RISCVISelLowering.cpp

Go to the documentation of this file.

//===-- RISCVISelLowering.cpp - RISCV 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 RISCV uses to lower LLVM code into a
// selection DAG.
//
//===----------------------------------------------------------------------===//
 
#include "RISCVISelLowering.h"
#include "MCTargetDesc/RISCVMatInt.h"
#include "RISCV.h"
#include "RISCVMachineFunctionInfo.h"
#include "RISCVRegisterInfo.h"
#include "RISCVSubtarget.h"
#include "RISCVTargetMachine.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/MemoryLocation.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
#include "llvm/CodeGen/ValueTypes.h"
#include "llvm/IR/DiagnosticInfo.h"
#include "llvm/IR/DiagnosticPrinter.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/IntrinsicsRISCV.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
 
using namespace llvm;
 
#define DEBUG_TYPE "riscv-lower"
 
STATISTIC(NumTailCalls, "Number of tail calls");
 
RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM,
                                         const RISCVSubtarget &STI)
    : TargetLowering(TM), Subtarget(STI) {
 
  if (Subtarget.isRV32E())
    report_fatal_error("Codegen not yet implemented for RV32E");
 
  RISCVABI::ABI ABI = Subtarget.getTargetABI();
  assert(ABI != RISCVABI::ABI_Unknown && "Improperly initialised target ABI");
 
  if ((ABI == RISCVABI::ABI_ILP32F || ABI == RISCVABI::ABI_LP64F) &&
      !Subtarget.hasStdExtF()) {
    errs() << "Hard-float 'f' ABI can't be used for a target that "
                "doesn't support the F instruction set extension (ignoring "
                          "target-abi)\n";
    ABI = Subtarget.is64Bit() ? RISCVABI::ABI_LP64 : RISCVABI::ABI_ILP32;
  } else if ((ABI == RISCVABI::ABI_ILP32D || ABI == RISCVABI::ABI_LP64D) &&
             !Subtarget.hasStdExtD()) {
    errs() << "Hard-float 'd' ABI can't be used for a target that "
              "doesn't support the D instruction set extension (ignoring "
              "target-abi)\n";
    ABI = Subtarget.is64Bit() ? RISCVABI::ABI_LP64 : RISCVABI::ABI_ILP32;
  }
 
  switch (ABI) {
  default:
    report_fatal_error("Don't know how to lower this ABI");
  case RISCVABI::ABI_ILP32:
  case RISCVABI::ABI_ILP32F:
  case RISCVABI::ABI_ILP32D:
  case RISCVABI::ABI_LP64:
  case RISCVABI::ABI_LP64F:
  case RISCVABI::ABI_LP64D:
    break;
  }
 
  MVT XLenVT = Subtarget.getXLenVT();
 
  // Set up the register classes.
  addRegisterClass(XLenVT, &RISCV::GPRRegClass);
 
  if (Subtarget.hasStdExtZfh())
    addRegisterClass(MVT::f16, &RISCV::FPR16RegClass);
  if (Subtarget.hasStdExtF())
    addRegisterClass(MVT::f32, &RISCV::FPR32RegClass);
  if (Subtarget.hasStdExtD())
    addRegisterClass(MVT::f64, &RISCV::FPR64RegClass);
 
  static const MVT::SimpleValueType BoolVecVTs[] = {
      MVT::nxv1i1,  MVT::nxv2i1,  MVT::nxv4i1, MVT::nxv8i1,
      MVT::nxv16i1, MVT::nxv32i1, MVT::nxv64i1};
  static const MVT::SimpleValueType IntVecVTs[] = {
      MVT::nxv1i8,  MVT::nxv2i8,   MVT::nxv4i8,   MVT::nxv8i8,  MVT::nxv16i8,
      MVT::nxv32i8, MVT::nxv64i8,  MVT::nxv1i16,  MVT::nxv2i16, MVT::nxv4i16,
      MVT::nxv8i16, MVT::nxv16i16, MVT::nxv32i16, MVT::nxv1i32, MVT::nxv2i32,
      MVT::nxv4i32, MVT::nxv8i32,  MVT::nxv16i32, MVT::nxv1i64, MVT::nxv2i64,
      MVT::nxv4i64, MVT::nxv8i64};
  static const MVT::SimpleValueType F16VecVTs[] = {
      MVT::nxv1f16, MVT::nxv2f16,  MVT::nxv4f16,
      MVT::nxv8f16, MVT::nxv16f16, MVT::nxv32f16};
  static const MVT::SimpleValueType F32VecVTs[] = {
      MVT::nxv1f32, MVT::nxv2f32, MVT::nxv4f32, MVT::nxv8f32, MVT::nxv16f32};
  static const MVT::SimpleValueType F64VecVTs[] = {
      MVT::nxv1f64, MVT::nxv2f64, MVT::nxv4f64, MVT::nxv8f64};
 
  if (Subtarget.hasVInstructions()) {
    auto addRegClassForRVV = [this](MVT VT) {
      unsigned Size = VT.getSizeInBits().getKnownMinValue();
      assert(Size <= 512 && isPowerOf2_32(Size));
      const TargetRegisterClass *RC;
      if (Size <= 64)
        RC = &RISCV::VRRegClass;
      else if (Size == 128)
        RC = &RISCV::VRM2RegClass;
      else if (Size == 256)
        RC = &RISCV::VRM4RegClass;
      else
        RC = &RISCV::VRM8RegClass;
 
      addRegisterClass(VT, RC);
    };
 
    for (MVT VT : BoolVecVTs)
      addRegClassForRVV(VT);
    for (MVT VT : IntVecVTs) {
      if (VT.getVectorElementType() == MVT::i64 &&
          !Subtarget.hasVInstructionsI64())
        continue;
      addRegClassForRVV(VT);
    }
 
    if (Subtarget.hasVInstructionsF16())
      for (MVT VT : F16VecVTs)
        addRegClassForRVV(VT);
 
    if (Subtarget.hasVInstructionsF32())
      for (MVT VT : F32VecVTs)
        addRegClassForRVV(VT);
 
    if (Subtarget.hasVInstructionsF64())
      for (MVT VT : F64VecVTs)
        addRegClassForRVV(VT);
 
    if (Subtarget.useRVVForFixedLengthVectors()) {
      auto addRegClassForFixedVectors = [this](MVT VT) {
        MVT ContainerVT = getContainerForFixedLengthVector(VT);
        unsigned RCID = getRegClassIDForVecVT(ContainerVT);
        const RISCVRegisterInfo &TRI = *Subtarget.getRegisterInfo();
        addRegisterClass(VT, TRI.getRegClass(RCID));
      };
      for (MVT VT : MVT::integer_fixedlen_vector_valuetypes())
        if (useRVVForFixedLengthVectorVT(VT))
          addRegClassForFixedVectors(VT);
 
      for (MVT VT : MVT::fp_fixedlen_vector_valuetypes())
        if (useRVVForFixedLengthVectorVT(VT))
          addRegClassForFixedVectors(VT);
    }
  }
 
  // Compute derived properties from the register classes.
  computeRegisterProperties(STI.getRegisterInfo());
 
  setStackPointerRegisterToSaveRestore(RISCV::X2);
 
  for (auto N : {ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD})
    setLoadExtAction(N, XLenVT, MVT::i1, Promote);
 
  // TODO: add all necessary setOperationAction calls.
  setOperationAction(ISD::DYNAMIC_STACKALLOC, XLenVT, Expand);
 
  setOperationAction(ISD::BR_JT, MVT::Other, Expand);
  setOperationAction(ISD::BR_CC, XLenVT, Expand);
  setOperationAction(ISD::BRCOND, MVT::Other, Custom);
  setOperationAction(ISD::SELECT_CC, XLenVT, Expand);
 
  setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
  setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
 
  setOperationAction(ISD::VASTART, MVT::Other, Custom);
  setOperationAction(ISD::VAARG, MVT::Other, Expand);
  setOperationAction(ISD::VACOPY, MVT::Other, Expand);
  setOperationAction(ISD::VAEND, MVT::Other, Expand);
 
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
  if (!Subtarget.hasStdExtZbb()) {
    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
  }
 
  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::UADDO, MVT::i32, Custom);
    setOperationAction(ISD::USUBO, MVT::i32, Custom);
    setOperationAction(ISD::UADDSAT, MVT::i32, Custom);
    setOperationAction(ISD::USUBSAT, MVT::i32, Custom);
  } else {
    setLibcallName(RTLIB::SHL_I128, nullptr);
    setLibcallName(RTLIB::SRL_I128, nullptr);
    setLibcallName(RTLIB::SRA_I128, nullptr);
    setLibcallName(RTLIB::MUL_I128, nullptr);
    setLibcallName(RTLIB::MULO_I64, nullptr);
  }
 
  if (!Subtarget.hasStdExtM()) {
    setOperationAction(ISD::MUL, XLenVT, Expand);
    setOperationAction(ISD::MULHS, XLenVT, Expand);
    setOperationAction(ISD::MULHU, XLenVT, Expand);
    setOperationAction(ISD::SDIV, XLenVT, Expand);
    setOperationAction(ISD::UDIV, XLenVT, Expand);
    setOperationAction(ISD::SREM, XLenVT, Expand);
    setOperationAction(ISD::UREM, XLenVT, Expand);
  } else {
    if (Subtarget.is64Bit()) {
      setOperationAction(ISD::MUL, MVT::i32, Custom);
      setOperationAction(ISD::MUL, MVT::i128, Custom);
 
      setOperationAction(ISD::SDIV, MVT::i8, Custom);
      setOperationAction(ISD::UDIV, MVT::i8, Custom);
      setOperationAction(ISD::UREM, MVT::i8, Custom);
      setOperationAction(ISD::SDIV, MVT::i16, Custom);
      setOperationAction(ISD::UDIV, MVT::i16, Custom);
      setOperationAction(ISD::UREM, MVT::i16, Custom);
      setOperationAction(ISD::SDIV, MVT::i32, Custom);
      setOperationAction(ISD::UDIV, MVT::i32, Custom);
      setOperationAction(ISD::UREM, MVT::i32, Custom);
    } else {
      setOperationAction(ISD::MUL, MVT::i64, Custom);
    }
  }
 
  setOperationAction(ISD::SDIVREM, XLenVT, Expand);
  setOperationAction(ISD::UDIVREM, XLenVT, Expand);
  setOperationAction(ISD::SMUL_LOHI, XLenVT, Expand);
  setOperationAction(ISD::UMUL_LOHI, XLenVT, Expand);
 
  setOperationAction(ISD::SHL_PARTS, XLenVT, Custom);
  setOperationAction(ISD::SRL_PARTS, XLenVT, Custom);
  setOperationAction(ISD::SRA_PARTS, XLenVT, Custom);
 
