W);
#undef IOCSRRD_CASE
    case Intrinsic::loongarch_cpucfg: {
      SDValue CPUCFGResults =
          DAG.getNode(LoongArchISD::CPUCFG, DL, {GRLenVT, MVT::Other},
                      {Chain, DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op2)});
      Results.push_back(
          DAG.getNode(ISD::TRUNCATE, DL, VT, CPUCFGResults.getValue(0)));
      Results.push_back(CPUCFGResults.getValue(1));
      break;
    }
    case Intrinsic::loongarch_lddir_d: {
      if (!Subtarget.is64Bit()) {
        emitErrorAndReplaceIntrinsicResults(N, Results, DAG, ErrorMsgReqLA64);
        return;
      }
      break;
    }
    }
    break;
  }
  case ISD::READ_REGISTER: {
    if (Subtarget.is64Bit())
      DAG.getContext()->emitError(
          "On LA64, only 64-bit registers can be read.");
    else
      DAG.getContext()->emitError(
          "On LA32, only 32-bit registers can be read.");
    Results.push_back(DAG.getUNDEF(VT));
    Results.push_back(N->getOperand(0));
    break;
  }
  case ISD::INTRINSIC_WO_CHAIN: {
    replaceINTRINSIC_WO_CHAINResults(N, Results, DAG, Subtarget);
    break;
  }
  case ISD::LROUND: {
    SDValue Op0 = N->getOperand(0);
    EVT OpVT = Op0.getValueType();
    RTLIB::Libcall LC =
        OpVT == MVT::f64 ? RTLIB::LROUND_F64 : RTLIB::LROUND_F32;
    MakeLibCallOptions CallOptions;
    CallOptions.setTypeListBeforeSoften(OpVT, MVT::i64);
    SDValue Result = makeLibCall(DAG, LC, MVT::i64, Op0, CallOptions, DL).first;
    Result = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Result);
    Results.push_back(Result);
    break;
  }
  case ISD::ATOMIC_CMP_SWAP: {
    replaceCMP_XCHG_128Results(N, Results, DAG);
    break;
  }
  case ISD::TRUNCATE: {
    MVT VT = N->getSimpleValueType(0);
    if (getTypeAction(*DAG.getContext(), VT) != TypeWidenVector)
      return;
 
    MVT WidenVT = getTypeToTransformTo(*DAG.getContext(), VT).getSimpleVT();
    SDValue In = N->getOperand(0);
    EVT InVT = In.getValueType();
    EVT InEltVT = InVT.getVectorElementType();
    EVT EltVT = VT.getVectorElementType();
    unsigned MinElts = VT.getVectorNumElements();
    unsigned WidenNumElts = WidenVT.getVectorNumElements();
    unsigned InBits = InVT.getSizeInBits();
 
    // v8i64 -> (v8i32) -> v8i8
    if (InVT == MVT::v8i64 && WidenVT.is128BitVector()) {
      InVT = MVT::getVectorVT(MVT::getIntegerVT(256 / MinElts), MinElts);
      In = DAG.getNode(N->getOpcode(), DL, InVT, In);
      InBits = 256;
    }
 
    // v8i32 -> v8i8 / v4i64 -> v4i16 / v4i64 -> v4i8
    if ((InVT == MVT::v8i32 || InVT == MVT::v4i64) &&
        WidenVT.is128BitVector()) {
      InVT = MVT::getVectorVT(MVT::getIntegerVT(128 / MinElts), MinElts);
      In = DAG.getNode(N->getOpcode(), DL, InVT, In);
      InBits = 128;
      InEltVT = InVT.getVectorElementType();
    }
 
    if ((128 % InBits) == 0 && WidenVT.is128BitVector()) {
      if ((InEltVT.getSizeInBits() % EltVT.getSizeInBits()) == 0) {
        int Scale = InEltVT.getSizeInBits() / EltVT.getSizeInBits();
        SmallVector<int, 16> TruncMask(WidenNumElts, -1);
        for (unsigned I = 0; I < MinElts; ++I)
          TruncMask[I] = Scale * I;
 
        unsigned WidenNumElts = 128 / In.getScalarValueSizeInBits();
        MVT SVT = In.getSimpleValueType().getScalarType();
        MVT VT = MVT::getVectorVT(SVT, WidenNumElts);
        SDValue WidenIn =
            DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), In,
                        DAG.getVectorIdxConstant(0, DL));
        assert(isTypeLegal(WidenVT) && isTypeLegal(WidenIn.getValueType()) &&
               "Illegal vector type in truncation");
        WidenIn = DAG.getBitcast(WidenVT, WidenIn);
        Results.push_back(
            DAG.getVectorShuffle(WidenVT, DL, WidenIn, WidenIn, TruncMask));
        return;
      }
    }
 
    break;
  }
  case ISD::SIGN_EXTEND: {
    // LASX has native VEXT2XV_* for sign extension.
    if (!Subtarget.hasExtLSX() || Subtarget.hasExtLASX())
      return;
 
    EVT DstVT = N->getValueType(0);
    SDValue Src = N->getOperand(0);
    MVT SrcVT = Src.getSimpleValueType();
 
    unsigned SrcEltBits = SrcVT.getScalarSizeInBits();
    unsigned DstEltBits = DstVT.getScalarSizeInBits();
    unsigned NumElts = DstVT.getVectorNumElements();
 
    if (SrcVT.getSizeInBits() > 128)
      return;
 
    if (!DstVT.isVector() || DstVT.getSizeInBits() <= 128)
      return;
 
    // Legalize and extend the src to 128-bit first.
    if (SrcVT.getSizeInBits() < 128) {
      unsigned WidenSrcElts = 128 / SrcEltBits;
      MVT WidenSrcVT = MVT::getVectorVT(SrcVT.getScalarType(), WidenSrcElts);
      Src = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WidenSrcVT,
                        DAG.getUNDEF(WidenSrcVT), Src,
                        DAG.getVectorIdxConstant(0, DL));
      SrcVT = WidenSrcVT;
 
      unsigned FirstStageEltBits = 128 / NumElts;
      MVT FirstStageEltVT = MVT::getIntegerVT(FirstStageEltBits);
      MVT FirstStageVT = MVT::getVectorVT(FirstStageEltVT, NumElts);
      Src = DAG.getNode(ISD::SIGN_EXTEND_VECTOR_INREG, DL, FirstStageVT, Src);
      SrcVT = FirstStageVT;
      SrcEltBits = FirstStageEltBits;
    }
 
    SmallVector<SDValue, 8> Blocks;
    Blocks.push_back(Src);
 
    // Sign-extend the src by using SLTI + VILVL + VILVH recursively.
    while (SrcEltBits < DstEltBits) {
      unsigned NextEltBits = SrcEltBits * 2;
      MVT NextEltVT = MVT::getIntegerVT(NextEltBits);
      unsigned CurEltsPerBlock = SrcVT.getVectorNumElements();
      unsigned NextEltsPerBlock = CurEltsPerBlock / 2;
      MVT NextBlockVT = MVT::getVectorVT(NextEltVT, NextEltsPerBlock);
 
      SmallVector<SDValue, 8> NextBlocks;
      NextBlocks.reserve(Blocks.size() * 2);
      for (SDValue Block : Blocks) {
        SDValue Zero = DAG.getConstant(0, DL, SrcVT);
        SDValue Mask = DAG.getNode(ISD::SETCC, DL, SrcVT, Block, Zero,
                                   DAG.getCondCode(ISD::SETLT));
        SDValue LoInterleaved =
            DAG.getNode(LoongArchISD::VILVL, DL, SrcVT, Mask, Block);
        SDValue HiInterleaved =
            DAG.getNode(LoongArchISD::VILVH, DL, SrcVT, Mask, Block);
 
        NextBlocks.push_back(DAG.getBitcast(NextBlockVT, LoInterleaved));
        NextBlocks.push_back(DAG.getBitcast(NextBlockVT, HiInterleaved));
      }
 
      Blocks = std::move(NextBlocks);
      SrcVT = NextBlockVT;
      SrcEltBits = NextEltBits;
    }
 
    Results.push_back(DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Blocks));
    break;
  }
  case ISD::FP_EXTEND:
    // FP_EXTEND may reach here due to the Custom action for v2f32 results, but
    // no target-specific lowering is required. Leave it unchanged and rely on
    // the default type legalization.
    break;
  }
}
 
/// Try to fold: (and (xor X, -1), Y) -> (vandn X, Y).
static SDValue combineAndNotIntoVANDN(SDNode *N, const SDLoc &DL,
                                      SelectionDAG &DAG) {
  assert(N->getOpcode() == ISD::AND && "Unexpected opcode combine into ANDN");
 
  MVT VT = N->getSimpleValueType(0);
  if (!VT.is128BitVector() && !VT.is256BitVector())
    return SDValue();
 
  SDValue X, Y;
  SDValue N0 = N->getOperand(0);
  SDValue N1 = N->getOperand(1);
 
  if (SDValue Not = isNOT(N0, DAG)) {
    X = Not;
    Y = N1;
  } else if (SDValue Not = isNOT(N1, DAG)) {
    X = Not;
    Y = N0;
  } else
    return SDValue();
 
  X = DAG.getBitcast(VT, X);
  Y = DAG.getBitcast(VT, Y);
  return DAG.getNode(LoongArchISD::VANDN, DL, VT, X, Y);
}
 
static bool isConstantSplatVector(SDValue N, APInt &SplatValue,
                                  unsigned MinSizeInBits) {
  N = peekThroughBitcasts(N);
  BuildVectorSDNode *Node = dyn_cast<BuildVectorSDNode>(N);
 
  if (!Node)
    return false;
 
  APInt SplatUndef;
  unsigned SplatBitSize;
  bool HasAnyUndefs;
 
  return Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
                               HasAnyUndefs, MinSizeInBits,
                               /*IsBigEndian=*/false);
}
 
static SDValue matchDeinterleaveBuildVector(SDValue N, unsigned &StartIndex) {
  auto *BV = dyn_cast<BuildVectorSDNode>(N);
  if (!BV)
    return SDValue();
 
  SDValue Src;
  int Start = -1;
 
  for (unsigned i = 0, NumElts = BV->getNumOperands(); i < NumElts; ++i) {
    SDValue Op = BV->getOperand(i);
    if (Op.isUndef())
      continue;
    if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
      return SDValue();
 
    auto *IdxC = dyn_cast<ConstantSDNode>(Op.getOperand(1));
    if (!IdxC)
      return SDValue();
 
    unsigned EltIdx = IdxC->getZExtValue();
    if (Start < 0)
      Start = (int)EltIdx - (int)(i * 2);
    if (Start < 0 || Start > 1 || EltIdx != (unsigned)(Start + (int)(i * 2)))
      return SDValue();
 
    SDValue CurSrc = Op.getOperand(0);
    if (!Src)
      Src = CurSrc;
    else if (Src != CurSrc)
      return SDValue();
  }
 
  if (!Src || Start < 0)
    return SDValue();
 
  StartIndex = (unsigned)Start;
  return Src;
}
 
static SDValue
performHorizWideningCombine(SDNode *N, SelectionDAG &DAG,
                            const LoongArchSubtarget &Subtarget) {
  if (!Subtarget.hasExtLSX())
    return SDValue();
 
  unsigned Opc = N->getOpcode();
  assert((Opc == ISD::ADD || Opc == ISD::SUB) && "Unexpected opcode");
 
  EVT VT = N->getValueType(0);
  SDLoc DL(N);
 
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
 
  bool isSigned;
  unsigned ExtOpc = LHS.getOpcode();
  if (ExtOpc == ISD::SIGN_EXTEND)
    isSigned = true;
  else if (ExtOpc == ISD::ZERO_EXTEND)
    isSigned = false;
  else
    return SDValue();
 
  if (ExtOpc != RHS.getOpcode())
    return SDValue();
 
  if (!LHS.hasOneUse() || !RHS.hasOneUse())
    return SDValue();
 
  unsigned OddIdx, EvenIdx;
  SDValue LHSVec = matchDeinterleaveBuildVector(LHS.getOperand(0), OddIdx);
  SDValue RHSVec = matchDeinterleaveBuildVector(RHS.getOperand(0), EvenIdx);
 
  if (!LHSVec || !RHSVec)
    return SDValue();
  if (OddIdx != 1 || EvenIdx != 0)
    return SDValue();
  if (LHSVec.getValueType() != RHSVec.getValueType())
    return SDValue();
 
  EVT SrcVT = LHSVec.getValueType();
  EVT SrcEltVT = SrcVT.getVectorElementType();
  EVT DstEltVT = VT.getVectorElementType();
  auto &TLI = DAG.getTargetLoweringInfo();
 
  if (!TLI.isTypeLegal(VT) || !TLI.isTypeLegal(SrcVT))
    return SDValue();
  if (!SrcVT.isVector() || !VT.isVector())
    return SDValue();
  if (SrcVT.getSizeInBits() != VT.getSizeInBits())
    return SDValue();
  if (DstEltVT.getSizeInBits() != SrcEltVT.getSizeInBits() * 2)
    return SDValue();
  if (!SrcEltVT.isInteger() || SrcEltVT.getSizeInBits() > 32)
    return SDValue();
 
  unsigned TargetOpc;
  if (Opc == ISD::ADD)
    TargetOpc = isSigned ? LoongArchISD::VHADDW : LoongArchISD::VHADDW_U;
  else
    TargetOpc = isSigned ? LoongArchISD::VHSUBW : LoongArchISD::VHSUBW_U;
 
  return DAG.getNode(TargetOpc, DL, VT, LHSVec, RHSVec);
}
 
static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
                                 TargetLowering::DAGCombinerInfo &DCI,
                                 const LoongArchSubtarget &Subtarget) {
  if (SDValue V = performHorizWideningCombine(N, DAG, Subtarget))
    return V;
 
  if (DCI.isBeforeLegalizeOps())
    return SDValue();
 
  EVT VT = N->getValueType(0);
  if (!VT.isVector())
    return SDValue();
 
  if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
    return SDValue();
 
  EVT EltVT = VT.getVectorElementType();
  if (!EltVT.isInteger())
    return SDValue();
 
  // match:
  //
  //   add
  //     (and
  //        (srl X, shift-1) / X
  //        1)
  //     (srl/sra X, shift)
 
  SDValue Add0 = N->getOperand(0);
  SDValue Add1 = N->getOperand(1);
  SDValue And;
  SDValue Shr;
 
  if (Add0.getOpcode() == ISD::AND) {
    And = Add0;
    Shr = Add1;
  } else if (Add1.getOpcode() == ISD::AND) {
    And = Add1;
    Shr = Add0;
  } else {
    return SDValue();
  }
 
  // match:
  //
  //   srl/sra X, shift
 
  if (Shr.getOpcode() != ISD::SRL && Shr.getOpcode() != ISD::SRA)
    return SDValue();
 
  SDValue X = Shr.getOperand(0);
  SDValue Shift = Shr.getOperand(1);
  APInt ShiftVal;
 
  if (!isConstantSplatVector(Shift, ShiftVal, EltVT.getSizeInBits()))
    return SDValue();
 
  if (ShiftVal == 0)
    return SDValue();
 
  // match:
  //
  //   and
  //      (srl X, shift-1) / X
  //      1
 
  SDValue One = And.getOperand(1);
  APInt SplatVal;
 
  if (!isConstantSplatVector(One, SplatVal, EltVT.getSizeInBits()))
    return SDValue();
 
  if (SplatVal != 1)
    return SDValue();
 
  if (And.getOperand(0) == X) {
    // match:
    //
    // shift == 1
 
    if (ShiftVal != 1)
      return SDValue();
  } else {
    // match:
    //
    // srl X, shift-1
 
    SDValue Srl = And.getOperand(0);
 
    if (Srl.getOpcode() != ISD::SRL)
      return SDValue();
 
    if (Srl.getOperand(0) != X)
      return SDValue();
 
    // match:
    //
    // shift-1
 
    SDValue ShiftMinus1 = Srl.getOperand(1);
 
    if (!isConstantSplatVector(ShiftMinus1, SplatVal, EltVT.getSizeInBits()))
      return SDValue();
 
    if (ShiftVal != (SplatVal + 1))
      return SDValue();
  }
 
  // We matched a rounded right shift pattern and can lower it
  // to a single vector rounded shift instruction.
 