  if (Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbp() ||
      Subtarget.hasStdExtZbkb()) {
    if (Subtarget.is64Bit()) {
      setOperationAction(ISD::ROTL, MVT::i32, Custom);
      setOperationAction(ISD::ROTR, MVT::i32, Custom);
    }
  } else {
    setOperationAction(ISD::ROTL, XLenVT, Expand);
    setOperationAction(ISD::ROTR, XLenVT, Expand);
  }
 
  if (Subtarget.hasStdExtZbp()) {
    // Custom lower bswap/bitreverse so we can convert them to GREVI to enable
    // more combining.
    setOperationAction(ISD::BITREVERSE, XLenVT,   Custom);
    setOperationAction(ISD::BSWAP,      XLenVT,   Custom);
    setOperationAction(ISD::BITREVERSE, MVT::i8,  Custom);
    // BSWAP i8 doesn't exist.
    setOperationAction(ISD::BITREVERSE, MVT::i16, Custom);
    setOperationAction(ISD::BSWAP,      MVT::i16, Custom);
 
    if (Subtarget.is64Bit()) {
      setOperationAction(ISD::BITREVERSE, MVT::i32, Custom);
      setOperationAction(ISD::BSWAP,      MVT::i32, Custom);
    }
  } else {
    // With Zbb we have an XLen rev8 instruction, but not GREVI. So we'll
    // pattern match it directly in isel.
    setOperationAction(ISD::BSWAP, XLenVT,
                       (Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbkb())
                           ? Legal
                           : Expand);
  }
 
  if (Subtarget.hasStdExtZbb()) {
    setOperationAction(ISD::SMIN, XLenVT, Legal);
    setOperationAction(ISD::SMAX, XLenVT, Legal);
    setOperationAction(ISD::UMIN, XLenVT, Legal);
    setOperationAction(ISD::UMAX, XLenVT, Legal);
 
    if (Subtarget.is64Bit()) {
      setOperationAction(ISD::CTTZ, MVT::i32, Custom);
      setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Custom);
      setOperationAction(ISD::CTLZ, MVT::i32, Custom);
      setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Custom);
    }
  } else {
    setOperationAction(ISD::CTTZ, XLenVT, Expand);
    setOperationAction(ISD::CTLZ, XLenVT, Expand);
    setOperationAction(ISD::CTPOP, XLenVT, Expand);
  }
 
  if (Subtarget.hasStdExtZbt()) {
    setOperationAction(ISD::FSHL, XLenVT, Custom);
    setOperationAction(ISD::FSHR, XLenVT, Custom);
    setOperationAction(ISD::SELECT, XLenVT, Legal);
 
    if (Subtarget.is64Bit()) {
      setOperationAction(ISD::FSHL, MVT::i32, Custom);
      setOperationAction(ISD::FSHR, MVT::i32, Custom);
    }
  } else {
    setOperationAction(ISD::SELECT, XLenVT, Custom);
  }
 
  static const ISD::CondCode FPCCToExpand[] = {
      ISD::SETOGT, ISD::SETOGE, ISD::SETONE, ISD::SETUEQ, ISD::SETUGT,
      ISD::SETUGE, ISD::SETULT, ISD::SETULE, ISD::SETUNE, ISD::SETGT,
      ISD::SETGE,  ISD::SETNE,  ISD::SETO,   ISD::SETUO};
 
  static const ISD::NodeType FPOpToExpand[] = {
      ISD::FSIN, ISD::FCOS,       ISD::FSINCOS,   ISD::FPOW,
      ISD::FREM, ISD::FP16_TO_FP, ISD::FP_TO_FP16};
 
  if (Subtarget.hasStdExtZfh())
    setOperationAction(ISD::BITCAST, MVT::i16, Custom);
 
  if (Subtarget.hasStdExtZfh()) {
    setOperationAction(ISD::FMINNUM, MVT::f16, Legal);
    setOperationAction(ISD::FMAXNUM, MVT::f16, Legal);
    setOperationAction(ISD::LRINT, MVT::f16, Legal);
    setOperationAction(ISD::LLRINT, MVT::f16, Legal);
    setOperationAction(ISD::LROUND, MVT::f16, Legal);
    setOperationAction(ISD::LLROUND, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_LRINT, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_LLRINT, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_LROUND, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_LLROUND, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_FADD, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_FMA, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_FSUB, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_FMUL, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_FDIV, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_FP_ROUND, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FSQRT, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_FSETCC, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_FSETCCS, MVT::f16, Legal);
    for (auto CC : FPCCToExpand)
      setCondCodeAction(CC, MVT::f16, Expand);
    setOperationAction(ISD::SELECT_CC, MVT::f16, Expand);
    setOperationAction(ISD::SELECT, MVT::f16, Custom);
    setOperationAction(ISD::BR_CC, MVT::f16, Expand);
 
    setOperationAction(ISD::FREM,       MVT::f16, Promote);
    setOperationAction(ISD::FCEIL,      MVT::f16, Promote);
    setOperationAction(ISD::FFLOOR,     MVT::f16, Promote);
    setOperationAction(ISD::FNEARBYINT, MVT::f16, Promote);
    setOperationAction(ISD::FRINT,      MVT::f16, Promote);
    setOperationAction(ISD::FROUND,     MVT::f16, Promote);
    setOperationAction(ISD::FROUNDEVEN, MVT::f16, Promote);
    setOperationAction(ISD::FTRUNC,     MVT::f16, Promote);
    setOperationAction(ISD::FPOW,       MVT::f16, Promote);
    setOperationAction(ISD::FPOWI,      MVT::f16, Promote);
    setOperationAction(ISD::FCOS,       MVT::f16, Promote);
    setOperationAction(ISD::FSIN,       MVT::f16, Promote);
    setOperationAction(ISD::FSINCOS,    MVT::f16, Promote);
    setOperationAction(ISD::FEXP,       MVT::f16, Promote);
    setOperationAction(ISD::FEXP2,      MVT::f16, Promote);
    setOperationAction(ISD::FLOG,       MVT::f16, Promote);
    setOperationAction(ISD::FLOG2,      MVT::f16, Promote);
    setOperationAction(ISD::FLOG10,     MVT::f16, Promote);
 
    // FIXME: Need to promote f16 STRICT_* to f32 libcalls, but we don't have
    // complete support for all operations in LegalizeDAG.
 
    // We need to custom promote this.
    if (Subtarget.is64Bit())
      setOperationAction(ISD::FPOWI, MVT::i32, Custom);
  }
 
  if (Subtarget.hasStdExtF()) {
    setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
    setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
    setOperationAction(ISD::LRINT, MVT::f32, Legal);
    setOperationAction(ISD::LLRINT, MVT::f32, Legal);
    setOperationAction(ISD::LROUND, MVT::f32, Legal);
    setOperationAction(ISD::LLROUND, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_LRINT, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_LLRINT, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_LROUND, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_LLROUND, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FADD, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FMA, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FSUB, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FMUL, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FDIV, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FSQRT, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FSETCC, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FSETCCS, MVT::f32, Legal);
    for (auto CC : FPCCToExpand)
      setCondCodeAction(CC, MVT::f32, Expand);
    setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
    setOperationAction(ISD::SELECT, MVT::f32, Custom);
    setOperationAction(ISD::BR_CC, MVT::f32, Expand);
    for (auto Op : FPOpToExpand)
      setOperationAction(Op, MVT::f32, Expand);
    setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
    setTruncStoreAction(MVT::f32, MVT::f16, Expand);
  }
 
  if (Subtarget.hasStdExtF() && Subtarget.is64Bit())
    setOperationAction(ISD::BITCAST, MVT::i32, Custom);
 
  if (Subtarget.hasStdExtD()) {
    setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
    setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
    setOperationAction(ISD::LRINT, MVT::f64, Legal);
    setOperationAction(ISD::LLRINT, MVT::f64, Legal);
    setOperationAction(ISD::LROUND, MVT::f64, Legal);
    setOperationAction(ISD::LLROUND, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_LRINT, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_LLRINT, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_LROUND, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_LLROUND, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_FMA, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_FADD, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_FSUB, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_FMUL, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_FDIV, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_FSQRT, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_FSETCC, MVT::f64, Legal);
    setOperationAction(ISD::STRICT_FSETCCS, MVT::f64, Legal);
    for (auto CC : FPCCToExpand)
      setCondCodeAction(CC, MVT::f64, Expand);
    setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
    setOperationAction(ISD::SELECT, MVT::f64, Custom);
    setOperationAction(ISD::BR_CC, MVT::f64, Expand);
    setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
    setTruncStoreAction(MVT::f64, MVT::f32, Expand);
    for (auto Op : FPOpToExpand)
      setOperationAction(Op, MVT::f64, Expand);
    setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
    setTruncStoreAction(MVT::f64, MVT::f16, Expand);
  }
 
  if (Subtarget.is64Bit()) {
    setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
    setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
    setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Custom);
    setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Custom);
  }
 
  if (Subtarget.hasStdExtF()) {
    setOperationAction(ISD::FP_TO_UINT_SAT, XLenVT, Custom);
    setOperationAction(ISD::FP_TO_SINT_SAT, XLenVT, Custom);
 
    setOperationAction(ISD::STRICT_FP_TO_UINT, XLenVT, Legal);
    setOperationAction(ISD::STRICT_FP_TO_SINT, XLenVT, Legal);
    setOperationAction(ISD::STRICT_UINT_TO_FP, XLenVT, Legal);
    setOperationAction(ISD::STRICT_SINT_TO_FP, XLenVT, Legal);
 
    setOperationAction(ISD::FLT_ROUNDS_, XLenVT, Custom);
    setOperationAction(ISD::SET_ROUNDING, MVT::Other, Custom);
  }
 
  setOperationAction(ISD::GlobalAddress, XLenVT, Custom);
  setOperationAction(ISD::BlockAddress, XLenVT, Custom);
  setOperationAction(ISD::ConstantPool, XLenVT, Custom);
  setOperationAction(ISD::JumpTable, XLenVT, Custom);
 
  setOperationAction(ISD::GlobalTLSAddress, XLenVT, Custom);
 
  // TODO: On M-mode only targets, the cycle[h] CSR may not be present.
  // Unfortunately this can't be determined just from the ISA naming string.
  setOperationAction(ISD::READCYCLECOUNTER, MVT::i64,
                     Subtarget.is64Bit() ? Legal : Custom);
 
  setOperationAction(ISD::TRAP, MVT::Other, Legal);
  setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
  if (Subtarget.is64Bit())
    setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i32, Custom);
 
  if (Subtarget.hasStdExtA()) {
    setMaxAtomicSizeInBitsSupported(Subtarget.getXLen());
    setMinCmpXchgSizeInBits(32);
  } else {
    setMaxAtomicSizeInBitsSupported(0);
  }
 
  setBooleanContents(ZeroOrOneBooleanContent);
 
  if (Subtarget.hasVInstructions()) {
    setBooleanVectorContents(ZeroOrOneBooleanContent);
 
    setOperationAction(ISD::VSCALE, XLenVT, Custom);
 