  SDLoc DL(N);
  return DAG.getNode(Shr.getOpcode() == ISD::SRL ? LoongArchISD::VSRLR
                                                 : LoongArchISD::VSRAR,
                     DL, VT, X, Shift);
}
 
static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
                                 TargetLowering::DAGCombinerInfo &DCI,
                                 const LoongArchSubtarget &Subtarget) {
  if (DCI.isBeforeLegalizeOps())
    return SDValue();
 
  SDValue FirstOperand = N->getOperand(0);
  SDValue SecondOperand = N->getOperand(1);
  unsigned FirstOperandOpc = FirstOperand.getOpcode();
  EVT ValTy = N->getValueType(0);
  SDLoc DL(N);
  uint64_t lsb, msb;
  unsigned SMIdx, SMLen;
  ConstantSDNode *CN;
  SDValue NewOperand;
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  if (SDValue R = combineAndNotIntoVANDN(N, DL, DAG))
    return R;
 
  // BSTRPICK requires the 32S feature.
  if (!Subtarget.has32S())
    return SDValue();
 
  // Op's second operand must be a shifted mask.
  if (!(CN = dyn_cast<ConstantSDNode>(SecondOperand)) ||
      !isShiftedMask_64(CN->getZExtValue(), SMIdx, SMLen))
    return SDValue();
 
  if (FirstOperandOpc == ISD::SRA || FirstOperandOpc == ISD::SRL) {
    // Pattern match BSTRPICK.
    //  $dst = and ((sra or srl) $src , lsb), (2**len - 1)
    //  => BSTRPICK $dst, $src, msb, lsb
    //  where msb = lsb + len - 1
 
    // The second operand of the shift must be an immediate.
    if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))))
      return SDValue();
 
    lsb = CN->getZExtValue();
 
    // Return if the shifted mask does not start at bit 0 or the sum of its
    // length and lsb exceeds the word's size.
    if (SMIdx != 0 || lsb + SMLen > ValTy.getSizeInBits())
      return SDValue();
 
    NewOperand = FirstOperand.getOperand(0);
  } else {
    // Pattern match BSTRPICK.
    //  $dst = and $src, (2**len- 1) , if len > 12
    //  => BSTRPICK $dst, $src, msb, lsb
    //  where lsb = 0 and msb = len - 1
 
    // If the mask is <= 0xfff, andi can be used instead.
    if (CN->getZExtValue() <= 0xfff)
      return SDValue();
 
    // Return if the MSB exceeds.
    if (SMIdx + SMLen > ValTy.getSizeInBits())
      return SDValue();
 
    if (SMIdx > 0) {
      // Omit if the constant has more than 2 uses. This a conservative
      // decision. Whether it is a win depends on the HW microarchitecture.
      // However it should always be better for 1 and 2 uses.
      if (CN->use_size() > 2)
        return SDValue();
      // Return if the constant can be composed by a single LU12I.W.
      if ((CN->getZExtValue() & 0xfff) == 0)
        return SDValue();
      // Return if the constand can be composed by a single ADDI with
      // the zero register.
      if (CN->getSExtValue() >= -2048 && CN->getSExtValue() < 0)
        return SDValue();
    }
 
    lsb = SMIdx;
    NewOperand = FirstOperand;
  }
 
  msb = lsb + SMLen - 1;
  SDValue NR0 = DAG.getNode(LoongArchISD::BSTRPICK, DL, ValTy, NewOperand,
                            DAG.getConstant(msb, DL, GRLenVT),
                            DAG.getConstant(lsb, DL, GRLenVT));
  if (FirstOperandOpc == ISD::SRA || FirstOperandOpc == ISD::SRL || lsb == 0)
    return NR0;
  // Try to optimize to
  //   bstrpick $Rd, $Rs, msb, lsb
  //   slli     $Rd, $Rd, lsb
  return DAG.getNode(ISD::SHL, DL, ValTy, NR0,
                     DAG.getConstant(lsb, DL, GRLenVT));
}
 
// Return the original source vector if N consists of the half
// of each 128-bit lane.
static SDValue matchHalfOf128BitLanes(SDValue N, bool isLow) {
  N = peekThroughBitcasts(N);
 
  EVT DstVT = N.getValueType();
  if (!DstVT.isVector())
    return SDValue();
 
  unsigned NumElts = DstVT.getVectorNumElements();
 
  // LSX canonical form:
  if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
    SDValue Src = N.getOperand(0);
    EVT SrcVT = Src.getValueType();
 
    if (!SrcVT.isVector() || !SrcVT.is128BitVector())
      return SDValue();
    if (SrcVT.getSizeInBits() != DstVT.getSizeInBits() * 2)
      return SDValue();
    if (SrcVT.getVectorNumElements() != NumElts * 2)
      return SDValue();
    if (N.getConstantOperandVal(1) != (isLow ? 0 : NumElts))
      return SDValue();
 
    return Src;
  }
 
  // LASX canonical form:
  auto *BV = dyn_cast<BuildVectorSDNode>(N);
  if (!BV)
    return SDValue();
 
  if (NumElts % 2 != 0)
    return SDValue();
 
  SDValue Src;
  EVT SrcVT;
 
  for (unsigned I = 0; I != NumElts; ++I) {
    SDValue Elt = BV->getOperand(I);
    if (Elt.isUndef())
      continue;
    if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
      return SDValue();
 
    SDValue ThisSrc = Elt.getOperand(0);
    SDValue Idx = Elt.getOperand(1);
    auto *CI = dyn_cast<ConstantSDNode>(Idx);
    if (!CI)
      return SDValue();
 
    if (!Src) {
      Src = ThisSrc;
      SrcVT = Src.getValueType();
      if (!SrcVT.isVector())
        return SDValue();
 
      if (!SrcVT.is256BitVector())
        return SDValue();
      if (SrcVT.getSizeInBits() != DstVT.getSizeInBits() * 2)
        return SDValue();
      if (SrcVT.getVectorNumElements() != NumElts * 2)
        return SDValue();
    } else if (ThisSrc != Src) {
      return SDValue();
    }
 
    unsigned Half = NumElts / 2;
    unsigned ExpectedIdx = (I < Half) ? I : (I + Half);
    ExpectedIdx += isLow ? 0 : Half;
 
    if (CI->getZExtValue() != ExpectedIdx)
      return SDValue();
  }
 
  return Src;
}
 
static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG,
                                 TargetLowering::DAGCombinerInfo &DCI,
                                 const LoongArchSubtarget &Subtarget) {
  assert(N->getOpcode() == ISD::SHL && "Unexpected opcode");
 
  EVT VT = N->getValueType(0);
  SDLoc DL(N);
 
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
 
  bool isSigned;
  unsigned ExtOpc = LHS.getOpcode();
  if (ExtOpc == ISD::SIGN_EXTEND)
    isSigned = true;
  else if (ExtOpc == ISD::ZERO_EXTEND)
    isSigned = false;
  else
    return SDValue();
 
  if (!LHS.hasOneUse())
    return SDValue();
 
  if (!DAG.getTargetLoweringInfo().isTypeLegal(VT) ||
      N->getValueSizeInBits(0) != LHS->getOperand(0).getValueSizeInBits() * 2)
    return SDValue();
 
  SDValue Vec = matchHalfOf128BitLanes(LHS.getOperand(0), /*isLow=*/true);
  if (!Vec)
    return SDValue();
 
  EVT SrcVT = Vec.getValueType();
  EVT SrcEltVT = SrcVT.getVectorElementType();
  EVT DstEltVT = VT.getVectorElementType();
  APInt Imm;
  if (!isConstantSplatVector(RHS, Imm, DstEltVT.getSizeInBits()))
    return SDValue();
  if (!Imm.ult(SrcEltVT.getSizeInBits()))
    return SDValue();
 
  unsigned Opc = isSigned ? LoongArchISD::VSLLWIL : LoongArchISD::VSLLWIL_U;
  SDValue Sht = DAG.getConstant(Imm.getZExtValue(), DL, Subtarget.getGRLenVT());
  return DAG.getNode(Opc, DL, VT, Vec, Sht);
}
 
static SDValue performSRLCombine(SDNode *N, SelectionDAG &DAG,
                                 TargetLowering::DAGCombinerInfo &DCI,
                                 const LoongArchSubtarget &Subtarget) {
  // BSTRPICK requires the 32S feature.
  if (!Subtarget.has32S())
    return SDValue();
 
  if (DCI.isBeforeLegalizeOps())
    return SDValue();
 
  // $dst = srl (and $src, Mask), Shamt
  // =>
  // BSTRPICK $dst, $src, MaskIdx+MaskLen-1, Shamt
  // when Mask is a shifted mask, and MaskIdx <= Shamt <= MaskIdx+MaskLen-1
  //
 
  SDValue FirstOperand = N->getOperand(0);
  ConstantSDNode *CN;
  EVT ValTy = N->getValueType(0);
  SDLoc DL(N);
  MVT GRLenVT = Subtarget.getGRLenVT();
  unsigned MaskIdx, MaskLen;
  uint64_t Shamt;
 
  // The first operand must be an AND and the second operand of the AND must be
  // a shifted mask.
  if (FirstOperand.getOpcode() != ISD::AND ||
      !(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))) ||
      !isShiftedMask_64(CN->getZExtValue(), MaskIdx, MaskLen))
    return SDValue();
 
  // The second operand (shift amount) must be an immediate.
  if (!(CN = dyn_cast<ConstantSDNode>(N->getOperand(1))))
    return SDValue();
 
  Shamt = CN->getZExtValue();
  if (MaskIdx <= Shamt && Shamt <= MaskIdx + MaskLen - 1)
    return DAG.getNode(LoongArchISD::BSTRPICK, DL, ValTy,
                       FirstOperand->getOperand(0),
                       DAG.getConstant(MaskIdx + MaskLen - 1, DL, GRLenVT),
                       DAG.getConstant(Shamt, DL, GRLenVT));
 
  return SDValue();
}
 
static SDValue performSUBCombine(SDNode *N, SelectionDAG &DAG,
                                 TargetLowering::DAGCombinerInfo &DCI,
                                 const LoongArchSubtarget &Subtarget) {
  if (SDValue V = performHorizWideningCombine(N, DAG, Subtarget))
    return V;
 
  return SDValue();
}
 
// Helper to peek through bitops/trunc/setcc to determine size of source vector.
// Allows BITCASTCombine to determine what size vector generated a <X x i1>.
static bool checkBitcastSrcVectorSize(SDValue Src, unsigned Size,
                                      unsigned Depth) {
  // Limit recursion.
  if (Depth >= SelectionDAG::MaxRecursionDepth)
    return false;
  switch (Src.getOpcode()) {
  case ISD::SETCC:
  case ISD::TRUNCATE:
    return Src.getOperand(0).getValueSizeInBits() == Size;
  case ISD::FREEZE:
    return checkBitcastSrcVectorSize(Src.getOperand(0), Size, Depth + 1);
  case ISD::AND:
  case ISD::XOR:
  case ISD::OR:
    return checkBitcastSrcVectorSize(Src.getOperand(0), Size, Depth + 1) &&
           checkBitcastSrcVectorSize(Src.getOperand(1), Size, Depth + 1);
  case ISD::SELECT:
  case ISD::VSELECT:
    return Src.getOperand(0).getScalarValueSizeInBits() == 1 &&
           checkBitcastSrcVectorSize(Src.getOperand(1), Size, Depth + 1) &&
           checkBitcastSrcVectorSize(Src.getOperand(2), Size, Depth + 1);
  case ISD::BUILD_VECTOR:
    return ISD::isBuildVectorAllZeros(Src.getNode()) ||
           ISD::isBuildVectorAllOnes(Src.getNode());
  }
  return false;
}
 