    // RVV intrinsics may have illegal operands.
    // We also need to custom legalize vmv.x.s.
    setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i8, Custom);
    setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i16, Custom);
    setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i8, Custom);
    setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i16, Custom);
    if (Subtarget.is64Bit()) {
      setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i32, Custom);
    } else {
      setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
      setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
    }
 
    setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
    setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
 
    static const unsigned IntegerVPOps[] = {
        ISD::VP_ADD,         ISD::VP_SUB,         ISD::VP_MUL,
        ISD::VP_SDIV,        ISD::VP_UDIV,        ISD::VP_SREM,
        ISD::VP_UREM,        ISD::VP_AND,         ISD::VP_OR,
        ISD::VP_XOR,         ISD::VP_ASHR,        ISD::VP_LSHR,
        ISD::VP_SHL,         ISD::VP_REDUCE_ADD,  ISD::VP_REDUCE_AND,
        ISD::VP_REDUCE_OR,   ISD::VP_REDUCE_XOR,  ISD::VP_REDUCE_SMAX,
        ISD::VP_REDUCE_SMIN, ISD::VP_REDUCE_UMAX, ISD::VP_REDUCE_UMIN,
        ISD::VP_MERGE,       ISD::VP_SELECT};
 
    static const unsigned FloatingPointVPOps[] = {
        ISD::VP_FADD,        ISD::VP_FSUB,        ISD::VP_FMUL,
        ISD::VP_FDIV,        ISD::VP_REDUCE_FADD, ISD::VP_REDUCE_SEQ_FADD,
        ISD::VP_REDUCE_FMIN, ISD::VP_REDUCE_FMAX, ISD::VP_MERGE,
        ISD::VP_SELECT};
 
    if (!Subtarget.is64Bit()) {
      // We must custom-lower certain vXi64 operations on RV32 due to the vector
      // element type being illegal.
      setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::i64, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::i64, Custom);
 
      setOperationAction(ISD::VECREDUCE_ADD, MVT::i64, Custom);
      setOperationAction(ISD::VECREDUCE_AND, MVT::i64, Custom);
      setOperationAction(ISD::VECREDUCE_OR, MVT::i64, Custom);
      setOperationAction(ISD::VECREDUCE_XOR, MVT::i64, Custom);
      setOperationAction(ISD::VECREDUCE_SMAX, MVT::i64, Custom);
      setOperationAction(ISD::VECREDUCE_SMIN, MVT::i64, Custom);
      setOperationAction(ISD::VECREDUCE_UMAX, MVT::i64, Custom);
      setOperationAction(ISD::VECREDUCE_UMIN, MVT::i64, Custom);
 
      setOperationAction(ISD::VP_REDUCE_ADD, MVT::i64, Custom);
      setOperationAction(ISD::VP_REDUCE_AND, MVT::i64, Custom);
      setOperationAction(ISD::VP_REDUCE_OR, MVT::i64, Custom);
      setOperationAction(ISD::VP_REDUCE_XOR, MVT::i64, Custom);
      setOperationAction(ISD::VP_REDUCE_SMAX, MVT::i64, Custom);
      setOperationAction(ISD::VP_REDUCE_SMIN, MVT::i64, Custom);
      setOperationAction(ISD::VP_REDUCE_UMAX, MVT::i64, Custom);
      setOperationAction(ISD::VP_REDUCE_UMIN, MVT::i64, Custom);
    }
 
    for (MVT VT : BoolVecVTs) {
      setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
 
      // Mask VTs are custom-expanded into a series of standard nodes
      setOperationAction(ISD::TRUNCATE, VT, Custom);
      setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
      setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
 
      setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
 
      setOperationAction(ISD::SELECT, VT, Custom);
      setOperationAction(ISD::SELECT_CC, VT, Expand);
      setOperationAction(ISD::VSELECT, VT, Expand);
      setOperationAction(ISD::VP_MERGE, VT, Expand);
      setOperationAction(ISD::VP_SELECT, VT, Expand);
 
      setOperationAction(ISD::VP_AND, VT, Custom);
      setOperationAction(ISD::VP_OR, VT, Custom);
      setOperationAction(ISD::VP_XOR, VT, Custom);
 
      setOperationAction(ISD::VECREDUCE_AND, VT, Custom);
      setOperationAction(ISD::VECREDUCE_OR, VT, Custom);
      setOperationAction(ISD::VECREDUCE_XOR, VT, Custom);
 
      setOperationAction(ISD::VP_REDUCE_AND, VT, Custom);
      setOperationAction(ISD::VP_REDUCE_OR, VT, Custom);
      setOperationAction(ISD::VP_REDUCE_XOR, VT, Custom);
 
      // RVV has native int->float & float->int conversions where the
      // element type sizes are within one power-of-two of each other. Any
      // wider distances between type sizes have to be lowered as sequences
      // which progressively narrow the gap in stages.
      setOperationAction(ISD::SINT_TO_FP, VT, Custom);
      setOperationAction(ISD::UINT_TO_FP, VT, Custom);
      setOperationAction(ISD::FP_TO_SINT, VT, Custom);
      setOperationAction(ISD::FP_TO_UINT, VT, Custom);
 
      // Expand all extending loads to types larger than this, and truncating
      // stores from types larger than this.
      for (MVT OtherVT : MVT::integer_scalable_vector_valuetypes()) {
        setTruncStoreAction(OtherVT, VT, Expand);
        setLoadExtAction(ISD::EXTLOAD, OtherVT, VT, Expand);
        setLoadExtAction(ISD::SEXTLOAD, OtherVT, VT, Expand);
        setLoadExtAction(ISD::ZEXTLOAD, OtherVT, VT, Expand);
      }
    }
 
    for (MVT VT : IntVecVTs) {
      if (VT.getVectorElementType() == MVT::i64 &&
          !Subtarget.hasVInstructionsI64())
        continue;
 
      setOperationAction(ISD::SPLAT_VECTOR, VT, Legal);
      setOperationAction(ISD::SPLAT_VECTOR_PARTS, VT, Custom);
 
      // Vectors implement MULHS/MULHU.
      setOperationAction(ISD::SMUL_LOHI, VT, Expand);
      setOperationAction(ISD::UMUL_LOHI, VT, Expand);
 
      // nxvXi64 MULHS/MULHU requires the V extension instead of Zve64*.
      if (VT.getVectorElementType() == MVT::i64 && !Subtarget.hasStdExtV()) {
        setOperationAction(ISD::MULHU, VT, Expand);
        setOperationAction(ISD::MULHS, VT, Expand);
      }
 
      setOperationAction(ISD::SMIN, VT, Legal);
      setOperationAction(ISD::SMAX, VT, Legal);
      setOperationAction(ISD::UMIN, VT, Legal);
      setOperationAction(ISD::UMAX, VT, Legal);
 
      setOperationAction(ISD::ROTL, VT, Expand);
      setOperationAction(ISD::ROTR, VT, Expand);
 
      setOperationAction(ISD::CTTZ, VT, Expand);
      setOperationAction(ISD::CTLZ, VT, Expand);
      setOperationAction(ISD::CTPOP, VT, Expand);
 
      setOperationAction(ISD::BSWAP, VT, Expand);
 
      // Custom-lower extensions and truncations from/to mask types.
      setOperationAction(ISD::ANY_EXTEND, VT, Custom);
      setOperationAction(ISD::SIGN_EXTEND, VT, Custom);
      setOperationAction(ISD::ZERO_EXTEND, VT, Custom);
 
      // RVV has native int->float & float->int conversions where the
      // element type sizes are within one power-of-two of each other. Any
      // wider distances between type sizes have to be lowered as sequences
      // which progressively narrow the gap in stages.
      setOperationAction(ISD::SINT_TO_FP, VT, Custom);
      setOperationAction(ISD::UINT_TO_FP, VT, Custom);
      setOperationAction(ISD::FP_TO_SINT, VT, Custom);
      setOperationAction(ISD::FP_TO_UINT, VT, Custom);
 
      setOperationAction(ISD::SADDSAT, VT, Legal);
      setOperationAction(ISD::UADDSAT, VT, Legal);
      setOperationAction(ISD::SSUBSAT, VT, Legal);
      setOperationAction(ISD::USUBSAT, VT, Legal);
 
      // Integer VTs are lowered as a series of "RISCVISD::TRUNCATE_VECTOR_VL"
      // nodes which truncate by one power of two at a time.
      setOperationAction(ISD::TRUNCATE, VT, Custom);
 
      // Custom-lower insert/extract operations to simplify patterns.
      setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
 
      // Custom-lower reduction operations to set up the corresponding custom
      // nodes' operands.
      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);
 
      for (unsigned VPOpc : IntegerVPOps)
        setOperationAction(VPOpc, VT, Custom);
 
      setOperationAction(ISD::LOAD, VT, Custom);
      setOperationAction(ISD::STORE, VT, Custom);
 
      setOperationAction(ISD::MLOAD, VT, Custom);
      setOperationAction(ISD::MSTORE, VT, Custom);
      setOperationAction(ISD::MGATHER, VT, Custom);
      setOperationAction(ISD::MSCATTER, VT, Custom);
 
      setOperationAction(ISD::VP_LOAD, VT, Custom);
      setOperationAction(ISD::VP_STORE, VT, Custom);
      setOperationAction(ISD::VP_GATHER, VT, Custom);
      setOperationAction(ISD::VP_SCATTER, VT, Custom);
 
      setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
      setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
 
      setOperationAction(ISD::SELECT, VT, Custom);
      setOperationAction(ISD::SELECT_CC, VT, Expand);
 
      setOperationAction(ISD::STEP_VECTOR, VT, Custom);
      setOperationAction(ISD::VECTOR_REVERSE, VT, Custom);
 
      for (MVT OtherVT : MVT::integer_scalable_vector_valuetypes()) {
        setTruncStoreAction(VT, OtherVT, Expand);
        setLoadExtAction(ISD::EXTLOAD, OtherVT, VT, Expand);
        setLoadExtAction(ISD::SEXTLOAD, OtherVT, VT, Expand);
        setLoadExtAction(ISD::ZEXTLOAD, OtherVT, VT, Expand);
      }
 
      // Lower CTLZ_ZERO_UNDEF and CTTZ_ZERO_UNDEF if we have a floating point
      // type that can represent the value exactly.
      if (VT.getVectorElementType() != MVT::i64) {
        MVT FloatEltVT =
            VT.getVectorElementType() == MVT::i32 ? MVT::f64 : MVT::f32;
        EVT FloatVT = MVT::getVectorVT(FloatEltVT, VT.getVectorElementCount());
        if (isTypeLegal(FloatVT)) {
          setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Custom);
          setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Custom);
        }
      }
    }
 
    // Expand various CCs to best match the RVV ISA, which natively supports UNE
    // but no other unordered comparisons, and supports all ordered comparisons
    // except ONE. Additionally, we expand GT,OGT,GE,OGE for optimization
    // purposes; they are expanded to their swapped-operand CCs (LT,OLT,LE,OLE),
    // and we pattern-match those back to the "original", swapping operands once
    // more. This way we catch both operations and both "vf" and "fv" forms with
    // fewer patterns.
    static const ISD::CondCode VFPCCToExpand[] = {
        ISD::SETO,   ISD::SETONE, ISD::SETUEQ, ISD::SETUGT,
        ISD::SETUGE, ISD::SETULT, ISD::SETULE, ISD::SETUO,
        ISD::SETGT,  ISD::SETOGT, ISD::SETGE,  ISD::SETOGE,
    };
 
    // Sets common operation actions on RVV floating-point vector types.
    const auto SetCommonVFPActions = [&](MVT VT) {
      setOperationAction(ISD::SPLAT_VECTOR, VT, Legal);
      // RVV has native FP_ROUND & FP_EXTEND conversions where the element type
      // sizes are within one power-of-two of each other. Therefore conversions
      // between vXf16 and vXf64 must be lowered as sequences which convert via
      // vXf32.
      setOperationAction(ISD::FP_ROUND, VT, Custom);
      setOperationAction(ISD::FP_EXTEND, VT, Custom);
      // Custom-lower insert/extract operations to simplify patterns.
      setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
      // Expand various condition codes (explained above).
      for (auto CC : VFPCCToExpand)
        setCondCodeAction(CC, VT, Expand);
 
      setOperationAction(ISD::FMINNUM, VT, Legal);
      setOperationAction(ISD::FMAXNUM, VT, Legal);
 
      setOperationAction(ISD::FTRUNC, VT, Custom);
      setOperationAction(ISD::FCEIL, VT, Custom);
      setOperationAction(ISD::FFLOOR, VT, Custom);
 
      setOperationAction(ISD::VECREDUCE_FADD, VT, Custom);
      setOperationAction(ISD::VECREDUCE_SEQ_FADD, VT, Custom);
      setOperationAction(ISD::VECREDUCE_FMIN, VT, Custom);
      setOperationAction(ISD::VECREDUCE_FMAX, VT, Custom);
 
      setOperationAction(ISD::FCOPYSIGN, VT, Legal);
 
      setOperationAction(ISD::LOAD, VT, Custom);
      setOperationAction(ISD::STORE, VT, Custom);
 
      setOperationAction(ISD::MLOAD, VT, Custom);
      setOperationAction(ISD::MSTORE, VT, Custom);
      setOperationAction(ISD::MGATHER, VT, Custom);
      setOperationAction(ISD::MSCATTER, VT, Custom);
 
      setOperationAction(ISD::VP_LOAD, VT, Custom);
      setOperationAction(ISD::VP_STORE, VT, Custom);
      setOperationAction(ISD::VP_GATHER, VT, Custom);
      setOperationAction(ISD::VP_SCATTER, VT, Custom);
 
      setOperationAction(ISD::SELECT, VT, Custom);
      setOperationAction(ISD::SELECT_CC, VT, Expand);
 
      setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
      setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
 
      setOperationAction(ISD::VECTOR_REVERSE, VT, Custom);
 
      for (unsigned VPOpc : FloatingPointVPOps)
        setOperationAction(VPOpc, VT, Custom);
    };
 
    // Sets common extload/truncstore actions on RVV floating-point vector
    // types.
    const auto SetCommonVFPExtLoadTruncStoreActions =
        [&](MVT VT, ArrayRef<MVT::SimpleValueType> SmallerVTs) {
          for (auto SmallVT : SmallerVTs) {
            setTruncStoreAction(VT, SmallVT, Expand);
            setLoadExtAction(ISD::EXTLOAD, VT, SmallVT, Expand);
          }
        };
 
    if (Subtarget.hasVInstructionsF16())
      for (MVT VT : F16VecVTs)
        SetCommonVFPActions(VT);
 
    for (MVT VT : F32VecVTs) {
      if (Subtarget.hasVInstructionsF32())
        SetCommonVFPActions(VT);
      SetCommonVFPExtLoadTruncStoreActions(VT, F16VecVTs);
    }
 
    for (MVT VT : F64VecVTs) {
      if (Subtarget.hasVInstructionsF64())
        SetCommonVFPActions(VT);
      SetCommonVFPExtLoadTruncStoreActions(VT, F16VecVTs);
      SetCommonVFPExtLoadTruncStoreActions(VT, F32VecVTs);
    }
 
    if (Subtarget.useRVVForFixedLengthVectors()) {
      for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) {
        if (!useRVVForFixedLengthVectorVT(VT))
          continue;
 
        // By default everything must be expanded.
        for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op)
          setOperationAction(Op, VT, Expand);
        for (MVT OtherVT : MVT::integer_fixedlen_vector_valuetypes()) {
          setTruncStoreAction(VT, OtherVT, Expand);
          setLoadExtAction(ISD::EXTLOAD, OtherVT, VT, Expand);
          setLoadExtAction(ISD::SEXTLOAD, OtherVT, VT, Expand);
          setLoadExtAction(ISD::ZEXTLOAD, OtherVT, VT, Expand);
        }
 
        // We use EXTRACT_SUBVECTOR as a "cast" from scalable to fixed.
        setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
        setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
 
        setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
        setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
 
        setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
        setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
 
        setOperationAction(ISD::LOAD, VT, Custom);
        setOperationAction(ISD::STORE, VT, Custom);
 
        setOperationAction(ISD::SETCC, VT, Custom);
 
        setOperationAction(ISD::SELECT, VT, Custom);
 
        setOperationAction(ISD::TRUNCATE, VT, Custom);
 
        setOperationAction(ISD::BITCAST, VT, Custom);
 
        setOperationAction(ISD::VECREDUCE_AND, VT, Custom);
        setOperationAction(ISD::VECREDUCE_OR, VT, Custom);
        setOperationAction(ISD::VECREDUCE_XOR, VT, Custom);
 
        setOperationAction(ISD::VP_REDUCE_AND, VT, Custom);
        setOperationAction(ISD::VP_REDUCE_OR, VT, Custom);
        setOperationAction(ISD::VP_REDUCE_XOR, VT, Custom);
 
        setOperationAction(ISD::SINT_TO_FP, VT, Custom);
        setOperationAction(ISD::UINT_TO_FP, VT, Custom);
        setOperationAction(ISD::FP_TO_SINT, VT, Custom);
        setOperationAction(ISD::FP_TO_UINT, VT, Custom);
 
        // Operations below are different for between masks and other vectors.
        if (VT.getVectorElementType() == MVT::i1) {
          setOperationAction(ISD::VP_AND, VT, Custom);
          setOperationAction(ISD::VP_OR, VT, Custom);
          setOperationAction(ISD::VP_XOR, VT, Custom);
          setOperationAction(ISD::AND, VT, Custom);
          setOperationAction(ISD::OR, VT, Custom);
          setOperationAction(ISD::XOR, VT, Custom);
          continue;
        }
 
        // Use SPLAT_VECTOR to prevent type legalization from destroying the
        // splats when type legalizing i64 scalar on RV32.
        // FIXME: Use SPLAT_VECTOR for all types? DAGCombine probably needs
        // improvements first.
        if (!Subtarget.is64Bit() && VT.getVectorElementType() == MVT::i64) {
          setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
          setOperationAction(ISD::SPLAT_VECTOR_PARTS, VT, Custom);
        }
 
        setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
        setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
 
        setOperationAction(ISD::MLOAD, VT, Custom);
        setOperationAction(ISD::MSTORE, VT, Custom);
        setOperationAction(ISD::MGATHER, VT, Custom);
        setOperationAction(ISD::MSCATTER, VT, Custom);
 
        setOperationAction(ISD::VP_LOAD, VT, Custom);
        setOperationAction(ISD::VP_STORE, VT, Custom);
        setOperationAction(ISD::VP_GATHER, VT, Custom);
        setOperationAction(ISD::VP_SCATTER, VT, Custom);
 
        setOperationAction(ISD::ADD, VT, Custom);
        setOperationAction(ISD::MUL, VT, Custom);
        setOperationAction(ISD::SUB, VT, Custom);
        setOperationAction(ISD::AND, VT, Custom);
        setOperationAction(ISD::OR, VT, Custom);
        setOperationAction(ISD::XOR, VT, Custom);
        setOperationAction(ISD::SDIV, VT, Custom);
        setOperationAction(ISD::SREM, VT, Custom);
        setOperationAction(ISD::UDIV, VT, Custom);
        setOperationAction(ISD::UREM, VT, Custom);
        setOperationAction(ISD::SHL, VT, Custom);
        setOperationAction(ISD::SRA, VT, Custom);
        setOperationAction(ISD::SRL, VT, Custom);
 
        setOperationAction(ISD::SMIN, VT, Custom);
        setOperationAction(ISD::SMAX, VT, Custom);
        setOperationAction(ISD::UMIN, VT, Custom);
        setOperationAction(ISD::UMAX, VT, Custom);
        setOperationAction(ISD::ABS,  VT, Custom);
 
        // vXi64 MULHS/MULHU requires the V extension instead of Zve64*.
        if (VT.getVectorElementType() != MVT::i64 || Subtarget.hasStdExtV()) {
          setOperationAction(ISD::MULHS, VT, Custom);
          setOperationAction(ISD::MULHU, VT, Custom);
        }
 
        setOperationAction(ISD::SADDSAT, VT, Custom);
        setOperationAction(ISD::UADDSAT, VT, Custom);
        setOperationAction(ISD::SSUBSAT, VT, Custom);
        setOperationAction(ISD::USUBSAT, VT, Custom);
 
        setOperationAction(ISD::VSELECT, VT, Custom);
        setOperationAction(ISD::SELECT_CC, VT, Expand);
 
        setOperationAction(ISD::ANY_EXTEND, VT, Custom);
        setOperationAction(ISD::SIGN_EXTEND, VT, Custom);
        setOperationAction(ISD::ZERO_EXTEND, VT, Custom);
 