// Helper to push sign extension of vXi1 SETCC result through bitops.
static SDValue signExtendBitcastSrcVector(SelectionDAG &DAG, EVT SExtVT,
                                          SDValue Src, const SDLoc &DL) {
  switch (Src.getOpcode()) {
  case ISD::SETCC:
  case ISD::FREEZE:
  case ISD::TRUNCATE:
  case ISD::BUILD_VECTOR:
    return DAG.getNode(ISD::SIGN_EXTEND, DL, SExtVT, Src);
  case ISD::AND:
  case ISD::XOR:
  case ISD::OR:
    return DAG.getNode(
        Src.getOpcode(), DL, SExtVT,
        signExtendBitcastSrcVector(DAG, SExtVT, Src.getOperand(0), DL),
        signExtendBitcastSrcVector(DAG, SExtVT, Src.getOperand(1), DL));
  case ISD::SELECT:
  case ISD::VSELECT:
    return DAG.getSelect(
        DL, SExtVT, Src.getOperand(0),
        signExtendBitcastSrcVector(DAG, SExtVT, Src.getOperand(1), DL),
        signExtendBitcastSrcVector(DAG, SExtVT, Src.getOperand(2), DL));
  }
  llvm_unreachable("Unexpected node type for vXi1 sign extension");
}
 
static SDValue
performSETCC_BITCASTCombine(SDNode *N, SelectionDAG &DAG,
                            TargetLowering::DAGCombinerInfo &DCI,
                            const LoongArchSubtarget &Subtarget) {
  SDLoc DL(N);
  EVT VT = N->getValueType(0);
  SDValue Src = N->getOperand(0);
  EVT SrcVT = Src.getValueType();
 
  if (Src.getOpcode() != ISD::SETCC || !Src.hasOneUse())
    return SDValue();
 
  bool UseLASX;
  unsigned Opc = ISD::DELETED_NODE;
  EVT CmpVT = Src.getOperand(0).getValueType();
  EVT EltVT = CmpVT.getVectorElementType();
 
  if (Subtarget.hasExtLSX() && CmpVT.getSizeInBits() == 128)
    UseLASX = false;
  else if (Subtarget.has32S() && Subtarget.hasExtLASX() &&
           CmpVT.getSizeInBits() == 256)
    UseLASX = true;
  else
    return SDValue();
 
  SDValue SrcN1 = Src.getOperand(1);
  switch (cast<CondCodeSDNode>(Src.getOperand(2))->get()) {
  default:
    break;
  case ISD::SETEQ:
    // x == 0 => not (vmsknez.b x)
    if (ISD::isBuildVectorAllZeros(SrcN1.getNode()) && EltVT == MVT::i8)
      Opc = UseLASX ? LoongArchISD::XVMSKEQZ : LoongArchISD::VMSKEQZ;
    break;
  case ISD::SETGT:
    // x > -1 => vmskgez.b x
    if (ISD::isBuildVectorAllOnes(SrcN1.getNode()) && EltVT == MVT::i8)
      Opc = UseLASX ? LoongArchISD::XVMSKGEZ : LoongArchISD::VMSKGEZ;
    break;
  case ISD::SETGE:
    // x >= 0 => vmskgez.b x
    if (ISD::isBuildVectorAllZeros(SrcN1.getNode()) && EltVT == MVT::i8)
      Opc = UseLASX ? LoongArchISD::XVMSKGEZ : LoongArchISD::VMSKGEZ;
    break;
  case ISD::SETLT:
    // x < 0 => vmskltz.{b,h,w,d} x
    if (ISD::isBuildVectorAllZeros(SrcN1.getNode()) &&
        (EltVT == MVT::i8 || EltVT == MVT::i16 || EltVT == MVT::i32 ||
         EltVT == MVT::i64))
      Opc = UseLASX ? LoongArchISD::XVMSKLTZ : LoongArchISD::VMSKLTZ;
    break;
  case ISD::SETLE:
    // x <= -1 => vmskltz.{b,h,w,d} x
    if (ISD::isBuildVectorAllOnes(SrcN1.getNode()) &&
        (EltVT == MVT::i8 || EltVT == MVT::i16 || EltVT == MVT::i32 ||
         EltVT == MVT::i64))
      Opc = UseLASX ? LoongArchISD::XVMSKLTZ : LoongArchISD::VMSKLTZ;
    break;
  case ISD::SETNE:
    // x != 0 => vmsknez.b x
    if (ISD::isBuildVectorAllZeros(SrcN1.getNode()) && EltVT == MVT::i8)
      Opc = UseLASX ? LoongArchISD::XVMSKNEZ : LoongArchISD::VMSKNEZ;
    break;
  }
 
  if (Opc == ISD::DELETED_NODE)
    return SDValue();
 
  SDValue V = DAG.getNode(Opc, DL, Subtarget.getGRLenVT(), Src.getOperand(0));
  EVT T = EVT::getIntegerVT(*DAG.getContext(), SrcVT.getVectorNumElements());
  V = DAG.getZExtOrTrunc(V, DL, T);
  return DAG.getBitcast(VT, V);
}
 
static SDValue performBITCASTCombine(SDNode *N, SelectionDAG &DAG,
                                     TargetLowering::DAGCombinerInfo &DCI,
                                     const LoongArchSubtarget &Subtarget) {
  SDLoc DL(N);
  EVT VT = N->getValueType(0);
  SDValue Src = N->getOperand(0);
  EVT SrcVT = Src.getValueType();
  MVT GRLenVT = Subtarget.getGRLenVT();
 
  if (!DCI.isBeforeLegalizeOps())
    return SDValue();
 
  if (!SrcVT.isSimple() || SrcVT.getScalarType() != MVT::i1)
    return SDValue();
 
  // Combine SETCC and BITCAST into [X]VMSK{LT,GE,NE} when possible
  SDValue Res = performSETCC_BITCASTCombine(N, DAG, DCI, Subtarget);
  if (Res)
    return Res;
 
  // Generate vXi1 using [X]VMSKLTZ
  MVT SExtVT;
  unsigned Opc;
  bool UseLASX = false;
  bool PropagateSExt = false;
 
  if (Src.getOpcode() == ISD::SETCC && Src.hasOneUse()) {
    EVT CmpVT = Src.getOperand(0).getValueType();
    if (CmpVT.getSizeInBits() > 256)
      return SDValue();
  }
 
  switch (SrcVT.getSimpleVT().SimpleTy) {
  default:
    return SDValue();
  case MVT::v2i1:
    SExtVT = MVT::v2i64;
    break;
  case MVT::v4i1:
    SExtVT = MVT::v4i32;
    if (Subtarget.hasExtLASX() && checkBitcastSrcVectorSize(Src, 256, 0)) {
      SExtVT = MVT::v4i64;
      UseLASX = true;
      PropagateSExt = true;
    }
    break;
  case MVT::v8i1:
    SExtVT = MVT::v8i16;
    if (Subtarget.hasExtLASX() && checkBitcastSrcVectorSize(Src, 256, 0)) {
      SExtVT = MVT::v8i32;
      UseLASX = true;
      PropagateSExt = true;
    }
    break;
  case MVT::v16i1:
    SExtVT = MVT::v16i8;
    if (Subtarget.hasExtLASX() && checkBitcastSrcVectorSize(Src, 256, 0)) {
      SExtVT = MVT::v16i16;
      UseLASX = true;
      PropagateSExt = true;
    }
    break;
  case MVT::v32i1:
    SExtVT = MVT::v32i8;
    UseLASX = true;
    break;
  };
  Src = PropagateSExt ? signExtendBitcastSrcVector(DAG, SExtVT, Src, DL)
                      : DAG.getNode(ISD::SIGN_EXTEND, DL, SExtVT, Src);
 
  SDValue V;
  if (!Subtarget.has32S() || !Subtarget.hasExtLASX()) {
    if (Src.getSimpleValueType() == MVT::v32i8) {
      SDValue Lo, Hi;
      std::tie(Lo, Hi) = DAG.SplitVector(Src, DL);
      Lo = DAG.getNode(LoongArchISD::VMSKLTZ, DL, GRLenVT, Lo);
      Hi = DAG.getNode(LoongArchISD::VMSKLTZ, DL, GRLenVT, Hi);
      Hi = DAG.getNode(ISD::SHL, DL, GRLenVT, Hi,
                       DAG.getShiftAmountConstant(16, GRLenVT, DL));
      V = DAG.getNode(ISD::OR, DL, GRLenVT, Lo, Hi);
    } else if (UseLASX) {
      return SDValue();
    }
  }
 
  if (!V) {
    Opc = UseLASX ? LoongArchISD::XVMSKLTZ : LoongArchISD::VMSKLTZ;
    V = DAG.getNode(Opc, DL, GRLenVT, Src);
  }
 
  EVT T = EVT::getIntegerVT(*DAG.getContext(), SrcVT.getVectorNumElements());
  V = DAG.getZExtOrTrunc(V, DL, T);
  return DAG.getBitcast(VT, V);
}
 
static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
                                TargetLowering::DAGCombinerInfo &DCI,
                                const LoongArchSubtarget &Subtarget) {
  MVT GRLenVT = Subtarget.getGRLenVT();
  EVT ValTy = N->getValueType(0);
  SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
  ConstantSDNode *CN0, *CN1;
  SDLoc DL(N);
  unsigned ValBits = ValTy.getSizeInBits();
  unsigned MaskIdx0, MaskLen0, MaskIdx1, MaskLen1;
  unsigned Shamt;
  bool SwapAndRetried = false;
 
  // BSTRPICK requires the 32S feature.
  if (!Subtarget.has32S())
    return SDValue();
 
  if (DCI.isBeforeLegalizeOps())
    return SDValue();
 
  if (ValBits != 32 && ValBits != 64)
    return SDValue();
 
Retry:
  // 1st pattern to match BSTRINS:
  //  R = or (and X, mask0), (and (shl Y, lsb), mask1)
  //  where mask1 = (2**size - 1) << lsb, mask0 = ~mask1
  //  =>
  //  R = BSTRINS X, Y, msb, lsb (where msb = lsb + size - 1)
  if (N0.getOpcode() == ISD::AND &&
      (CN0 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) &&
      isShiftedMask_64(~CN0->getSExtValue(), MaskIdx0, MaskLen0) &&
      N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL &&
      (CN1 = dyn_cast<ConstantSDNode>(N1.getOperand(1))) &&
      isShiftedMask_64(CN1->getZExtValue(), MaskIdx1, MaskLen1) &&
      MaskIdx0 == MaskIdx1 && MaskLen0 == MaskLen1 &&
      (CN1 = dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(1))) &&
      (Shamt = CN1->getZExtValue()) == MaskIdx0 &&
      (MaskIdx0 + MaskLen0 <= ValBits)) {
    LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 1\n");
    return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0.getOperand(0),
                       N1.getOperand(0).getOperand(0),
                       DAG.getConstant((MaskIdx0 + MaskLen0 - 1), DL, GRLenVT),
                       DAG.getConstant(MaskIdx0, DL, GRLenVT));
  }
 
  // 2nd pattern to match BSTRINS:
  //  R = or (and X, mask0), (shl (and Y, mask1), lsb)
  //  where mask1 = (2**size - 1), mask0 = ~(mask1 << lsb)
  //  =>
  //  R = BSTRINS X, Y, msb, lsb (where msb = lsb + size - 1)
  if (N0.getOpcode() == ISD::AND &&
      (CN0 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) &&
      isShiftedMask_64(~CN0->getSExtValue(), MaskIdx0, MaskLen0) &&
      N1.getOpcode() == ISD::SHL && N1.getOperand(0).getOpcode() == ISD::AND &&
      (CN1 = dyn_cast<ConstantSDNode>(N1.getOperand(1))) &&
      (Shamt = CN1->getZExtValue()) == MaskIdx0 &&
      (CN1 = dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(1))) &&
      isShiftedMask_64(CN1->getZExtValue(), MaskIdx1, MaskLen1) &&
      MaskLen0 == MaskLen1 && MaskIdx1 == 0 &&
      (MaskIdx0 + MaskLen0 <= ValBits)) {
    LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 2\n");
    return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0.getOperand(0),
                       N1.getOperand(0).getOperand(0),
                       DAG.getConstant((MaskIdx0 + MaskLen0 - 1), DL, GRLenVT),
                       DAG.getConstant(MaskIdx0, DL, GRLenVT));
  }
 
  // 3rd pattern to match BSTRINS:
  //  R = or (and X, mask0), (and Y, mask1)
  //  where ~mask0 = (2**size - 1) << lsb, mask0 & mask1 = 0
  //  =>
  //  R = BSTRINS X, (shr (and Y, mask1), lsb), msb, lsb
  //  where msb = lsb + size - 1
  if (N0.getOpcode() == ISD::AND && N1.getOpcode() == ISD::AND &&
      (CN0 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) &&
      isShiftedMask_64(~CN0->getSExtValue(), MaskIdx0, MaskLen0) &&
      (MaskIdx0 + MaskLen0 <= 64) &&
      (CN1 = dyn_cast<ConstantSDNode>(N1->getOperand(1))) &&
      (CN1->getSExtValue() & CN0->getSExtValue()) == 0) {
    LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 3\n");
    return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0.getOperand(0),
                       DAG.getNode(ISD::SRL, DL, N1->getValueType(0), N1,
                                   DAG.getConstant(MaskIdx0, DL, GRLenVT)),
                       DAG.getConstant(ValBits == 32
                                           ? (MaskIdx0 + (MaskLen0 & 31) - 1)
                                           : (MaskIdx0 + MaskLen0 - 1),
                                       DL, GRLenVT),
                       DAG.getConstant(MaskIdx0, DL, GRLenVT));
  }
 
  // 4th pattern to match BSTRINS:
  //  R = or (and X, mask), (shl Y, shamt)
  //  where mask = (2**shamt - 1)
  //  =>
  //  R = BSTRINS X, Y, ValBits - 1, shamt
  //  where ValBits = 32 or 64
  if (N0.getOpcode() == ISD::AND && N1.getOpcode() == ISD::SHL &&
      (CN0 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) &&
      isShiftedMask_64(CN0->getZExtValue(), MaskIdx0, MaskLen0) &&
      MaskIdx0 == 0 && (CN1 = dyn_cast<ConstantSDNode>(N1.getOperand(1))) &&
      (Shamt = CN1->getZExtValue()) == MaskLen0 &&
      (MaskIdx0 + MaskLen0 <= ValBits)) {
    LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 4\n");
    return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0.getOperand(0),
                       N1.getOperand(0),
                       DAG.getConstant((ValBits - 1), DL, GRLenVT),
                       DAG.getConstant(Shamt, DL, GRLenVT));
  }
 
  // 5th pattern to match BSTRINS:
  //  R = or (and X, mask), const
  //  where ~mask = (2**size - 1) << lsb, mask & const = 0
  //  =>
  //  R = BSTRINS X, (const >> lsb), msb, lsb
  //  where msb = lsb + size - 1
  if (N0.getOpcode() == ISD::AND &&
      (CN0 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) &&
      isShiftedMask_64(~CN0->getSExtValue(), MaskIdx0, MaskLen0) &&
      (CN1 = dyn_cast<ConstantSDNode>(N1)) &&
      (CN1->getSExtValue() & CN0->getSExtValue()) == 0) {
    LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 5\n");
    return DAG.getNode(
        LoongArchISD::BSTRINS, DL, ValTy, N0.getOperand(0),
        DAG.getSignedConstant(CN1->getSExtValue() >> MaskIdx0, DL, ValTy),
        DAG.getConstant(ValBits == 32 ? (MaskIdx0 + (MaskLen0 & 31) - 1)
                                      : (MaskIdx0 + MaskLen0 - 1),
                        DL, GRLenVT),
        DAG.getConstant(MaskIdx0, DL, GRLenVT));
  }
 