        // Custom-lower reduction operations to set up the corresponding custom
        // nodes' operands.
        setOperationAction(ISD::VECREDUCE_ADD, 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);
 
        for (unsigned VPOpc : IntegerVPOps)
          setOperationAction(VPOpc, VT, Custom);
 
        // Lower CTLZ_ZERO_UNDEF and CTTZ_ZERO_UNDEF if we have a floating point
        // type that can represent the value exactly.
        if (VT.getVectorElementType() != MVT::i64) {
          MVT FloatEltVT =
              VT.getVectorElementType() == MVT::i32 ? MVT::f64 : MVT::f32;
          EVT FloatVT =
              MVT::getVectorVT(FloatEltVT, VT.getVectorElementCount());
          if (isTypeLegal(FloatVT)) {
            setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Custom);
            setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Custom);
          }
        }
      }
 
      for (MVT VT : MVT::fp_fixedlen_vector_valuetypes()) {
        if (!useRVVForFixedLengthVectorVT(VT))
          continue;
 
        // By default everything must be expanded.
        for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op)
          setOperationAction(Op, VT, Expand);
        for (MVT OtherVT : MVT::fp_fixedlen_vector_valuetypes()) {
          setLoadExtAction(ISD::EXTLOAD, OtherVT, VT, Expand);
          setTruncStoreAction(VT, OtherVT, Expand);
        }
 
        // We use EXTRACT_SUBVECTOR as a "cast" from scalable to fixed.
        setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
        setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
 
        setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
        setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
        setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
        setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
        setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
 
        setOperationAction(ISD::LOAD, VT, Custom);
        setOperationAction(ISD::STORE, VT, Custom);
        setOperationAction(ISD::MLOAD, VT, Custom);
        setOperationAction(ISD::MSTORE, VT, Custom);
        setOperationAction(ISD::MGATHER, VT, Custom);
        setOperationAction(ISD::MSCATTER, VT, Custom);
 
        setOperationAction(ISD::VP_LOAD, VT, Custom);
        setOperationAction(ISD::VP_STORE, VT, Custom);
        setOperationAction(ISD::VP_GATHER, VT, Custom);
        setOperationAction(ISD::VP_SCATTER, VT, Custom);
 
        setOperationAction(ISD::FADD, VT, Custom);
        setOperationAction(ISD::FSUB, VT, Custom);
        setOperationAction(ISD::FMUL, VT, Custom);
        setOperationAction(ISD::FDIV, VT, Custom);
        setOperationAction(ISD::FNEG, VT, Custom);
        setOperationAction(ISD::FABS, VT, Custom);
        setOperationAction(ISD::FCOPYSIGN, VT, Custom);
        setOperationAction(ISD::FSQRT, VT, Custom);
        setOperationAction(ISD::FMA, VT, Custom);
        setOperationAction(ISD::FMINNUM, VT, Custom);
        setOperationAction(ISD::FMAXNUM, VT, Custom);
 
        setOperationAction(ISD::FP_ROUND, VT, Custom);
        setOperationAction(ISD::FP_EXTEND, VT, Custom);
 
        setOperationAction(ISD::FTRUNC, VT, Custom);
        setOperationAction(ISD::FCEIL, VT, Custom);
        setOperationAction(ISD::FFLOOR, VT, Custom);
 
        for (auto CC : VFPCCToExpand)
          setCondCodeAction(CC, VT, Expand);
 
        setOperationAction(ISD::VSELECT, VT, Custom);
        setOperationAction(ISD::SELECT, VT, Custom);
        setOperationAction(ISD::SELECT_CC, VT, Expand);
 
        setOperationAction(ISD::BITCAST, VT, Custom);
 
        setOperationAction(ISD::VECREDUCE_FADD, VT, Custom);
        setOperationAction(ISD::VECREDUCE_SEQ_FADD, VT, Custom);
        setOperationAction(ISD::VECREDUCE_FMIN, VT, Custom);
        setOperationAction(ISD::VECREDUCE_FMAX, VT, Custom);
 
        for (unsigned VPOpc : FloatingPointVPOps)
          setOperationAction(VPOpc, VT, Custom);
      }
 
      // Custom-legalize bitcasts from fixed-length vectors to scalar types.
      setOperationAction(ISD::BITCAST, MVT::i8, Custom);
      setOperationAction(ISD::BITCAST, MVT::i16, Custom);
      setOperationAction(ISD::BITCAST, MVT::i32, Custom);
      setOperationAction(ISD::BITCAST, MVT::i64, Custom);
      if (Subtarget.hasStdExtZfh())
        setOperationAction(ISD::BITCAST, MVT::f16, Custom);
      if (Subtarget.hasStdExtF())
        setOperationAction(ISD::BITCAST, MVT::f32, Custom);
      if (Subtarget.hasStdExtD())
        setOperationAction(ISD::BITCAST, MVT::f64, Custom);
    }
  }
 
  // Function alignments.
  const Align FunctionAlignment(Subtarget.hasStdExtC() ? 2 : 4);
  setMinFunctionAlignment(FunctionAlignment);
  setPrefFunctionAlignment(FunctionAlignment);
 
  setMinimumJumpTableEntries(5);
 
  // Jumps are expensive, compared to logic
  setJumpIsExpensive();
 
  setTargetDAGCombine(ISD::ADD);
  setTargetDAGCombine(ISD::SUB);
  setTargetDAGCombine(ISD::AND);
  setTargetDAGCombine(ISD::OR);
  setTargetDAGCombine(ISD::XOR);
  setTargetDAGCombine(ISD::ANY_EXTEND);
  if (Subtarget.hasStdExtF()) {
    setTargetDAGCombine(ISD::ZERO_EXTEND);
    setTargetDAGCombine(ISD::FP_TO_SINT);
    setTargetDAGCombine(ISD::FP_TO_UINT);
    setTargetDAGCombine(ISD::FP_TO_SINT_SAT);
    setTargetDAGCombine(ISD::FP_TO_UINT_SAT);
  }
  if (Subtarget.hasVInstructions()) {
    setTargetDAGCombine(ISD::FCOPYSIGN);
    setTargetDAGCombine(ISD::MGATHER);
    setTargetDAGCombine(ISD::MSCATTER);
    setTargetDAGCombine(ISD::VP_GATHER);
    setTargetDAGCombine(ISD::VP_SCATTER);
    setTargetDAGCombine(ISD::SRA);
    setTargetDAGCombine(ISD::SRL);
    setTargetDAGCombine(ISD::SHL);
    setTargetDAGCombine(ISD::STORE);
  }
}
 
EVT RISCVTargetLowering::getSetCCResultType(const DataLayout &DL,
                                            LLVMContext &Context,
                                            EVT VT) const {
  if (!VT.isVector())
    return getPointerTy(DL);
  if (Subtarget.hasVInstructions() &&
      (VT.isScalableVector() || Subtarget.useRVVForFixedLengthVectors()))
    return EVT::getVectorVT(Context, MVT::i1, VT.getVectorElementCount());
  return VT.changeVectorElementTypeToInteger();
}
 
MVT RISCVTargetLowering::getVPExplicitVectorLengthTy() const {
  return Subtarget.getXLenVT();
}
 
bool RISCVTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
                                             const CallInst &I,
                                             MachineFunction &MF,
                                             unsigned Intrinsic) const {
  auto &DL = I.getModule()->getDataLayout();
  switch (Intrinsic) {
  default:
    return false;
  case Intrinsic::riscv_masked_atomicrmw_xchg_i32:
  case Intrinsic::riscv_masked_atomicrmw_add_i32:
  case Intrinsic::riscv_masked_atomicrmw_sub_i32:
  case Intrinsic::riscv_masked_atomicrmw_nand_i32:
  case Intrinsic::riscv_masked_atomicrmw_max_i32:
  case Intrinsic::riscv_masked_atomicrmw_min_i32:
  case Intrinsic::riscv_masked_atomicrmw_umax_i32:
  case Intrinsic::riscv_masked_atomicrmw_umin_i32:
  case Intrinsic::riscv_masked_cmpxchg_i32: {
    PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    Info.memVT = MVT::getVT(PtrTy->getPointerElementType());
    Info.ptrVal = I.getArgOperand(0);
    Info.offset = 0;
    Info.align = Align(4);
    Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore |
                 MachineMemOperand::MOVolatile;
    return true;
  }
  case Intrinsic::riscv_masked_strided_load:
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    Info.ptrVal = I.getArgOperand(1);
    Info.memVT = getValueType(DL, I.getType()->getScalarType());
    Info.align = Align(DL.getTypeSizeInBits(I.getType()->getScalarType()) / 8);
    Info.size = MemoryLocation::UnknownSize;
    Info.flags |= MachineMemOperand::MOLoad;
    return true;
  case Intrinsic::riscv_masked_strided_store:
    Info.opc = ISD::INTRINSIC_VOID;
    Info.ptrVal = I.getArgOperand(1);
    Info.memVT =
        getValueType(DL, I.getArgOperand(0)->getType()->getScalarType());
    Info.align = Align(
        DL.getTypeSizeInBits(I.getArgOperand(0)->getType()->getScalarType()) /
        8);
    Info.size = MemoryLocation::UnknownSize;
    Info.flags |= MachineMemOperand::MOStore;
    return true;
  }
}
 
bool RISCVTargetLowering::isLegalAddressingMode(const DataLayout &DL,
                                                const AddrMode &AM, Type *Ty,
                                                unsigned AS,
                                                Instruction *I) const {
  // No global is ever allowed as a base.
  if (AM.BaseGV)
    return false;
 
  // Require a 12-bit signed offset.
  if (!isInt<12>(AM.BaseOffs))
    return false;
 
  switch (AM.Scale) {
  case 0: // "r+i" or just "i", depending on HasBaseReg.
    break;
  case 1:
    if (!AM.HasBaseReg) // allow "r+i".
      break;
    return false; // disallow "r+r" or "r+r+i".
  default:
    return false;
  }
 
  return true;
}
 
bool RISCVTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
  return isInt<12>(Imm);
}
 
bool RISCVTargetLowering::isLegalAddImmediate(int64_t Imm) const {
  return isInt<12>(Imm);
}
 