  // 6th pattern.
  // a = b | ((c & mask) << shamt), where all positions in b to be overwritten
  // by the incoming bits are known to be zero.
  // =>
  // a = BSTRINS b, c, shamt + MaskLen - 1, shamt
  //
  // Note that the 1st pattern is a special situation of the 6th, i.e. the 6th
  // pattern is more common than the 1st. So we put the 1st before the 6th in
  // order to match as many nodes as possible.
  ConstantSDNode *CNMask, *CNShamt;
  unsigned MaskIdx, MaskLen;
  if (N1.getOpcode() == ISD::SHL && N1.getOperand(0).getOpcode() == ISD::AND &&
      (CNMask = dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(1))) &&
      isShiftedMask_64(CNMask->getZExtValue(), MaskIdx, MaskLen) &&
      MaskIdx == 0 && (CNShamt = dyn_cast<ConstantSDNode>(N1.getOperand(1))) &&
      CNShamt->getZExtValue() + MaskLen <= ValBits) {
    Shamt = CNShamt->getZExtValue();
    APInt ShMask(ValBits, CNMask->getZExtValue() << Shamt);
    if (ShMask.isSubsetOf(DAG.computeKnownBits(N0).Zero)) {
      LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 6\n");
      return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0,
                         N1.getOperand(0).getOperand(0),
                         DAG.getConstant(Shamt + MaskLen - 1, DL, GRLenVT),
                         DAG.getConstant(Shamt, DL, GRLenVT));
    }
  }
 
  // 7th pattern.
  // a = b | ((c << shamt) & shifted_mask), where all positions in b to be
  // overwritten by the incoming bits are known to be zero.
  // =>
  // a = BSTRINS b, c, MaskIdx + MaskLen - 1, MaskIdx
  //
  // Similarly, the 7th pattern is more common than the 2nd. So we put the 2nd
  // before the 7th in order to match as many nodes as possible.
  if (N1.getOpcode() == ISD::AND &&
      (CNMask = dyn_cast<ConstantSDNode>(N1.getOperand(1))) &&
      isShiftedMask_64(CNMask->getZExtValue(), MaskIdx, MaskLen) &&
      N1.getOperand(0).getOpcode() == ISD::SHL &&
      (CNShamt = dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(1))) &&
      CNShamt->getZExtValue() == MaskIdx) {
    APInt ShMask(ValBits, CNMask->getZExtValue());
    if (ShMask.isSubsetOf(DAG.computeKnownBits(N0).Zero)) {
      LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 7\n");
      return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0,
                         N1.getOperand(0).getOperand(0),
                         DAG.getConstant(MaskIdx + MaskLen - 1, DL, GRLenVT),
                         DAG.getConstant(MaskIdx, DL, GRLenVT));
    }
  }
 
  // (or a, b) and (or b, a) are equivalent, so swap the operands and retry.
  if (!SwapAndRetried) {
    std::swap(N0, N1);
    SwapAndRetried = true;
    goto Retry;
  }
 
  SwapAndRetried = false;
Retry2:
  // 8th pattern.
  // a = b | (c & shifted_mask), where all positions in b to be overwritten by
  // the incoming bits are known to be zero.
  // =>
  // a = BSTRINS b, c >> MaskIdx, MaskIdx + MaskLen - 1, MaskIdx
  //
  // Similarly, the 8th pattern is more common than the 4th and 5th patterns. So
  // we put it here in order to match as many nodes as possible or generate less
  // instructions.
  if (N1.getOpcode() == ISD::AND &&
      (CNMask = dyn_cast<ConstantSDNode>(N1.getOperand(1))) &&
      isShiftedMask_64(CNMask->getZExtValue(), MaskIdx, MaskLen)) {
    APInt ShMask(ValBits, CNMask->getZExtValue());
    if (ShMask.isSubsetOf(DAG.computeKnownBits(N0).Zero)) {
      LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 8\n");
      return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0,
                         DAG.getNode(ISD::SRL, DL, N1->getValueType(0),
                                     N1->getOperand(0),
                                     DAG.getConstant(MaskIdx, DL, GRLenVT)),
                         DAG.getConstant(MaskIdx + MaskLen - 1, DL, GRLenVT),
                         DAG.getConstant(MaskIdx, DL, GRLenVT));
    }
  }
  // Swap N0/N1 and retry.
  if (!SwapAndRetried) {
    std::swap(N0, N1);
    SwapAndRetried = true;
    goto Retry2;
  }
 
  return SDValue();
}
 
static bool checkValueWidth(SDValue V, ISD::LoadExtType &ExtType) {
  ExtType = ISD::NON_EXTLOAD;
 
  switch (V.getNode()->getOpcode()) {
  case ISD::LOAD: {
    LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode());
    if ((LoadNode->getMemoryVT() == MVT::i8) ||
        (LoadNode->getMemoryVT() == MVT::i16)) {
      ExtType = LoadNode->getExtensionType();
      return true;
    }
    return false;
  }
  case ISD::AssertSext: {
    VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
    if ((TypeNode->getVT() == MVT::i8) || (TypeNode->getVT() == MVT::i16)) {
      ExtType = ISD::SEXTLOAD;
      return true;
    }
    return false;
  }
  case ISD::AssertZext: {
    VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
    if ((TypeNode->getVT() == MVT::i8) || (TypeNode->getVT() == MVT::i16)) {
      ExtType = ISD::ZEXTLOAD;
      return true;
    }
    return false;
  }
  default:
    return false;
  }
 
  return false;
}
 
// Eliminate redundant truncation and zero-extension nodes.
// * Case 1:
//  +------------+ +------------+ +------------+
//  |   Input1   | |   Input2   | |     CC     |
//  +------------+ +------------+ +------------+
//         |              |              |
//         V              V              +----+
//  +------------+ +------------+             |
//  |  TRUNCATE  | |  TRUNCATE  |             |
//  +------------+ +------------+             |
//         |              |                   |
//         V              V                   |
//  +------------+ +------------+             |
//  |  ZERO_EXT  | |  ZERO_EXT  |             |
//  +------------+ +------------+             |
//         |              |                   |
//         |              +-------------+     |
//         V              V             |     |
//        +----------------+            |     |
//        |      AND       |            |     |
//        +----------------+            |     |
//                |                     |     |
//                +---------------+     |     |
//                                |     |     |
//                                V     V     V
//                               +-------------+
//                               |     CMP     |
//                               +-------------+
// * Case 2:
//  +------------+ +------------+ +-------------+ +------------+ +------------+
//  |   Input1   | |   Input2   | | Constant -1 | | Constant 0 | |     CC     |
//  +------------+ +------------+ +-------------+ +------------+ +------------+
//         |              |             |               |               |
//         V              |             |               |               |
//  +------------+        |             |               |               |
//  |     XOR    |<---------------------+               |               |
//  +------------+        |                             |               |
//         |              |                             |               |
//         V              V             +---------------+               |
//  +------------+ +------------+       |                               |
//  |  TRUNCATE  | |  TRUNCATE  |       |     +-------------------------+
//  +------------+ +------------+       |     |
//         |              |             |     |
//         V              V             |     |
//  +------------+ +------------+       |     |
//  |  ZERO_EXT  | |  ZERO_EXT  |       |     |
//  +------------+ +------------+       |     |
//         |              |             |     |
//         V              V             |     |
//        +----------------+            |     |
//        |      AND       |            |     |
//        +----------------+            |     |
//                |                     |     |
//                +---------------+     |     |
//                                |     |     |
//                                V     V     V
//                               +-------------+
//                               |     CMP     |
//                               +-------------+
static SDValue performSETCCCombine(SDNode *N, SelectionDAG &DAG,
                                   TargetLowering::DAGCombinerInfo &DCI,
                                   const LoongArchSubtarget &Subtarget) {
  ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
 
  SDNode *AndNode = N->getOperand(0).getNode();
  if (AndNode->getOpcode() != ISD::AND)
    return SDValue();
 
  SDValue AndInputValue2 = AndNode->getOperand(1);
  if (AndInputValue2.getOpcode() != ISD::ZERO_EXTEND)
    return SDValue();
 
  SDValue CmpInputValue = N->getOperand(1);
  SDValue AndInputValue1 = AndNode->getOperand(0);
  if (AndInputValue1.getOpcode() == ISD::XOR) {
    if (CC != ISD::SETEQ && CC != ISD::SETNE)
      return SDValue();
    ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndInputValue1.getOperand(1));
    if (!CN || !CN->isAllOnes())
      return SDValue();
    CN = dyn_cast<ConstantSDNode>(CmpInputValue);
    if (!CN || !CN->isZero())
      return SDValue();
    AndInputValue1 = AndInputValue1.getOperand(0);
    if (AndInputValue1.getOpcode() != ISD::ZERO_EXTEND)
      return SDValue();
  } else if (AndInputValue1.getOpcode() == ISD::ZERO_EXTEND) {
    if (AndInputValue2 != CmpInputValue)
      return SDValue();
  } else {
    return SDValue();
  }
 
  SDValue TruncValue1 = AndInputValue1.getNode()->getOperand(0);
  if (TruncValue1.getOpcode() != ISD::TRUNCATE)
    return SDValue();
 
  SDValue TruncValue2 = AndInputValue2.getNode()->getOperand(0);
  if (TruncValue2.getOpcode() != ISD::TRUNCATE)
    return SDValue();
 
  SDValue TruncInputValue1 = TruncValue1.getNode()->getOperand(0);
  SDValue TruncInputValue2 = TruncValue2.getNode()->getOperand(0);
  ISD::LoadExtType ExtType1;
  ISD::LoadExtType ExtType2;
 
  if (!checkValueWidth(TruncInputValue1, ExtType1) ||
      !checkValueWidth(TruncInputValue2, ExtType2))
    return SDValue();
 
  if (TruncInputValue1->getValueType(0) != TruncInputValue2->getValueType(0) ||
      AndNode->getValueType(0) != TruncInputValue1->getValueType(0))
    return SDValue();
 
  if ((ExtType2 != ISD::ZEXTLOAD) &&
      ((ExtType2 != ISD::SEXTLOAD) && (ExtType1 != ISD::SEXTLOAD)))
    return SDValue();
 
  // These truncation and zero-extension nodes are not necessary, remove them.
  SDValue NewAnd = DAG.getNode(ISD::AND, SDLoc(N), AndNode->getValueType(0),
                               TruncInputValue1, TruncInputValue2);
  SDValue NewSetCC =
      DAG.getSetCC(SDLoc(N), N->getValueType(0), NewAnd, TruncInputValue2, CC);
  DAG.ReplaceAllUsesWith(N, NewSetCC.getNode());
  return SDValue(N, 0);
}
 
// Combine (loongarch_bitrev_w (loongarch_revb_2w X)) to loongarch_bitrev_4b.
static SDValue performBITREV_WCombine(SDNode *N, SelectionDAG &DAG,
                                      TargetLowering::DAGCombinerInfo &DCI,
                                      const LoongArchSubtarget &Subtarget) {
  if (DCI.isBeforeLegalizeOps())
    return SDValue();
 
  SDValue Src = N->getOperand(0);
  if (Src.getOpcode() != LoongArchISD::REVB_2W)
    return SDValue();
 
  return DAG.getNode(LoongArchISD::BITREV_4B, SDLoc(N), N->getValueType(0),
                     Src.getOperand(0));
}
 
// Perform common combines for BR_CC and SELECT_CC conditions.
static bool combine_CC(SDValue &LHS, SDValue &RHS, SDValue &CC, const SDLoc &DL,
                       SelectionDAG &DAG, const LoongArchSubtarget &Subtarget) {
  ISD::CondCode CCVal = cast<CondCodeSDNode>(CC)->get();
 
  // As far as arithmetic right shift always saves the sign,
  // shift can be omitted.
  // Fold setlt (sra X, N), 0 -> setlt X, 0 and
  // setge (sra X, N), 0 -> setge X, 0
  if (isNullConstant(RHS) && (CCVal == ISD::SETGE || CCVal == ISD::SETLT) &&
      LHS.getOpcode() == ISD::SRA) {
    LHS = LHS.getOperand(0);
    return true;
  }
 
  if (!ISD::isIntEqualitySetCC(CCVal))
    return false;
 
  // Fold ((setlt X, Y), 0, ne) -> (X, Y, lt)
  // Sometimes the setcc is introduced after br_cc/select_cc has been formed.
  if (LHS.getOpcode() == ISD::SETCC && isNullConstant(RHS) &&
      LHS.getOperand(0).getValueType() == Subtarget.getGRLenVT()) {
    // If we're looking for eq 0 instead of ne 0, we need to invert the
    // condition.
    bool Invert = CCVal == ISD::SETEQ;
    CCVal = cast<CondCodeSDNode>(LHS.getOperand(2))->get();
    if (Invert)
      CCVal = ISD::getSetCCInverse(CCVal, LHS.getValueType());
 
    RHS = LHS.getOperand(1);
    LHS = LHS.getOperand(0);
    translateSetCCForBranch(DL, LHS, RHS, CCVal, DAG);
 
    CC = DAG.getCondCode(CCVal);
    return true;
  }
 
  // Fold ((srl (and X, 1<<C), C), 0, eq/ne) -> ((shl X, GRLen-1-C), 0, ge/lt)
  if (isNullConstant(RHS) && LHS.getOpcode() == ISD::SRL && LHS.hasOneUse() &&
      LHS.getOperand(1).getOpcode() == ISD::Constant) {
    SDValue LHS0 = LHS.getOperand(0);
    if (LHS0.getOpcode() == ISD::AND &&
        LHS0.getOperand(1).getOpcode() == ISD::Constant) {
      uint64_t Mask = LHS0.getConstantOperandVal(1);
      uint64_t ShAmt = LHS.getConstantOperandVal(1);
      if (isPowerOf2_64(Mask) && Log2_64(Mask) == ShAmt) {
        CCVal = CCVal == ISD::SETEQ ? ISD::SETGE : ISD::SETLT;
        CC = DAG.getCondCode(CCVal);
 
        ShAmt = LHS.getValueSizeInBits() - 1 - ShAmt;
        LHS = LHS0.getOperand(0);
        if (ShAmt != 0)
          LHS =
              DAG.getNode(ISD::SHL, DL, LHS.getValueType(), LHS0.getOperand(0),
                          DAG.getConstant(ShAmt, DL, LHS.getValueType()));
        return true;
      }
    }
  }
 