// On RV32, 64-bit integers are split into their high and low parts and held
// in two different registers, so the trunc is free since the low register can
// just be used.
bool RISCVTargetLowering::isTruncateFree(Type *SrcTy, Type *DstTy) const {
  if (Subtarget.is64Bit() || !SrcTy->isIntegerTy() || !DstTy->isIntegerTy())
    return false;
  unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
  unsigned DestBits = DstTy->getPrimitiveSizeInBits();
  return (SrcBits == 64 && DestBits == 32);
}
 
bool RISCVTargetLowering::isTruncateFree(EVT SrcVT, EVT DstVT) const {
  if (Subtarget.is64Bit() || SrcVT.isVector() || DstVT.isVector() ||
      !SrcVT.isInteger() || !DstVT.isInteger())
    return false;
  unsigned SrcBits = SrcVT.getSizeInBits();
  unsigned DestBits = DstVT.getSizeInBits();
  return (SrcBits == 64 && DestBits == 32);
}
 
bool RISCVTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
  // Zexts are free if they can be combined with a load.
  // Don't advertise i32->i64 zextload as being free for RV64. It interacts
  // poorly with type legalization of compares preferring sext.
  if (auto *LD = dyn_cast<LoadSDNode>(Val)) {
    EVT MemVT = LD->getMemoryVT();
    if ((MemVT == MVT::i8 || MemVT == MVT::i16) &&
        (LD->getExtensionType() == ISD::NON_EXTLOAD ||
         LD->getExtensionType() == ISD::ZEXTLOAD))
      return true;
  }
 
  return TargetLowering::isZExtFree(Val, VT2);
}
 
bool RISCVTargetLowering::isSExtCheaperThanZExt(EVT SrcVT, EVT DstVT) const {
  return Subtarget.is64Bit() && SrcVT == MVT::i32 && DstVT == MVT::i64;
}
 
bool RISCVTargetLowering::isCheapToSpeculateCttz() const {
  return Subtarget.hasStdExtZbb();
}
 
bool RISCVTargetLowering::isCheapToSpeculateCtlz() const {
  return Subtarget.hasStdExtZbb();
}
 
bool RISCVTargetLowering::hasAndNotCompare(SDValue Y) const {
  EVT VT = Y.getValueType();
 
  // FIXME: Support vectors once we have tests.
  if (VT.isVector())
    return false;
 
  return (Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbp() ||
          Subtarget.hasStdExtZbkb()) &&
         !isa<ConstantSDNode>(Y);
}
 
/// Check if sinking \p I's operands to I's basic block is profitable, because
/// the operands can be folded into a target instruction, e.g.
/// splats of scalars can fold into vector instructions.
bool RISCVTargetLowering::shouldSinkOperands(
    Instruction *I, SmallVectorImpl<Use *> &Ops) const {
  using namespace llvm::PatternMatch;
 
  if (!I->getType()->isVectorTy() || !Subtarget.hasVInstructions())
    return false;
 
  auto IsSinker = [&](Instruction *I, int Operand) {
    switch (I->getOpcode()) {
    case Instruction::Add:
    case Instruction::Sub:
    case Instruction::Mul:
    case Instruction::And:
    case Instruction::Or:
    case Instruction::Xor:
    case Instruction::FAdd:
    case Instruction::FSub:
    case Instruction::FMul:
    case Instruction::FDiv:
    case Instruction::ICmp:
    case Instruction::FCmp:
      return true;
    case Instruction::Shl:
    case Instruction::LShr:
    case Instruction::AShr:
    case Instruction::UDiv:
    case Instruction::SDiv:
    case Instruction::URem:
    case Instruction::SRem:
      return Operand == 1;
    case Instruction::Call:
      if (auto *II = dyn_cast<IntrinsicInst>(I)) {
        switch (II->getIntrinsicID()) {
        case Intrinsic::fma:
          return Operand == 0 || Operand == 1;
        // FIXME: Our patterns can only match vx/vf instructions when the splat
        // it on the RHS, because TableGen doesn't recognize our VP operations
        // as commutative.
        case Intrinsic::vp_add:
        case Intrinsic::vp_mul:
        case Intrinsic::vp_and:
        case Intrinsic::vp_or:
        case Intrinsic::vp_xor:
        case Intrinsic::vp_fadd:
        case Intrinsic::vp_fmul:
        case Intrinsic::vp_shl:
        case Intrinsic::vp_lshr:
        case Intrinsic::vp_ashr:
        case Intrinsic::vp_udiv:
        case Intrinsic::vp_sdiv:
        case Intrinsic::vp_urem:
        case Intrinsic::vp_srem:
          return Operand == 1;
        // ... with the exception of vp.sub/vp.fsub/vp.fdiv, which have
        // explicit patterns for both LHS and RHS (as 'vr' versions).
        case Intrinsic::vp_sub:
        case Intrinsic::vp_fsub:
        case Intrinsic::vp_fdiv:
          return Operand == 0 || Operand == 1;
        default:
          return false;
        }
      }
      return false;
    default:
      return false;
    }
  };
 
  for (auto OpIdx : enumerate(I->operands())) {
    if (!IsSinker(I, OpIdx.index()))
      continue;
 
    Instruction *Op = dyn_cast<Instruction>(OpIdx.value().get());
    // Make sure we are not already sinking this operand
    if (!Op || any_of(Ops, [&](Use *U) { return U->get() == Op; }))
      continue;
 
    // We are looking for a splat that can be sunk.
    if (!match(Op, m_Shuffle(m_InsertElt(m_Undef(), m_Value(), m_ZeroInt()),
                             m_Undef(), m_ZeroMask())))
      continue;
 
    // All uses of the shuffle should be sunk to avoid duplicating it across gpr
    // and vector registers
    for (Use &U : Op->uses()) {
      Instruction *Insn = cast<Instruction>(U.getUser());
      if (!IsSinker(Insn, U.getOperandNo()))
        return false;
    }
 
    Ops.push_back(&Op->getOperandUse(0));
    Ops.push_back(&OpIdx.value());
  }
  return true;
}
 
bool RISCVTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
                                       bool ForCodeSize) const {
  // FIXME: Change to Zfhmin once f16 becomes a legal type with Zfhmin.
  if (VT == MVT::f16 && !Subtarget.hasStdExtZfh())
    return false;
  if (VT == MVT::f32 && !Subtarget.hasStdExtF())
    return false;
  if (VT == MVT::f64 && !Subtarget.hasStdExtD())
    return false;
  return Imm.isZero();
}
 
bool RISCVTargetLowering::hasBitPreservingFPLogic(EVT VT) const {
  return (VT == MVT::f16 && Subtarget.hasStdExtZfh()) ||
         (VT == MVT::f32 && Subtarget.hasStdExtF()) ||
         (VT == MVT::f64 && Subtarget.hasStdExtD());
}
 
MVT RISCVTargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
                                                      CallingConv::ID CC,
                                                      EVT VT) const {
  // Use f32 to pass f16 if it is legal and Zfh is not enabled.
  // We might still end up using a GPR but that will be decided based on ABI.
  // FIXME: Change to Zfhmin once f16 becomes a legal type with Zfhmin.
  if (VT == MVT::f16 && Subtarget.hasStdExtF() && !Subtarget.hasStdExtZfh())
    return MVT::f32;
 
  return TargetLowering::getRegisterTypeForCallingConv(Context, CC, VT);
}
 
unsigned RISCVTargetLowering::getNumRegistersForCallingConv(LLVMContext &Context,
                                                           CallingConv::ID CC,
                                                           EVT VT) const {
  // Use f32 to pass f16 if it is legal and Zfh is not enabled.
  // We might still end up using a GPR but that will be decided based on ABI.
  // FIXME: Change to Zfhmin once f16 becomes a legal type with Zfhmin.
  if (VT == MVT::f16 && Subtarget.hasStdExtF() && !Subtarget.hasStdExtZfh())
    return 1;
 
  return TargetLowering::getNumRegistersForCallingConv(Context, CC, VT);
}
 
// Changes the condition code and swaps operands if necessary, so the SetCC
// operation matches one of the comparisons supported directly by branches
// in the RISC-V 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) {
  // Convert X > -1 to X >= 0.
  if (CC == ISD::SETGT && isAllOnesConstant(RHS)) {
    RHS = DAG.getConstant(0, DL, RHS.getValueType());
    CC = ISD::SETGE;
    return;
  }
  // Convert X < 1 to 0 >= X.
  if (CC == ISD::SETLT && isOneConstant(RHS)) {
    RHS = LHS;
    LHS = DAG.getConstant(0, DL, RHS.getValueType());
    CC = ISD::SETGE;
    return;
  }
 
  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;
  }
}
 
RISCVII::VLMUL RISCVTargetLowering::getLMUL(MVT VT) {
  assert(VT.isScalableVector() && "Expecting a scalable vector type");
  unsigned KnownSize = VT.getSizeInBits().getKnownMinValue();
  if (VT.getVectorElementType() == MVT::i1)
    KnownSize *= 8;
 
  switch (KnownSize) {
  default:
    llvm_unreachable("Invalid LMUL.");
  case 8:
    return RISCVII::VLMUL::LMUL_F8;
  case 16:
    return RISCVII::VLMUL::LMUL_F4;
  case 32:
    return RISCVII::VLMUL::LMUL_F2;
  case 64:
    return RISCVII::VLMUL::LMUL_1;
  case 128:
    return RISCVII::VLMUL::LMUL_2;
  case 256:
    return RISCVII::VLMUL::LMUL_4;
  case 512:
    return RISCVII::VLMUL::LMUL_8;
  }
}
 
unsigned RISCVTargetLowering::getRegClassIDForLMUL(RISCVII::VLMUL LMul) {
  switch (LMul) {
  default:
    llvm_unreachable("Invalid LMUL.");
  case RISCVII::VLMUL::LMUL_F8:
  case RISCVII::VLMUL::LMUL_F4:
  case RISCVII::VLMUL::LMUL_F2:
  case RISCVII::VLMUL::LMUL_1:
    return RISCV::VRRegClassID;
  case RISCVII::VLMUL::LMUL_2:
    return RISCV::VRM2RegClassID;
  case RISCVII::VLMUL::LMUL_4:
    return RISCV::VRM4RegClassID;
  case RISCVII::VLMUL::LMUL_8:
    return RISCV::VRM8RegClassID;
  }
}
 
unsigned RISCVTargetLowering::getSubregIndexByMVT(MVT VT, unsigned Index) {
  RISCVII::VLMUL LMUL = getLMUL(VT);
  if (LMUL == RISCVII::VLMUL::LMUL_F8 ||
      LMUL == RISCVII::VLMUL::LMUL_F4 ||
      LMUL == RISCVII::VLMUL::LMUL_F2 ||
      LMUL == RISCVII::VLMUL::LMUL_1) {
    static_assert(RISCV::sub_vrm1_7 == RISCV::sub_vrm1_0 + 7,
                  "Unexpected subreg numbering");
    return RISCV::sub_vrm1_0 + Index;
  }
  if (LMUL == RISCVII::VLMUL::LMUL_2) {
    static_assert(RISCV::sub_vrm2_3 == RISCV::sub_vrm2_0 + 3,
                  "Unexpected subreg numbering");
    return RISCV::sub_vrm2_0 + Index;
  }
  if (LMUL == RISCVII::VLMUL::LMUL_4) {
    static_assert(RISCV::sub_vrm4_1 == RISCV::sub_vrm4_0 + 1,
                  "Unexpected subreg numbering");
    return RISCV::sub_vrm4_0 + Index;
  }
  llvm_unreachable("Invalid vector type.");
}
 
unsigned RISCVTargetLowering::getRegClassIDForVecVT(MVT VT) {
  if (VT.getVectorElementType() == MVT::i1)
    return RISCV::VRRegClassID;
  return getRegClassIDForLMUL(getLMUL(VT));
}
 
// Attempt to decompose a subvector insert/extract between VecVT and
// SubVecVT via subregister indices. Returns the subregister index that
// can perform the subvector insert/extract with the given element index, as
// well as the index corresponding to any leftover subvectors that must be
// further inserted/extracted within the register class for SubVecVT.
std::pair<unsigned, unsigned>
RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
    MVT VecVT, MVT SubVecVT, unsigned InsertExtractIdx,
    const RISCVRegisterInfo *TRI) {
  static_assert((RISCV::VRM8RegClassID > RISCV::VRM4RegClassID &&
                 RISCV::VRM4RegClassID > RISCV::VRM2RegClassID &&
                 RISCV::VRM2RegClassID > RISCV::VRRegClassID),
                "Register classes not ordered");
  unsigned VecRegClassID = getRegClassIDForVecVT(VecVT);
  unsigned SubRegClassID = getRegClassIDForVecVT(SubVecVT);
  // Try to compose a subregister index that takes us from the incoming
  // LMUL>1 register class down to the outgoing one. At each step we half
  // the LMUL:
  //   nxv16i32@12 -> nxv2i32: sub_vrm4_1_then_sub_vrm2_1_then_sub_vrm1_0
  // Note that this is not guaranteed to find a subregister index, such as
  // when we are extracting from one VR type to another.
  unsigned SubRegIdx = RISCV::NoSubRegister;
  for (const unsigned RCID :
       {RISCV::VRM4RegClassID, RISCV::VRM2RegClassID, RISCV::VRRegClassID})
    if (VecRegClassID > RCID && SubRegClassID <= RCID) {
      VecVT = VecVT.getHalfNumVectorElementsVT();
      bool IsHi =
          InsertExtractIdx >= VecVT.getVectorElementCount().getKnownMinValue();
      SubRegIdx = TRI->composeSubRegIndices(SubRegIdx,
                                            getSubregIndexByMVT(VecVT, IsHi));
      if (IsHi)
        InsertExtractIdx -= VecVT.getVectorElementCount().getKnownMinValue();
    }
  return {SubRegIdx, InsertExtractIdx};
}
 
// Permit combining of mask vectors as BUILD_VECTOR never expands to scalar
// stores for those types.
bool RISCVTargetLowering::mergeStoresAfterLegalization(EVT VT) const {
  return !Subtarget.useRVVForFixedLengthVectors() ||
         (VT.isFixedLengthVector() && VT.getVectorElementType() == MVT::i1);
}
 
bool RISCVTargetLowering::isLegalElementTypeForRVV(Type *ScalarTy) const {
  if (ScalarTy->isPointerTy())
    return true;
 
  if (ScalarTy->isIntegerTy(8) || ScalarTy->isIntegerTy(16) ||
      ScalarTy->isIntegerTy(32))
    return true;
 
  if (ScalarTy->isIntegerTy(64))
    return Subtarget.hasVInstructionsI64();
 
  if (ScalarTy->isHalfTy())
    return Subtarget.hasVInstructionsF16();
  if (ScalarTy->isFloatTy())
    return Subtarget.hasVInstructionsF32();
  if (ScalarTy->isDoubleTy())
    return Subtarget.hasVInstructionsF64();
 
  return false;
}
 
static SDValue getVLOperand(SDValue Op) {
  assert((Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
          Op.getOpcode() == ISD::INTRINSIC_W_CHAIN) &&
         "Unexpected opcode");
  bool HasChain = Op.getOpcode() == ISD::INTRINSIC_W_CHAIN;
  unsigned IntNo = Op.getConstantOperandVal(HasChain ? 1 : 0);
  const RISCVVIntrinsicsTable::RISCVVIntrinsicInfo *II =
      RISCVVIntrinsicsTable::getRISCVVIntrinsicInfo(IntNo);
  if (!II)
    return SDValue();
  return Op.getOperand(II->VLOperand + 1 + HasChain);
}
 
static bool useRVVForFixedLengthVectorVT(MVT VT,
                                         const RISCVSubtarget &Subtarget) {
  assert(VT.isFixedLengthVector() && "Expected a fixed length vector type!");
  if (!Subtarget.useRVVForFixedLengthVectors())
    return false;
 
  // We only support a set of vector types with a consistent maximum fixed size
  // across all supported vector element types to avoid legalization issues.
  // Therefore -- since the largest is v1024i8/v512i16/etc -- the largest
  // fixed-length vector type we support is 1024 bytes.
  if (VT.getFixedSizeInBits() > 1024 * 8)
    return false;
 
  unsigned MinVLen = Subtarget.getMinRVVVectorSizeInBits();
 
  MVT EltVT = VT.getVectorElementType();
 
  // Don't use RVV for vectors we cannot scalarize if required.
  switch (EltVT.SimpleTy) {
  // i1 is supported but has different rules.
  default:
    return false;
  case MVT::i1:
    // Masks can only use a single register.
    if (VT.getVectorNumElements() > MinVLen)
      return false;
    MinVLen /= 8;
    break;
  case MVT::i8:
  case MVT::i16:
  case MVT::i32:
    break;
  case MVT::i64:
    if (!Subtarget.hasVInstructionsI64())
      return false;
    break;
  case MVT::f16:
    if (!Subtarget.hasVInstructionsF16())
      return false;
    break;
  case MVT::f32:
    if (!Subtarget.hasVInstructionsF32())
      return false;
    break;
  case MVT::f64:
    if (!Subtarget.hasVInstructionsF64())
      return false;
    break;
  }
 
  // Reject elements larger than ELEN.
  if (EltVT.getSizeInBits() > Subtarget.getMaxELENForFixedLengthVectors())
    return false;
 
  unsigned LMul = divideCeil(VT.getSizeInBits(), MinVLen);
  // Don't use RVV for types that don't fit.
  if (LMul > Subtarget.getMaxLMULForFixedLengthVectors())
    return false;
 
  // TODO: Perhaps an artificial restriction, but worth having whilst getting
  // the base fixed length RVV support in place.
  if (!VT.isPow2VectorType())
    return false;
 
  return true;
}
 
bool RISCVTargetLowering::useRVVForFixedLengthVectorVT(MVT VT) const {
  return ::useRVVForFixedLengthVectorVT(VT, Subtarget);
}
 
// Return the largest legal scalable vector type that matches VT's element type.
static MVT getContainerForFixedLengthVector(const TargetLowering &TLI, MVT VT,
                                            const RISCVSubtarget &Subtarget) {
  // This may be called before legal types are setup.
  assert(((VT.isFixedLengthVector() && TLI.isTypeLegal(VT)) ||
          useRVVForFixedLengthVectorVT(VT, Subtarget)) &&
         "Expected legal fixed length vector!");
 
  unsigned MinVLen = Subtarget.getMinRVVVectorSizeInBits();
  unsigned MaxELen = Subtarget.getMaxELENForFixedLengthVectors();
 
  MVT EltVT = VT.getVectorElementType();
  switch (EltVT.SimpleTy) {
  default:
    llvm_unreachable("unexpected element type for RVV container");
  case MVT::i1:
  case MVT::i8:
  case MVT::i16:
  case MVT::i32:
  case MVT::i64:
  case MVT::f16:
  case MVT::f32:
  case MVT::f64: {
    // We prefer to use LMUL=1 for VLEN sized types. Use fractional lmuls for
    // narrower types. The smallest fractional LMUL we support is 8/ELEN. Within
    // each fractional LMUL we support SEW between 8 and LMUL*ELEN.
    unsigned NumElts =
        (VT.getVectorNumElements() * RISCV::RVVBitsPerBlock) / MinVLen;
    NumElts = std::max(NumElts, RISCV::RVVBitsPerBlock / MaxELen);
    assert(isPowerOf2_32(NumElts) && "Expected power of 2 NumElts");
    return MVT::getScalableVectorVT(EltVT, NumElts);
  }
  }
}
 
static MVT getContainerForFixedLengthVector(SelectionDAG &DAG, MVT VT,
                                            const RISCVSubtarget &Subtarget) {
  return getContainerForFixedLengthVector(DAG.getTargetLoweringInfo(), VT,
                                          Subtarget);
}
 
MVT RISCVTargetLowering::getContainerForFixedLengthVector(MVT VT) const {
  return ::getContainerForFixedLengthVector(*this, VT, getSubtarget());
}
 
// Grow V to consume an entire RVV register.
static SDValue convertToScalableVector(EVT VT, SDValue V, SelectionDAG &DAG,
                                       const RISCVSubtarget &Subtarget) {
  assert(VT.isScalableVector() &&
         "Expected to convert into a scalable vector!");
  assert(V.getValueType().isFixedLengthVector() &&
         "Expected a fixed length vector operand!");
  SDLoc DL(V);
  SDValue Zero = DAG.getConstant(0, DL, Subtarget.getXLenVT());
  return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), V, Zero);
}
 
// Shrink V so it's just big enough to maintain a VT's worth of data.
static SDValue convertFromScalableVector(EVT VT, SDValue V, SelectionDAG &DAG,
                                         const RISCVSubtarget &Subtarget) {
  assert(VT.isFixedLengthVector() &&
         "Expected to convert into a fixed length vector!");
  assert(V.getValueType().isScalableVector() &&
         "Expected a scalable vector operand!");
  SDLoc DL(V);
  SDValue Zero = DAG.getConstant(0, DL, Subtarget.getXLenVT());
  return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V, Zero);
}
 
// Gets the two common "VL" operands: an all-ones mask and the vector length.
// VecVT is a vector type, either fixed-length or scalable, and ContainerVT is
// the vector type that it is contained in.
static std::pair<SDValue, SDValue>
getDefaultVLOps(MVT VecVT, MVT ContainerVT, SDLoc DL, SelectionDAG &DAG,
                const RISCVSubtarget &Subtarget) {
  assert(ContainerVT.isScalableVector() && "Expecting scalable container type");
  MVT XLenVT = Subtarget.getXLenVT();
  SDValue VL = VecVT.isFixedLengthVector()
                   ? DAG.getConstant(VecVT.getVectorNumElements(), DL, XLenVT)
                   : DAG.getTargetConstant(RISCV::VLMaxSentinel, DL, XLenVT);
  MVT MaskVT = MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount());
  SDValue Mask = DAG.getNode(RISCVISD::VMSET_VL, DL, MaskVT, VL);
  return {Mask, VL};
}
 