  // (X, 1, setne) -> (X, 0, seteq) if we can prove X is 0/1.
  // This can occur when legalizing some floating point comparisons.
  APInt Mask = APInt::getBitsSetFrom(LHS.getValueSizeInBits(), 1);
  if (isOneConstant(RHS) && DAG.MaskedValueIsZero(LHS, Mask)) {
    CCVal = ISD::getSetCCInverse(CCVal, LHS.getValueType());
    CC = DAG.getCondCode(CCVal);
    RHS = DAG.getConstant(0, DL, LHS.getValueType());
    return true;
  }
 
  // Fold ((shl (extract_vector_elt X, I), GRLen - EleBits)), 0, eq/ne) ->
  //      ((extract_vector_elt X, I), 0, eq/ne)
  if (isNullConstant(RHS) && (CCVal == ISD::SETEQ || CCVal == ISD::SETNE) &&
      LHS.getOpcode() == ISD::SHL && LHS.hasOneUse() &&
      isa<ConstantSDNode>(LHS.getOperand(1))) {
    SDValue Ext = LHS.getOperand(0);
    unsigned Sht = LHS.getConstantOperandVal(1);
    if (Ext.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
      SDValue Vec = Ext.getOperand(0);
      unsigned EleBits = Vec.getScalarValueSizeInBits();
      if ((EleBits + Sht) == Subtarget.getGRLen()) {
        LHS = Ext;
        return true;
      }
    }
  }
 
  return false;
}
 
static SDValue performBR_CCCombine(SDNode *N, SelectionDAG &DAG,
                                   TargetLowering::DAGCombinerInfo &DCI,
                                   const LoongArchSubtarget &Subtarget) {
  SDValue LHS = N->getOperand(1);
  SDValue RHS = N->getOperand(2);
  SDValue CC = N->getOperand(3);
  SDLoc DL(N);
 
  if (combine_CC(LHS, RHS, CC, DL, DAG, Subtarget))
    return DAG.getNode(LoongArchISD::BR_CC, DL, N->getValueType(0),
                       N->getOperand(0), LHS, RHS, CC, N->getOperand(4));
 
  return SDValue();
}
 
static SDValue performSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
                                       TargetLowering::DAGCombinerInfo &DCI,
                                       const LoongArchSubtarget &Subtarget) {
  // Transform
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
  SDValue CC = N->getOperand(2);
  ISD::CondCode CCVal = cast<CondCodeSDNode>(CC)->get();
  SDValue TrueV = N->getOperand(3);
  SDValue FalseV = N->getOperand(4);
  SDLoc DL(N);
  EVT VT = N->getValueType(0);
 
  // If the True and False values are the same, we don't need a select_cc.
  if (TrueV == FalseV)
    return TrueV;
 
  // (select (x < 0), y, z)  -> x >> (GRLEN - 1) & (y - z) + z
  // (select (x >= 0), y, z) -> x >> (GRLEN - 1) & (z - y) + y
  if (isa<ConstantSDNode>(TrueV) && isa<ConstantSDNode>(FalseV) &&
      isNullConstant(RHS) &&
      (CCVal == ISD::CondCode::SETLT || CCVal == ISD::CondCode::SETGE)) {
    if (CCVal == ISD::CondCode::SETGE)
      std::swap(TrueV, FalseV);
 
    int64_t TrueSImm = cast<ConstantSDNode>(TrueV)->getSExtValue();
    int64_t FalseSImm = cast<ConstantSDNode>(FalseV)->getSExtValue();
    // Only handle simm12, if it is not in this range, it can be considered as
    // register.
    if (isInt<12>(TrueSImm) && isInt<12>(FalseSImm) &&
        isInt<12>(TrueSImm - FalseSImm)) {
      SDValue SRA =
          DAG.getNode(ISD::SRA, DL, VT, LHS,
                      DAG.getConstant(Subtarget.getGRLen() - 1, DL, VT));
      SDValue AND =
          DAG.getNode(ISD::AND, DL, VT, SRA,
                      DAG.getSignedConstant(TrueSImm - FalseSImm, DL, VT));
      return DAG.getNode(ISD::ADD, DL, VT, AND, FalseV);
    }
 
    if (CCVal == ISD::CondCode::SETGE)
      std::swap(TrueV, FalseV);
  }
 
  if (combine_CC(LHS, RHS, CC, DL, DAG, Subtarget))
    return DAG.getNode(LoongArchISD::SELECT_CC, DL, N->getValueType(0),
                       {LHS, RHS, CC, TrueV, FalseV});
 
  return SDValue();
}
 
template <unsigned N>
static SDValue legalizeIntrinsicImmArg(SDNode *Node, unsigned ImmOp,
                                       SelectionDAG &DAG,
                                       const LoongArchSubtarget &Subtarget,
                                       bool IsSigned = false) {
  SDLoc DL(Node);
  auto *CImm = cast<ConstantSDNode>(Node->getOperand(ImmOp));
  // Check the ImmArg.
  if ((IsSigned && !isInt<N>(CImm->getSExtValue())) ||
      (!IsSigned && !isUInt<N>(CImm->getZExtValue()))) {
    DAG.getContext()->emitError(Node->getOperationName(0) +
                                ": argument out of range.");
    return DAG.getNode(ISD::UNDEF, DL, Subtarget.getGRLenVT());
  }
  return DAG.getConstant(CImm->getZExtValue(), DL, Subtarget.getGRLenVT());
}
 
template <unsigned N>
static SDValue lowerVectorSplatImm(SDNode *Node, unsigned ImmOp,
                                   SelectionDAG &DAG, bool IsSigned = false) {
  SDLoc DL(Node);
  EVT ResTy = Node->getValueType(0);
  auto *CImm = cast<ConstantSDNode>(Node->getOperand(ImmOp));
 
  // Check the ImmArg.
  if ((IsSigned && !isInt<N>(CImm->getSExtValue())) ||
      (!IsSigned && !isUInt<N>(CImm->getZExtValue()))) {
    DAG.getContext()->emitError(Node->getOperationName(0) +
                                ": argument out of range.");
    return DAG.getNode(ISD::UNDEF, DL, ResTy);
  }
  return DAG.getConstant(
      APInt(ResTy.getScalarType().getSizeInBits(),
            IsSigned ? CImm->getSExtValue() : CImm->getZExtValue(), IsSigned),
      DL, ResTy);
}
 
static SDValue truncateVecElts(SDNode *Node, SelectionDAG &DAG) {
  SDLoc DL(Node);
  EVT ResTy = Node->getValueType(0);
  SDValue Vec = Node->getOperand(2);
  SDValue Mask = DAG.getConstant(Vec.getScalarValueSizeInBits() - 1, DL, ResTy);
  return DAG.getNode(ISD::AND, DL, ResTy, Vec, Mask);
}
 
static SDValue lowerVectorBitClear(SDNode *Node, SelectionDAG &DAG) {
  SDLoc DL(Node);
  EVT ResTy = Node->getValueType(0);
  SDValue One = DAG.getConstant(1, DL, ResTy);
  SDValue Bit =
      DAG.getNode(ISD::SHL, DL, ResTy, One, truncateVecElts(Node, DAG));
 
  return DAG.getNode(ISD::AND, DL, ResTy, Node->getOperand(1),
                     DAG.getNOT(DL, Bit, ResTy));
}
 
template <unsigned N>
static SDValue lowerVectorBitClearImm(SDNode *Node, SelectionDAG &DAG) {
  SDLoc DL(Node);
  EVT ResTy = Node->getValueType(0);
  auto *CImm = cast<ConstantSDNode>(Node->getOperand(2));
  // Check the unsigned ImmArg.
  if (!isUInt<N>(CImm->getZExtValue())) {
    DAG.getContext()->emitError(Node->getOperationName(0) +
                                ": argument out of range.");
    return DAG.getNode(ISD::UNDEF, DL, ResTy);
  }
 
  APInt BitImm = APInt(ResTy.getScalarSizeInBits(), 1) << CImm->getAPIntValue();
  SDValue Mask = DAG.getConstant(~BitImm, DL, ResTy);
 
  return DAG.getNode(ISD::AND, DL, ResTy, Node->getOperand(1), Mask);
}
 
template <unsigned N>
static SDValue lowerVectorBitSetImm(SDNode *Node, SelectionDAG &DAG) {
  SDLoc DL(Node);
  EVT ResTy = Node->getValueType(0);
  auto *CImm = cast<ConstantSDNode>(Node->getOperand(2));
  // Check the unsigned ImmArg.
  if (!isUInt<N>(CImm->getZExtValue())) {
    DAG.getContext()->emitError(Node->getOperationName(0) +
                                ": argument out of range.");
    return DAG.getNode(ISD::UNDEF, DL, ResTy);
  }
 
  APInt Imm = APInt(ResTy.getScalarSizeInBits(), 1) << CImm->getAPIntValue();
  SDValue BitImm = DAG.getConstant(Imm, DL, ResTy);
  return DAG.getNode(ISD::OR, DL, ResTy, Node->getOperand(1), BitImm);
}
 
template <unsigned N>
static SDValue lowerVectorBitRevImm(SDNode *Node, SelectionDAG &DAG) {
  SDLoc DL(Node);
  EVT ResTy = Node->getValueType(0);
  auto *CImm = cast<ConstantSDNode>(Node->getOperand(2));
  // Check the unsigned ImmArg.
  if (!isUInt<N>(CImm->getZExtValue())) {
    DAG.getContext()->emitError(Node->getOperationName(0) +
                                ": argument out of range.");
    return DAG.getNode(ISD::UNDEF, DL, ResTy);
  }
 
  APInt Imm = APInt(ResTy.getScalarSizeInBits(), 1) << CImm->getAPIntValue();
  SDValue BitImm = DAG.getConstant(Imm, DL, ResTy);
  return DAG.getNode(ISD::XOR, DL, ResTy, Node->getOperand(1), BitImm);
}
 
template <unsigned W>
static SDValue lowerVectorPickVE2GR(SDNode *N, SelectionDAG &DAG,
                                    unsigned ResOp) {
  unsigned Imm = N->getConstantOperandVal(2);
  if (!isUInt<W>(Imm)) {
    const StringRef ErrorMsg = "argument out of range";
    DAG.getContext()->emitError(N->getOperationName(0) + ": " + ErrorMsg + ".");
    return DAG.getUNDEF(N->getValueType(0));
  }
  SDLoc DL(N);
  SDValue Vec = N->getOperand(1);
  SDValue Idx = DAG.getConstant(Imm, DL, MVT::i32);
  SDValue EltVT = DAG.getValueType(Vec.getValueType().getVectorElementType());
  return DAG.getNode(ResOp, DL, N->getValueType(0), Vec, Idx, EltVT);
}
 
static SDValue
performINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
                                 TargetLowering::DAGCombinerInfo &DCI,
                                 const LoongArchSubtarget &Subtarget) {
  SDLoc DL(N);
  switch (N->getConstantOperandVal(0)) {
  default:
    break;
  case Intrinsic::loongarch_lsx_vadd_b:
  case Intrinsic::loongarch_lsx_vadd_h:
  case Intrinsic::loongarch_lsx_vadd_w:
  case Intrinsic::loongarch_lsx_vadd_d:
  case Intrinsic::loongarch_lasx_xvadd_b:
  case Intrinsic::loongarch_lasx_xvadd_h:
  case Intrinsic::loongarch_lasx_xvadd_w:
  case Intrinsic::loongarch_lasx_xvadd_d:
    return DAG.getNode(ISD::ADD, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vaddi_bu:
  case Intrinsic::loongarch_lsx_vaddi_hu:
  case Intrinsic::loongarch_lsx_vaddi_wu:
  case Intrinsic::loongarch_lsx_vaddi_du:
  case Intrinsic::loongarch_lasx_xvaddi_bu:
  case Intrinsic::loongarch_lasx_xvaddi_hu:
  case Intrinsic::loongarch_lasx_xvaddi_wu:
  case Intrinsic::loongarch_lasx_xvaddi_du:
    return DAG.getNode(ISD::ADD, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<5>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vsub_b:
  case Intrinsic::loongarch_lsx_vsub_h:
  case Intrinsic::loongarch_lsx_vsub_w:
  case Intrinsic::loongarch_lsx_vsub_d:
  case Intrinsic::loongarch_lasx_xvsub_b:
  case Intrinsic::loongarch_lasx_xvsub_h:
  case Intrinsic::loongarch_lasx_xvsub_w:
  case Intrinsic::loongarch_lasx_xvsub_d:
    return DAG.getNode(ISD::SUB, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vsubi_bu:
  case Intrinsic::loongarch_lsx_vsubi_hu:
  case Intrinsic::loongarch_lsx_vsubi_wu:
  case Intrinsic::loongarch_lsx_vsubi_du:
  case Intrinsic::loongarch_lasx_xvsubi_bu:
  case Intrinsic::loongarch_lasx_xvsubi_hu:
  case Intrinsic::loongarch_lasx_xvsubi_wu:
  case Intrinsic::loongarch_lasx_xvsubi_du:
    return DAG.getNode(ISD::SUB, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<5>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vneg_b:
  case Intrinsic::loongarch_lsx_vneg_h:
  case Intrinsic::loongarch_lsx_vneg_w:
  case Intrinsic::loongarch_lsx_vneg_d:
  case Intrinsic::loongarch_lasx_xvneg_b:
  case Intrinsic::loongarch_lasx_xvneg_h:
  case Intrinsic::loongarch_lasx_xvneg_w:
  case Intrinsic::loongarch_lasx_xvneg_d:
    return DAG.getNode(
        ISD::SUB, DL, N->getValueType(0),
        DAG.getConstant(
            APInt(N->getValueType(0).getScalarType().getSizeInBits(), 0,
                  /*isSigned=*/true),
            SDLoc(N), N->getValueType(0)),
        N->getOperand(1));
  case Intrinsic::loongarch_lsx_vmax_b:
  case Intrinsic::loongarch_lsx_vmax_h:
  case Intrinsic::loongarch_lsx_vmax_w:
  case Intrinsic::loongarch_lsx_vmax_d:
  case Intrinsic::loongarch_lasx_xvmax_b:
  case Intrinsic::loongarch_lasx_xvmax_h:
  case Intrinsic::loongarch_lasx_xvmax_w:
  case Intrinsic::loongarch_lasx_xvmax_d:
    return DAG.getNode(ISD::SMAX, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vmax_bu:
  case Intrinsic::loongarch_lsx_vmax_hu:
  case Intrinsic::loongarch_lsx_vmax_wu:
  case Intrinsic::loongarch_lsx_vmax_du:
  case Intrinsic::loongarch_lasx_xvmax_bu:
  case Intrinsic::loongarch_lasx_xvmax_hu:
  case Intrinsic::loongarch_lasx_xvmax_wu:
  case Intrinsic::loongarch_lasx_xvmax_du:
    return DAG.getNode(ISD::UMAX, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vmaxi_b:
  case Intrinsic::loongarch_lsx_vmaxi_h:
  case Intrinsic::loongarch_lsx_vmaxi_w:
  case Intrinsic::loongarch_lsx_vmaxi_d:
  case Intrinsic::loongarch_lasx_xvmaxi_b:
  case Intrinsic::loongarch_lasx_xvmaxi_h:
  case Intrinsic::loongarch_lasx_xvmaxi_w:
  case Intrinsic::loongarch_lasx_xvmaxi_d:
    return DAG.getNode(ISD::SMAX, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<5>(N, 2, DAG, /*IsSigned=*/true));
  case Intrinsic::loongarch_lsx_vmaxi_bu:
  case Intrinsic::loongarch_lsx_vmaxi_hu:
  case Intrinsic::loongarch_lsx_vmaxi_wu:
  case Intrinsic::loongarch_lsx_vmaxi_du:
  case Intrinsic::loongarch_lasx_xvmaxi_bu:
  case Intrinsic::loongarch_lasx_xvmaxi_hu:
  case Intrinsic::loongarch_lasx_xvmaxi_wu:
  case Intrinsic::loongarch_lasx_xvmaxi_du:
    return DAG.getNode(ISD::UMAX, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<5>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vmin_b:
  case Intrinsic::loongarch_lsx_vmin_h:
  case Intrinsic::loongarch_lsx_vmin_w:
  case Intrinsic::loongarch_lsx_vmin_d:
  case Intrinsic::loongarch_lasx_xvmin_b:
  case Intrinsic::loongarch_lasx_xvmin_h:
  case Intrinsic::loongarch_lasx_xvmin_w:
  case Intrinsic::loongarch_lasx_xvmin_d:
    return DAG.getNode(ISD::SMIN, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vmin_bu:
  case Intrinsic::loongarch_lsx_vmin_hu:
  case Intrinsic::loongarch_lsx_vmin_wu:
  case Intrinsic::loongarch_lsx_vmin_du:
  case Intrinsic::loongarch_lasx_xvmin_bu:
  case Intrinsic::loongarch_lasx_xvmin_hu:
  case Intrinsic::loongarch_lasx_xvmin_wu:
  case Intrinsic::loongarch_lasx_xvmin_du:
    return DAG.getNode(ISD::UMIN, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vmini_b:
  case Intrinsic::loongarch_lsx_vmini_h:
  case Intrinsic::loongarch_lsx_vmini_w:
  case Intrinsic::loongarch_lsx_vmini_d:
  case Intrinsic::loongarch_lasx_xvmini_b:
  case Intrinsic::loongarch_lasx_xvmini_h:
  case Intrinsic::loongarch_lasx_xvmini_w:
  case Intrinsic::loongarch_lasx_xvmini_d:
    return DAG.getNode(ISD::SMIN, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<5>(N, 2, DAG, /*IsSigned=*/true));
  case Intrinsic::loongarch_lsx_vmini_bu:
  case Intrinsic::loongarch_lsx_vmini_hu:
  case Intrinsic::loongarch_lsx_vmini_wu:
  case Intrinsic::loongarch_lsx_vmini_du:
  case Intrinsic::loongarch_lasx_xvmini_bu:
  case Intrinsic::loongarch_lasx_xvmini_hu:
  case Intrinsic::loongarch_lasx_xvmini_wu:
  case Intrinsic::loongarch_lasx_xvmini_du:
    return DAG.getNode(ISD::UMIN, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<5>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vmul_b:
  case Intrinsic::loongarch_lsx_vmul_h:
  case Intrinsic::loongarch_lsx_vmul_w:
  case Intrinsic::loongarch_lsx_vmul_d:
  case Intrinsic::loongarch_lasx_xvmul_b:
  case Intrinsic::loongarch_lasx_xvmul_h:
  case Intrinsic::loongarch_lasx_xvmul_w:
  case Intrinsic::loongarch_lasx_xvmul_d:
    return DAG.getNode(ISD::MUL, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vmadd_b:
  case Intrinsic::loongarch_lsx_vmadd_h:
  case Intrinsic::loongarch_lsx_vmadd_w:
  case Intrinsic::loongarch_lsx_vmadd_d:
  case Intrinsic::loongarch_lasx_xvmadd_b:
  case Intrinsic::loongarch_lasx_xvmadd_h:
  case Intrinsic::loongarch_lasx_xvmadd_w:
  case Intrinsic::loongarch_lasx_xvmadd_d: {
    EVT ResTy = N->getValueType(0);
    return DAG.getNode(ISD::ADD, SDLoc(N), ResTy, N->getOperand(1),
                       DAG.getNode(ISD::MUL, SDLoc(N), ResTy, N->getOperand(2),
                                   N->getOperand(3)));
  }
  case Intrinsic::loongarch_lsx_vmsub_b:
  case Intrinsic::loongarch_lsx_vmsub_h:
  case Intrinsic::loongarch_lsx_vmsub_w:
  case Intrinsic::loongarch_lsx_vmsub_d:
  case Intrinsic::loongarch_lasx_xvmsub_b:
  case Intrinsic::loongarch_lasx_xvmsub_h:
  case Intrinsic::loongarch_lasx_xvmsub_w:
  case Intrinsic::loongarch_lasx_xvmsub_d: {
    EVT ResTy = N->getValueType(0);
    return DAG.getNode(ISD::SUB, SDLoc(N), ResTy, N->getOperand(1),
                       DAG.getNode(ISD::MUL, SDLoc(N), ResTy, N->getOperand(2),
                                   N->getOperand(3)));
  }
  case Intrinsic::loongarch_lsx_vdiv_b:
  case Intrinsic::loongarch_lsx_vdiv_h:
  case Intrinsic::loongarch_lsx_vdiv_w:
  case Intrinsic::loongarch_lsx_vdiv_d:
  case Intrinsic::loongarch_lasx_xvdiv_b:
  case Intrinsic::loongarch_lasx_xvdiv_h:
  case Intrinsic::loongarch_lasx_xvdiv_w:
  case Intrinsic::loongarch_lasx_xvdiv_d:
    return DAG.getNode(ISD::SDIV, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vdiv_bu:
  case Intrinsic::loongarch_lsx_vdiv_hu:
  case Intrinsic::loongarch_lsx_vdiv_wu:
  case Intrinsic::loongarch_lsx_vdiv_du:
  case Intrinsic::loongarch_lasx_xvdiv_bu:
  case Intrinsic::loongarch_lasx_xvdiv_hu:
  case Intrinsic::loongarch_lasx_xvdiv_wu:
  case Intrinsic::loongarch_lasx_xvdiv_du:
    return DAG.getNode(ISD::UDIV, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vmod_b:
  case Intrinsic::loongarch_lsx_vmod_h:
  case Intrinsic::loongarch_lsx_vmod_w:
  case Intrinsic::loongarch_lsx_vmod_d:
  case Intrinsic::loongarch_lasx_xvmod_b:
  case Intrinsic::loongarch_lasx_xvmod_h:
  case Intrinsic::loongarch_lasx_xvmod_w:
  case Intrinsic::loongarch_lasx_xvmod_d:
    return DAG.getNode(ISD::SREM, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vmod_bu:
  case Intrinsic::loongarch_lsx_vmod_hu:
  case Intrinsic::loongarch_lsx_vmod_wu:
  case Intrinsic::loongarch_lsx_vmod_du:
  case Intrinsic::loongarch_lasx_xvmod_bu:
  case Intrinsic::loongarch_lasx_xvmod_hu:
  case Intrinsic::loongarch_lasx_xvmod_wu:
  case Intrinsic::loongarch_lasx_xvmod_du:
    return DAG.getNode(ISD::UREM, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vand_v:
  case Intrinsic::loongarch_lasx_xvand_v:
    return DAG.getNode(ISD::AND, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vor_v:
  case Intrinsic::loongarch_lasx_xvor_v:
    return DAG.getNode(ISD::OR, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vxor_v:
  case Intrinsic::loongarch_lasx_xvxor_v:
    return DAG.getNode(ISD::XOR, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vnor_v:
  case Intrinsic::loongarch_lasx_xvnor_v: {
    SDValue Res = DAG.getNode(ISD::OR, DL, N->getValueType(0), N->getOperand(1),
                              N->getOperand(2));
    return DAG.getNOT(DL, Res, Res->getValueType(0));
  }
  case Intrinsic::loongarch_lsx_vandi_b:
  case Intrinsic::loongarch_lasx_xvandi_b:
    return DAG.getNode(ISD::AND, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<8>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vori_b:
  case Intrinsic::loongarch_lasx_xvori_b:
    return DAG.getNode(ISD::OR, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<8>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vxori_b:
  case Intrinsic::loongarch_lasx_xvxori_b:
    return DAG.getNode(ISD::XOR, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<8>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vsll_b:
  case Intrinsic::loongarch_lsx_vsll_h:
  case Intrinsic::loongarch_lsx_vsll_w:
  case Intrinsic::loongarch_lsx_vsll_d:
  case Intrinsic::loongarch_lasx_xvsll_b:
  case Intrinsic::loongarch_lasx_xvsll_h:
  case Intrinsic::loongarch_lasx_xvsll_w:
  case Intrinsic::loongarch_lasx_xvsll_d:
    return DAG.getNode(ISD::SHL, DL, N->getValueType(0), N->getOperand(1),
                       truncateVecElts(N, DAG));
  case Intrinsic::loongarch_lsx_vslli_b:
  case Intrinsic::loongarch_lasx_xvslli_b:
    return DAG.getNode(ISD::SHL, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<3>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vslli_h:
  case Intrinsic::loongarch_lasx_xvslli_h:
    return DAG.getNode(ISD::SHL, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<4>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vslli_w:
  case Intrinsic::loongarch_lasx_xvslli_w:
    return DAG.getNode(ISD::SHL, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<5>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vslli_d:
  case Intrinsic::loongarch_lasx_xvslli_d:
    return DAG.getNode(ISD::SHL, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<6>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vsrl_b:
  case Intrinsic::loongarch_lsx_vsrl_h:
  case Intrinsic::loongarch_lsx_vsrl_w:
  case Intrinsic::loongarch_lsx_vsrl_d:
  case Intrinsic::loongarch_lasx_xvsrl_b:
  case Intrinsic::loongarch_lasx_xvsrl_h:
  case Intrinsic::loongarch_lasx_xvsrl_w:
  case Intrinsic::loongarch_lasx_xvsrl_d:
    return DAG.getNode(ISD::SRL, DL, N->getValueType(0), N->getOperand(1),
                       truncateVecElts(N, DAG));
  case Intrinsic::loongarch_lsx_vsrli_b:
  case Intrinsic::loongarch_lasx_xvsrli_b:
    return DAG.getNode(ISD::SRL, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<3>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vsrli_h:
  case Intrinsic::loongarch_lasx_xvsrli_h:
    return DAG.getNode(ISD::SRL, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<4>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vsrli_w:
  case Intrinsic::loongarch_lasx_xvsrli_w:
    return DAG.getNode(ISD::SRL, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<5>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vsrli_d:
  case Intrinsic::loongarch_lasx_xvsrli_d:
    return DAG.getNode(ISD::SRL, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<6>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vsra_b:
  case Intrinsic::loongarch_lsx_vsra_h:
  case Intrinsic::loongarch_lsx_vsra_w:
  case Intrinsic::loongarch_lsx_vsra_d:
  case Intrinsic::loongarch_lasx_xvsra_b:
  case Intrinsic::loongarch_lasx_xvsra_h:
  case Intrinsic::loongarch_lasx_xvsra_w:
  case Intrinsic::loongarch_lasx_xvsra_d:
    return DAG.getNode(ISD::SRA, DL, N->getValueType(0), N->getOperand(1),
                       truncateVecElts(N, DAG));
  case Intrinsic::loongarch_lsx_vsrai_b:
  case Intrinsic::loongarch_lasx_xvsrai_b:
    return DAG.getNode(ISD::SRA, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<3>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vsrai_h:
  case Intrinsic::loongarch_lasx_xvsrai_h:
    return DAG.getNode(ISD::SRA, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<4>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vsrai_w:
  case Intrinsic::loongarch_lasx_xvsrai_w:
    return DAG.getNode(ISD::SRA, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<5>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vsrai_d:
  case Intrinsic::loongarch_lasx_xvsrai_d:
    return DAG.getNode(ISD::SRA, DL, N->getValueType(0), N->getOperand(1),
                       lowerVectorSplatImm<6>(N, 2, DAG));
  case Intrinsic::loongarch_lsx_vclz_b:
  case Intrinsic::loongarch_lsx_vclz_h:
  case Intrinsic::loongarch_lsx_vclz_w:
  case Intrinsic::loongarch_lsx_vclz_d:
  case Intrinsic::loongarch_lasx_xvclz_b:
  case Intrinsic::loongarch_lasx_xvclz_h:
  case Intrinsic::loongarch_lasx_xvclz_w:
  case Intrinsic::loongarch_lasx_xvclz_d:
    return DAG.getNode(ISD::CTLZ, DL, N->getValueType(0), N->getOperand(1));
  case Intrinsic::loongarch_lsx_vpcnt_b:
  case Intrinsic::loongarch_lsx_vpcnt_h:
  case Intrinsic::loongarch_lsx_vpcnt_w:
  case Intrinsic::loongarch_lsx_vpcnt_d:
  case Intrinsic::loongarch_lasx_xvpcnt_b:
  case Intrinsic::loongarch_lasx_xvpcnt_h:
  case Intrinsic::loongarch_lasx_xvpcnt_w:
  case Intrinsic::loongarch_lasx_xvpcnt_d:
    return DAG.getNode(ISD::CTPOP, DL, N->getValueType(0), N->getOperand(1));
  case Intrinsic::loongarch_lsx_vbitclr_b:
  case Intrinsic::loongarch_lsx_vbitclr_h:
  case Intrinsic::loongarch_lsx_vbitclr_w:
  case Intrinsic::loongarch_lsx_vbitclr_d:
  case Intrinsic::loongarch_lasx_xvbitclr_b:
  case Intrinsic::loongarch_lasx_xvbitclr_h:
  case Intrinsic::loongarch_lasx_xvbitclr_w:
  case Intrinsic::loongarch_lasx_xvbitclr_d:
    return lowerVectorBitClear(N, DAG);
  case Intrinsic::loongarch_lsx_vbitclri_b:
  case Intrinsic::loongarch_lasx_xvbitclri_b:
    return lowerVectorBitClearImm<3>(N, DAG);
  case Intrinsic::loongarch_lsx_vbitclri_h:
  case Intrinsic::loongarch_lasx_xvbitclri_h:
    return lowerVectorBitClearImm<4>(N, DAG);
  case Intrinsic::loongarch_lsx_vbitclri_w:
  case Intrinsic::loongarch_lasx_xvbitclri_w:
    return lowerVectorBitClearImm<5>(N, DAG);
  case Intrinsic::loongarch_lsx_vbitclri_d:
  case Intrinsic::loongarch_lasx_xvbitclri_d:
    return lowerVectorBitClearImm<6>(N, DAG);
  case Intrinsic::loongarch_lsx_vbitset_b:
  case Intrinsic::loongarch_lsx_vbitset_h:
  case Intrinsic::loongarch_lsx_vbitset_w:
  case Intrinsic::loongarch_lsx_vbitset_d:
  case Intrinsic::loongarch_lasx_xvbitset_b:
  case Intrinsic::loongarch_lasx_xvbitset_h:
  case Intrinsic::loongarch_lasx_xvbitset_w:
  case Intrinsic::loongarch_lasx_xvbitset_d: {
    EVT VecTy = N->getValueType(0);
    SDValue One = DAG.getConstant(1, DL, VecTy);
    return DAG.getNode(
        ISD::OR, DL, VecTy, N->getOperand(1),
        DAG.getNode(ISD::SHL, DL, VecTy, One, truncateVecElts(N, DAG)));
  }
  case Intrinsic::loongarch_lsx_vbitseti_b:
  case Intrinsic::loongarch_lasx_xvbitseti_b:
    return lowerVectorBitSetImm<3>(N, DAG);
  case Intrinsic::loongarch_lsx_vbitseti_h:
  case Intrinsic::loongarch_lasx_xvbitseti_h:
    return lowerVectorBitSetImm<4>(N, DAG);
  case Intrinsic::loongarch_lsx_vbitseti_w:
  case Intrinsic::loongarch_lasx_xvbitseti_w:
    return lowerVectorBitSetImm<5>(N, DAG);
  case Intrinsic::loongarch_lsx_vbitseti_d:
  case Intrinsic::loongarch_lasx_xvbitseti_d:
    return lowerVectorBitSetImm<6>(N, DAG);
  case Intrinsic::loongarch_lsx_vbitrev_b:
  case Intrinsic::loongarch_lsx_vbitrev_h:
  case Intrinsic::loongarch_lsx_vbitrev_w:
  case Intrinsic::loongarch_lsx_vbitrev_d:
  case Intrinsic::loongarch_lasx_xvbitrev_b:
  case Intrinsic::loongarch_lasx_xvbitrev_h:
  case Intrinsic::loongarch_lasx_xvbitrev_w:
  case Intrinsic::loongarch_lasx_xvbitrev_d: {
    EVT VecTy = N->getValueType(0);
    SDValue One = DAG.getConstant(1, DL, VecTy);
    return DAG.getNode(
        ISD::XOR, DL, VecTy, N->getOperand(1),
        DAG.getNode(ISD::SHL, DL, VecTy, One, truncateVecElts(N, DAG)));
  }
  case Intrinsic::loongarch_lsx_vbitrevi_b:
  case Intrinsic::loongarch_lasx_xvbitrevi_b:
    return lowerVectorBitRevImm<3>(N, DAG);
  case Intrinsic::loongarch_lsx_vbitrevi_h:
  case Intrinsic::loongarch_lasx_xvbitrevi_h:
    return lowerVectorBitRevImm<4>(N, DAG);
  case Intrinsic::loongarch_lsx_vbitrevi_w:
  case Intrinsic::loongarch_lasx_xvbitrevi_w:
    return lowerVectorBitRevImm<5>(N, DAG);
  case Intrinsic::loongarch_lsx_vbitrevi_d:
  case Intrinsic::loongarch_lasx_xvbitrevi_d:
    return lowerVectorBitRevImm<6>(N, DAG);
  case Intrinsic::loongarch_lsx_vfadd_s:
  case Intrinsic::loongarch_lsx_vfadd_d:
  case Intrinsic::loongarch_lasx_xvfadd_s:
  case Intrinsic::loongarch_lasx_xvfadd_d:
    return DAG.getNode(ISD::FADD, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vfsub_s:
  case Intrinsic::loongarch_lsx_vfsub_d:
  case Intrinsic::loongarch_lasx_xvfsub_s:
  case Intrinsic::loongarch_lasx_xvfsub_d:
    return DAG.getNode(ISD::FSUB, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vfmul_s:
  case Intrinsic::loongarch_lsx_vfmul_d:
  case Intrinsic::loongarch_lasx_xvfmul_s:
  case Intrinsic::loongarch_lasx_xvfmul_d:
    return DAG.getNode(ISD::FMUL, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vfdiv_s:
  case Intrinsic::loongarch_lsx_vfdiv_d:
  case Intrinsic::loongarch_lasx_xvfdiv_s:
  case Intrinsic::loongarch_lasx_xvfdiv_d:
    return DAG.getNode(ISD::FDIV, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2));
  case Intrinsic::loongarch_lsx_vfmadd_s:
  case Intrinsic::loongarch_lsx_vfmadd_d:
  case Intrinsic::loongarch_lasx_xvfmadd_s:
  case Intrinsic::loongarch_lasx_xvfmadd_d:
    return DAG.getNode(ISD::FMA, DL, N->getValueType(0), N->getOperand(1),
                       N->getOperand(2), N->getOperand(3));
  case Intrinsic::loongarch_lsx_vinsgr2vr_b:
    return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2),
                       legalizeIntrinsicImmArg<4>(N, 3, DAG, Subtarget));
  case Intrinsic::loongarch_lsx_vinsgr2vr_h:
  case Intrinsic::loongarch_lasx_xvinsgr2vr_w:
    return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2),
                       legalizeIntrinsicImmArg<3>(N, 3, DAG, Subtarget));
  case Intrinsic::loongarch_lsx_vinsgr2vr_w:
  case Intrinsic::loongarch_lasx_xvinsgr2vr_d:
    return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2),
                       legalizeIntrinsicImmArg<2>(N, 3, DAG, Subtarget));
  case Intrinsic::loongarch_lsx_vinsgr2vr_d:
    return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2),
                       legalizeIntrinsicImmArg<1>(N, 3, DAG, Subtarget));
  case Intrinsic::loongarch_lsx_vreplgr2vr_b:
  case Intrinsic::loongarch_lsx_vreplgr2vr_h:
  case Intrinsic::loongarch_lsx_vreplgr2vr_w:
  case Intrinsic::loongarch_lsx_vreplgr2vr_d:
  case Intrinsic::loongarch_lasx_xvreplgr2vr_b:
  case Intrinsic::loongarch_lasx_xvreplgr2vr_h:
  case Intrinsic::loongarch_lasx_xvreplgr2vr_w:
  case Intrinsic::loongarch_lasx_xvreplgr2vr_d:
    return DAG.getNode(LoongArchISD::VREPLGR2VR, DL, N->getValueType(0),
                       DAG.getNode(ISD::ANY_EXTEND, DL, Subtarget.getGRLenVT(),
                                   N->getOperand(1)));
  case Intrinsic::loongarch_lsx_vreplve_b:
  case Intrinsic::loongarch_lsx_vreplve_h:
  case Intrinsic::loongarch_lsx_vreplve_w:
  case Intrinsic::loongarch_lsx_vreplve_d:
  case Intrinsic::loongarch_lasx_xvreplve_b:
  case Intrinsic::loongarch_lasx_xvreplve_h:
  case Intrinsic::loongarch_lasx_xvreplve_w:
  case Intrinsic::loongarch_lasx_xvreplve_d:
    return DAG.getNode(LoongArchISD::VREPLVE, DL, N->getValueType(0),
                       N->getOperand(1),
                       DAG.getNode(ISD::ANY_EXTEND, DL, Subtarget.getGRLenVT(),
                                   N->getOperand(2)));
  case Intrinsic::loongarch_lsx_vpickve2gr_b:
    if (!Subtarget.is64Bit())
      return lowerVectorPickVE2GR<4>(N, DAG, LoongArchISD::VPICK_SEXT_ELT);
    break;
  case Intrinsic::loongarch_lsx_vpickve2gr_h:
  case Intrinsic::loongarch_lasx_xvpickve2gr_w:
    if (!Subtarget.is64Bit())
      return lowerVectorPickVE2GR<3>(N, DAG, LoongArchISD::VPICK_SEXT_ELT);
    break;
  case Intrinsic::loongarch_lsx_vpickve2gr_w:
    if (!Subtarget.is64Bit())
      return lowerVectorPickVE2GR<2>(N, DAG, LoongArchISD::VPICK_SEXT_ELT);
    break;
  case Intrinsic::loongarch_lsx_vpickve2gr_bu:
    if (!Subtarget.is64Bit())
      return lowerVectorPickVE2GR<4>(N, DAG, LoongArchISD::VPICK_ZEXT_ELT);
    break;
  case Intrinsic::loongarch_lsx_vpickve2gr_hu:
  case Intrinsic::loongarch_lasx_xvpickve2gr_wu:
    if (!Subtarget.is64Bit())
      return lowerVectorPickVE2GR<3>(N, DAG, LoongArchISD::VPICK_ZEXT_ELT);
    break;
  case Intrinsic::loongarch_lsx_vpickve2gr_wu:
    if (!Subtarget.is64Bit())
      return lowerVectorPickVE2GR<2>(N, DAG, LoongArchISD::VPICK_ZEXT_ELT);
    break;
  case Intrinsic::loongarch_lsx_bz_b:
  case Intrinsic::loongarch_lsx_bz_h:
  case Intrinsic::loongarch_lsx_bz_w:
  case Intrinsic::loongarch_lsx_bz_d:
  case Intrinsic::loongarch_lasx_xbz_b:
  case Intrinsic::loongarch_lasx_xbz_h:
  case Intrinsic::loongarch_lasx_xbz_w:
  case Intrinsic::loongarch_lasx_xbz_d:
    if (!Subtarget.is64Bit())
      return DAG.getNode(LoongArchISD::VALL_ZERO, DL, N->getValueType(0),
                         N->getOperand(1));
    break;
  case Intrinsic::loongarch_lsx_bz_v:
  case Intrinsic::loongarch_lasx_xbz_v:
    if (!Subtarget.is64Bit())
      return DAG.getNode(LoongArchISD::VANY_ZERO, DL, N->getValueType(0),
                         N->getOperand(1));
    break;
  case Intrinsic::loongarch_lsx_bnz_b:
  case Intrinsic::loongarch_lsx_bnz_h:
  case Intrinsic::loongarch_lsx_bnz_w:
  case Intrinsic::loongarch_lsx_bnz_d:
  case Intrinsic::loongarch_lasx_xbnz_b:
  case Intrinsic::loongarch_lasx_xbnz_h:
  case Intrinsic::loongarch_lasx_xbnz_w:
  case Intrinsic::loongarch_lasx_xbnz_d:
    if (!Subtarget.is64Bit())
      return DAG.getNode(LoongArchISD::VALL_NONZERO, DL, N->getValueType(0),
                         N->getOperand(1));
    break;
  case Intrinsic::loongarch_lsx_bnz_v:
  case Intrinsic::loongarch_lasx_xbnz_v:
    if (!Subtarget.is64Bit())
      return DAG.getNode(LoongArchISD::VANY_NONZERO, DL, N->getValueType(0),
                         N->getOperand(1));
    break;
  case Intrinsic::loongarch_lasx_concat_128_s:
  case Intrinsic::loongarch_lasx_concat_128_d:
  case Intrinsic::loongarch_lasx_concat_128:
    return DAG.getNode(ISD::CONCAT_VECTORS, DL, N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  }
  return SDValue();
}
 
static SDValue performMOVGR2FR_WCombine(SDNode *N, SelectionDAG &DAG,
                                        TargetLowering::DAGCombinerInfo &DCI,
                                        const LoongArchSubtarget &Subtarget) {
  // If the input to MOVGR2FR_W_LA64 is just MOVFR2GR_S_LA64 the the
  // conversion is unnecessary and can be replaced with the
  // MOVFR2GR_S_LA64 operand.
  SDValue Op0 = N->getOperand(0);
  if (Op0.getOpcode() == LoongArchISD::MOVFR2GR_S_LA64)
    return Op0.getOperand(0);
  return SDValue();
}
 
static SDValue performMOVFR2GR_SCombine(SDNode *N, SelectionDAG &DAG,
                                        TargetLowering::DAGCombinerInfo &DCI,
                                        const LoongArchSubtarget &Subtarget) {
  // If the input to MOVFR2GR_S_LA64 is just MOVGR2FR_W_LA64 then the
  // conversion is unnecessary and can be replaced with the MOVGR2FR_W_LA64
  // operand.
  SDValue Op0 = N->getOperand(0);
  if (Op0->getOpcode() == LoongArchISD::MOVGR2FR_W_LA64) {
    assert(Op0.getOperand(0).getValueType() == N->getSimpleValueType(0) &&
           "Unexpected value type!");
    return Op0.getOperand(0);
  }
  return SDValue();
}
 
static SDValue
performDemandedBitsCombine(SDNode *N, SelectionDAG &DAG,
                           TargetLowering::DAGCombinerInfo &DCI) {
  MVT VT = N->getSimpleValueType(0);
  unsigned NumBits = VT.getScalarSizeInBits();
 
  // Simplify the inputs.
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  APInt DemandedMask(APInt::getAllOnes(NumBits));
  if (TLI.SimplifyDemandedBits(SDValue(N, 0), DemandedMask, DCI))
    return SDValue(N, 0);
 
  return SDValue();
}
 
static SDValue
performSPLIT_PAIR_F64Combine(SDNode *N, SelectionDAG &DAG,
                             TargetLowering::DAGCombinerInfo &DCI,
                             const LoongArchSubtarget &Subtarget) {
  SDValue Op0 = N->getOperand(0);
  SDLoc DL(N);
 
  // If the input to SplitPairF64 is just BuildPairF64 then the operation is
  // redundant. Instead, use BuildPairF64's operands directly.
  if (Op0->getOpcode() == LoongArchISD::BUILD_PAIR_F64)
    return DCI.CombineTo(N, Op0.getOperand(0), Op0.getOperand(1));
 
  if (Op0->isUndef()) {
    SDValue Lo = DAG.getUNDEF(MVT::i32);
    SDValue Hi = DAG.getUNDEF(MVT::i32);
    return DCI.CombineTo(N, Lo, Hi);
  }
 
  // It's cheaper to materialise two 32-bit integers than to load a double
  // from the constant pool and transfer it to integer registers through the
  // stack.
  if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op0)) {
    APInt V = C->getValueAPF().bitcastToAPInt();
    SDValue Lo = DAG.getConstant(V.trunc(32), DL, MVT::i32);
    SDValue Hi = DAG.getConstant(V.lshr(32).trunc(32), DL, MVT::i32);
    return DCI.CombineTo(N, Lo, Hi);
  }
 
  return SDValue();
}
 
/// Do target-specific dag combines on LoongArchISD::VANDN nodes.
static SDValue performVANDNCombine(SDNode *N, SelectionDAG &DAG,
                                   TargetLowering::DAGCombinerInfo &DCI,
                                   const LoongArchSubtarget &Subtarget) {
  SDValue N0 = N->getOperand(0);
  SDValue N1 = N->getOperand(1);
  MVT VT = N->getSimpleValueType(0);
  SDLoc DL(N);
 
  // VANDN(undef, x) -> 0
  // VANDN(x, undef) -> 0
  if (N0.isUndef() || N1.isUndef())
    return DAG.getConstant(0, DL, VT);
 
  // VANDN(0, x) -> x
  if (ISD::isBuildVectorAllZeros(N0.getNode()))
    return N1;
 
  // VANDN(x, 0) -> 0
  if (ISD::isBuildVectorAllZeros(N1.getNode()))
    return DAG.getConstant(0, DL, VT);
 
  // VANDN(x, -1) -> NOT(x) -> XOR(x, -1)
  if (ISD::isBuildVectorAllOnes(N1.getNode()))
    return DAG.getNOT(DL, N0, VT);
 
  // Turn VANDN back to AND if input is inverted.
  if (SDValue Not = isNOT(N0, DAG))
    return DAG.getNode(ISD::AND, DL, VT, DAG.getBitcast(VT, Not), N1);
 
  // Folds for better commutativity:
  if (N1->hasOneUse()) {
    // VANDN(x,NOT(y)) -> AND(NOT(x),NOT(y)) -> NOT(OR(X,Y)).
    if (SDValue Not = isNOT(N1, DAG))
      return DAG.getNOT(
          DL, DAG.getNode(ISD::OR, DL, VT, N0, DAG.getBitcast(VT, Not)), VT);
 
    // VANDN(x, SplatVector(Imm)) -> AND(NOT(x), NOT(SplatVector(~Imm)))
    // -> NOT(OR(x, SplatVector(-Imm))
    // Combination is performed only when VT is v16i8/v32i8, using `vnori.b` to
    // gain benefits.
    if (!DCI.isBeforeLegalizeOps() && (VT == MVT::v16i8 || VT == MVT::v32i8) &&
        N1.getOpcode() == ISD::BUILD_VECTOR) {
      if (SDValue SplatValue =
              cast<BuildVectorSDNode>(N1.getNode())->getSplatValue()) {
        if (!N1->isOnlyUserOf(SplatValue.getNode()))
          return SDValue();
 
        if (auto *C = dyn_cast<ConstantSDNode>(SplatValue)) {
          uint8_t NCVal = static_cast<uint8_t>(~(C->getSExtValue()));
          SDValue Not =
              DAG.getSplat(VT, DL, DAG.getTargetConstant(NCVal, DL, MVT::i8));
          return DAG.getNOT(
              DL, DAG.getNode(ISD::OR, DL, VT, N0, DAG.getBitcast(VT, Not)),
              VT);
        }
      }
    }
  }
 
  return SDValue();
}
 
static SDValue ExtendSrcToDst(SDNode *N, SelectionDAG &DAG, unsigned ExtendOp) {
  SDLoc DL(N);
  EVT VT = N->getValueType(0);
  SDValue Src = N->getOperand(0);
  EVT SrcVT = Src.getValueType();
 
  unsigned DstElts = VT.getVectorNumElements();
  unsigned SrcEltBits = SrcVT.getScalarSizeInBits();
  unsigned DstEltBits = VT.getScalarSizeInBits();
 
  if (SrcEltBits >= DstEltBits)
    return SDValue();
 
  MVT WidenEltVT = MVT::getIntegerVT(DstEltBits);
  MVT WidenSrcVT = MVT::getVectorVT(WidenEltVT, DstElts);
 
  SDValue Extend = DAG.getNode(ExtendOp, DL, WidenSrcVT, Src);
  return DAG.getNode(N->getOpcode(), DL, VT, Extend);
}
 
// Merge two 64 to 32 convert instructions into one,
// e.g.
//   vffint.s.l $vr0, $vr1, $vr2
// will convert 4 si64 into 4 float at once.
// or
//   vftintrz.w.d $vr0, $vr1, $vr2
// which will convert 4 double into 4 si32 at once.
// also deal with their 256-bits LASX version.
static SDValue MergeBlocksConvert(SDNode *N, SelectionDAG &DAG, unsigned Opcode,
                                  unsigned BlockBits) {
  SDLoc DL(N);
  MVT DstVT = N->getSimpleValueType(0);
  SDValue Src = N->getOperand(0);
  MVT SrcVT = Src.getSimpleValueType();
  unsigned SrcBits = SrcVT.getSizeInBits();
 
  SmallVector<SDValue, 4> Blocks;
  unsigned BlockNumElts = BlockBits / SrcVT.getScalarSizeInBits();
  MVT BlockVT = MVT::getVectorVT(SrcVT.getScalarType(), BlockNumElts);
  if (Src.getOpcode() == ISD::CONCAT_VECTORS &&
      Src.getOperand(0).getValueType() == BlockVT) {
    for (unsigned i = 0; i < Src.getNumOperands(); ++i)
      Blocks.push_back(Src.getOperand(i));
  } else if (SrcBits > BlockBits) {
    // Wider than one register: extract each BlockBits-wide sub-vector.
    for (unsigned i = 0; i < SrcBits / BlockBits; ++i)
      Blocks.push_back(
          DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, BlockVT, Src,
                      DAG.getVectorIdxConstant(i * BlockNumElts, DL)));
  } else {
    BlockBits = SrcBits;
    Blocks.push_back(Src);
  }
 
  MVT NativeVecVT = MVT::getVectorVT(DstVT.getScalarType(),
                                     BlockBits / DstVT.getScalarSizeInBits());
  SmallVector<SDValue, 4> Parts;
  for (unsigned i = 0; i < Blocks.size(); i += 2) {
    SDValue Lo = Blocks[i];
    SDValue Hi = Blocks.size() > 1 ? Blocks[i + 1] : Lo;
    SDValue Res = DAG.getNode(Opcode, DL, NativeVecVT, Hi, Lo);
 
    if (BlockBits == 256) {
      SDValue Undef = DAG.getUNDEF(NativeVecVT);
      SmallVector<int, 8> Mask = {0, 1, 4, 5, 2, 3, 6, 7};
      Res = DAG.getVectorShuffle(NativeVecVT, DL, Res, Undef, Mask);
      Res = DAG.getBitcast(NativeVecVT, Res);
    }
 
    Parts.push_back(Res);
  }
 
  if (Blocks.size() == 1)
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, DstVT, Parts[0],
                       DAG.getVectorIdxConstant(0, DL));
  return DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Parts);
}
 
static SDValue performSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
                                        TargetLowering::DAGCombinerInfo &DCI,
                                        const LoongArchSubtarget &Subtarget) {
  SDLoc DL(N);
  EVT VT = N->getValueType(0);
  SDValue Src = N->getOperand(0);
  EVT SrcVT = Src.getValueType();
 
  if (VT.isVector()) {
    unsigned SrcEltBits = SrcVT.getScalarSizeInBits();
    unsigned DstEltBits = VT.getScalarSizeInBits();
    unsigned NumElts = VT.getVectorNumElements();
    unsigned BlockBits = Subtarget.hasExtLASX() ? 256 : 128;
 
    // Sign-extend src to avoid scalarization.
    if (SrcEltBits <= DstEltBits)
      return ExtendSrcToDst(N, DAG, ISD::SIGN_EXTEND);
 
    if (SrcEltBits != 64 || DstEltBits != 32 || !isPowerOf2_32(NumElts))
      return SDValue();
 
    if (!SrcVT.isSimple() || !VT.isSimple())
      return SDValue();
 
    // Combine [x]vffint.s.l for vector si64 to float conversion.
    return MergeBlocksConvert(N, DAG, LoongArchISD::VFFINT, BlockBits);
  }
 
  if (VT != MVT::f32 && VT != MVT::f64)
    return SDValue();
  if (VT == MVT::f32 && !Subtarget.hasBasicF())
    return SDValue();
  if (VT == MVT::f64 && !Subtarget.hasBasicD())
    return SDValue();
 
  // Only optimize when the source and destination types have the same width.
  if (VT.getSizeInBits() != N->getOperand(0).getValueSizeInBits())
    return SDValue();
 
  // If the result of an integer load is only used by an integer-to-float
  // conversion, use a fp load instead. This eliminates an integer-to-float-move
  // (movgr2fr) instruction.
  if (ISD::isNormalLoad(Src.getNode()) && Src.hasOneUse() &&
      // Do not change the width of a volatile load. This condition check is
      // inspired by AArch64.
      !cast<LoadSDNode>(Src)->isVolatile()) {
    LoadSDNode *LN0 = cast<LoadSDNode>(Src);
    SDValue Load = DAG.getLoad(VT, DL, LN0->getChain(), LN0->getBasePtr(),
                               LN0->getPointerInfo(), LN0->getAlign(),
                               LN0->getMemOperand()->getFlags());
 
    // Make sure successors of the original load stay after it by updating them
    // to use the new Chain.
    DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
    return DAG.getNode(LoongArchISD::SITOF, SDLoc(N), VT, Load);
  }
 
  return SDValue();
}
 
static SDValue performUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
                                        TargetLowering::DAGCombinerInfo &DCI,
                                        const LoongArchSubtarget &Subtarget) {
  SDLoc DL(N);
  EVT VT = N->getValueType(0);
 
  // Zero-extend src to avoid scalarization.
  if (VT.isVector())
    return ExtendSrcToDst(N, DAG, ISD::ZERO_EXTEND);
 
  return SDValue();
}
 
// Using [X]VFTINTRZ_W_D for double to signed 32-bit integer conversion.
// For example:
//   v4i32 = fp_to_sint (concat_vectors v2f64, v2f64)
// Can be combined into:
//   v4i32 = VFTINTRZ_W_D v2f64. v2f64
static SDValue performFP_TO_INTCombine(SDNode *N, SelectionDAG &DAG,
                                       TargetLowering::DAGCombinerInfo &DCI,
                                       const LoongArchSubtarget &Subtarget) {
  if (!Subtarget.hasExtLSX())
    return SDValue();
 
  SDLoc DL(N);
  EVT DstVT = N->getValueType(0);
  SDValue Src = N->getOperand(0);
  EVT SrcVT = Src.getValueType();
  bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
 
  if (!DstVT.isVector() || !DstVT.isSimple() || !SrcVT.isSimple())
    return SDValue();
 
  unsigned SrcEltBits = SrcVT.getScalarSizeInBits();
  unsigned SrcBits = SrcVT.getSizeInBits();
  unsigned DstEltBits = DstVT.getScalarSizeInBits();
  unsigned NumElts = DstVT.getVectorNumElements();
  unsigned BlockBits = Subtarget.hasExtLASX() ? 256 : 128;
 
  if (!isPowerOf2_32(NumElts) || !isPowerOf2_32(DstEltBits))
    return SDValue();
 
  if (SrcBits % BlockBits != 0 && SrcBits != 128)
    return SDValue();
 
  if (DstEltBits < 32) {
    MVT PromoteVT = MVT::getVectorVT(MVT::getIntegerVT(32), NumElts);
    SDValue Conv = DAG.getNode(N->getOpcode(), DL, PromoteVT, Src);
    return DAG.getNode(ISD::TRUNCATE, DL, DstVT, Conv);
  }
 
  if (SrcEltBits != 64 || DstEltBits != 32)
    return SDValue();
 
  if (!IsSigned) {
    // LASX already has pattern for double convert to uint32.
    if (Subtarget.hasExtLASX())
      return SDValue();
    MVT TmpVT = MVT::getVectorVT(MVT::i64, NumElts);
    SDValue Tmp = DAG.getNode(ISD::FP_TO_SINT, DL, TmpVT, Src);
    return DAG.getNode(ISD::TRUNCATE, DL, DstVT, Tmp);
  }
 
  return MergeBlocksConvert(N, DAG, LoongArchISD::VFTINTRZ, BlockBits);
}
 
// Try to widen AND, OR and XOR nodes to VT in order to remove casts around
// logical operations, like in the example below.
//   or (and (truncate x, truncate y)),
//      (xor (truncate z, build_vector (constants)))
// Given a target type \p VT, we generate
//   or (and x, y), (xor z, zext(build_vector (constants)))
// given x, y and z are of type \p VT. We can do so, if operands are either
// truncates from VT types, the second operand is a vector of constants, can
// be recursively promoted or is an existing extension we can extend further.
static SDValue PromoteMaskArithmetic(SDValue N, const SDLoc &DL, EVT VT,
                                     SelectionDAG &DAG,
                                     const LoongArchSubtarget &Subtarget,
                                     unsigned Depth) {
  // Limit recursion to avoid excessive compile times.
  if (Depth >= SelectionDAG::MaxRecursionDepth)
    return SDValue();
 
  if (!ISD::isBitwiseLogicOp(N.getOpcode()))
    return SDValue();
 
  SDValue N0 = N.getOperand(0);
  SDValue N1 = N.getOperand(1);
 
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  if (!TLI.isOperationLegalOrPromote(N.getOpcode(), VT))
    return SDValue();
 
  if (SDValue NN0 =
          PromoteMaskArithmetic(N0, DL, VT, DAG, Subtarget, Depth + 1))
    N0 = NN0;