// As above but assuming the given type is a scalable vector type.
static std::pair<SDValue, SDValue>
getDefaultScalableVLOps(MVT VecVT, SDLoc DL, SelectionDAG &DAG,
                        const RISCVSubtarget &Subtarget) {
  assert(VecVT.isScalableVector() && "Expecting a scalable vector");
  return getDefaultVLOps(VecVT, VecVT, DL, DAG, Subtarget);
}
 
// The state of RVV BUILD_VECTOR and VECTOR_SHUFFLE lowering is that very few
// of either is (currently) supported. This can get us into an infinite loop
// where we try to lower a BUILD_VECTOR as a VECTOR_SHUFFLE as a BUILD_VECTOR
// as a ..., etc.
// Until either (or both) of these can reliably lower any node, reporting that
// we don't want to expand BUILD_VECTORs via VECTOR_SHUFFLEs at least breaks
// the infinite loop. Note that this lowers BUILD_VECTOR through the stack,
// which is not desirable.
bool RISCVTargetLowering::shouldExpandBuildVectorWithShuffles(
    EVT VT, unsigned DefinedValues) const {
  return false;
}
 
bool RISCVTargetLowering::isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const {
  // Only splats are currently supported.
  if (ShuffleVectorSDNode::isSplatMask(M.data(), VT))
    return true;
 
  return false;
}
 
static SDValue lowerFP_TO_INT_SAT(SDValue Op, SelectionDAG &DAG,
                                  const RISCVSubtarget &Subtarget) {
  // RISCV FP-to-int conversions saturate to the destination register size, but
  // don't produce 0 for nan. We can use a conversion instruction and fix the
  // nan case with a compare and a select.
  SDValue Src = Op.getOperand(0);
 
  EVT DstVT = Op.getValueType();
  EVT SatVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
 
  bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT_SAT;
  unsigned Opc;
  if (SatVT == DstVT)
    Opc = IsSigned ? RISCVISD::FCVT_X : RISCVISD::FCVT_XU;
  else if (DstVT == MVT::i64 && SatVT == MVT::i32)
    Opc = IsSigned ? RISCVISD::FCVT_W_RV64 : RISCVISD::FCVT_WU_RV64;
  else
    return SDValue();
  // FIXME: Support other SatVTs by clamping before or after the conversion.
 
  SDLoc DL(Op);
  SDValue FpToInt = DAG.getNode(
      Opc, DL, DstVT, Src,
      DAG.getTargetConstant(RISCVFPRndMode::RTZ, DL, Subtarget.getXLenVT()));
 
  SDValue ZeroInt = DAG.getConstant(0, DL, DstVT);
  return DAG.getSelectCC(DL, Src, Src, ZeroInt, FpToInt, ISD::CondCode::SETUO);
}
 
// Expand vector FTRUNC, FCEIL, and FFLOOR by converting to the integer domain
// and back. Taking care to avoid converting values that are nan or already
// correct.
// TODO: Floor and ceil could be shorter by changing rounding mode, but we don't
// have FRM dependencies modeled yet.
static SDValue lowerFTRUNC_FCEIL_FFLOOR(SDValue Op, SelectionDAG &DAG) {
  MVT VT = Op.getSimpleValueType();
  assert(VT.isVector() && "Unexpected type");
 
  SDLoc DL(Op);
 
  // Freeze the source since we are increasing the number of uses.
  SDValue Src = DAG.getNode(ISD::FREEZE, DL, VT, Op.getOperand(0));
 
  // Truncate to integer and convert back to FP.
  MVT IntVT = VT.changeVectorElementTypeToInteger();
  SDValue Truncated = DAG.getNode(ISD::FP_TO_SINT, DL, IntVT, Src);
  Truncated = DAG.getNode(ISD::SINT_TO_FP, DL, VT, Truncated);
 
  MVT SetccVT = MVT::getVectorVT(MVT::i1, VT.getVectorElementCount());
 
  if (Op.getOpcode() == ISD::FCEIL) {
    // If the truncated value is the greater than or equal to the original
    // value, we've computed the ceil. Otherwise, we went the wrong way and
    // need to increase by 1.
    // FIXME: This should use a masked operation. Handle here or in isel?
    SDValue Adjust = DAG.getNode(ISD::FADD, DL, VT, Truncated,
                                 DAG.getConstantFP(1.0, DL, VT));
    SDValue NeedAdjust = DAG.getSetCC(DL, SetccVT, Truncated, Src, ISD::SETOLT);
    Truncated = DAG.getSelect(DL, VT, NeedAdjust, Adjust, Truncated);
  } else if (Op.getOpcode() == ISD::FFLOOR) {
    // If the truncated value is the less than or equal to the original value,
    // we've computed the floor. Otherwise, we went the wrong way and need to
    // decrease by 1.
    // FIXME: This should use a masked operation. Handle here or in isel?
    SDValue Adjust = DAG.getNode(ISD::FSUB, DL, VT, Truncated,
                                 DAG.getConstantFP(1.0, DL, VT));
    SDValue NeedAdjust = DAG.getSetCC(DL, SetccVT, Truncated, Src, ISD::SETOGT);
    Truncated = DAG.getSelect(DL, VT, NeedAdjust, Adjust, Truncated);
  }
 
  // Restore the original sign so that -0.0 is preserved.
  Truncated = DAG.getNode(ISD::FCOPYSIGN, DL, VT, Truncated, Src);
 
  // Determine the largest integer that can be represented exactly. This and
  // values larger than it don't have any fractional bits so don't need to
  // be converted.
  const fltSemantics &FltSem = DAG.EVTToAPFloatSemantics(VT);
  unsigned Precision = APFloat::semanticsPrecision(FltSem);
  APFloat MaxVal = APFloat(FltSem);
  MaxVal.convertFromAPInt(APInt::getOneBitSet(Precision, Precision - 1),
                          /*IsSigned*/ false, APFloat::rmNearestTiesToEven);
  SDValue MaxValNode = DAG.getConstantFP(MaxVal, DL, VT);
 
  // If abs(Src) was larger than MaxVal or nan, keep it.
  SDValue Abs = DAG.getNode(ISD::FABS, DL, VT, Src);
  SDValue Setcc = DAG.getSetCC(DL, SetccVT, Abs, MaxValNode, ISD::SETOLT);
  return DAG.getSelect(DL, VT, Setcc, Truncated, Src);
}
 
static SDValue lowerSPLAT_VECTOR(SDValue Op, SelectionDAG &DAG,
                                 const RISCVSubtarget &Subtarget) {
  MVT VT = Op.getSimpleValueType();
  assert(VT.isFixedLengthVector() && "Unexpected vector!");
 
  MVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
 
  SDLoc DL(Op);
  SDValue Mask, VL;
  std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
 
  unsigned Opc =
      VT.isFloatingPoint() ? RISCVISD::VFMV_V_F_VL : RISCVISD::VMV_V_X_VL;
  SDValue Splat = DAG.getNode(Opc, DL, ContainerVT, Op.getOperand(0), VL);
  return convertFromScalableVector(VT, Splat, DAG, Subtarget);
}
 
struct VIDSequence {
  int64_t StepNumerator;
  unsigned StepDenominator;
  int64_t Addend;
};
 
// Try to match an arithmetic-sequence BUILD_VECTOR [X,X+S,X+2*S,...,X+(N-1)*S]
// to the (non-zero) step S and start value X. This can be then lowered as the
// RVV sequence (VID * S) + X, for example.
// The step S is represented as an integer numerator divided by a positive
// denominator. Note that the implementation currently only identifies
// sequences in which either the numerator is +/- 1 or the denominator is 1. It
// cannot detect 2/3, for example.
// Note that this method will also match potentially unappealing index
// sequences, like <i32 0, i32 50939494>, however it is left to the caller to
// determine whether this is worth generating code for.
static Optional<VIDSequence> isSimpleVIDSequence(SDValue Op) {
  unsigned NumElts = Op.getNumOperands();
  assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unexpected BUILD_VECTOR");
  if (!Op.getValueType().isInteger())
    return None;
 
  Optional<unsigned> SeqStepDenom;
  Optional<int64_t> SeqStepNum, SeqAddend;
  Optional<std::pair<uint64_t, unsigned>> PrevElt;
  unsigned EltSizeInBits = Op.getValueType().getScalarSizeInBits();
  for (unsigned Idx = 0; Idx < NumElts; Idx++) {
    // Assume undef elements match the sequence; we just have to be careful
    // when interpolating across them.
    if (Op.getOperand(Idx).isUndef())
      continue;
    // The BUILD_VECTOR must be all constants.
    if (!isa<ConstantSDNode>(Op.getOperand(Idx)))
      return None;
 
    uint64_t Val = Op.getConstantOperandVal(Idx) &
                   maskTrailingOnes<uint64_t>(EltSizeInBits);
 
    if (PrevElt) {
      // Calculate the step since the last non-undef element, and ensure
      // it's consistent across the entire sequence.
      unsigned IdxDiff = Idx - PrevElt->second;
      int64_t ValDiff = SignExtend64(Val - PrevElt->first, EltSizeInBits);
 
      // A zero-value value difference means that we're somewhere in the middle
      // of a fractional step, e.g. <0,0,0*,0,1,1,1,1>. Wait until we notice a
      // step change before evaluating the sequence.
      if (ValDiff != 0) {
        int64_t Remainder = ValDiff % IdxDiff;
        // Normalize the step if it's greater than 1.
        if (Remainder != ValDiff) {
          // The difference must cleanly divide the element span.
          if (Remainder != 0)
            return None;
          ValDiff /= IdxDiff;
          IdxDiff = 1;
        }
 
        if (!SeqStepNum)
          SeqStepNum = ValDiff;
        else if (ValDiff != SeqStepNum)
          return None;
 
        if (!SeqStepDenom)
          SeqStepDenom = IdxDiff;
        else if (IdxDiff != *SeqStepDenom)
          return None;
      }
    }
 
    // Record and/or check any addend.
    if (SeqStepNum && SeqStepDenom) {
      uint64_t ExpectedVal =
          (int64_t)(Idx * (uint64_t)*SeqStepNum) / *SeqStepDenom;
      int64_t Addend = SignExtend64(Val - ExpectedVal, EltSizeInBits);
      if (!SeqAddend)
        SeqAddend = Addend;
      else if (SeqAddend != Addend)
        return None;
    }
 
    // Record this non-undef element for later.
    if (!PrevElt || PrevElt->first != Val)
      PrevElt = std::make_pair(Val, Idx);
  }
  // We need to have logged both a step and an addend for this to count as
  // a legal index sequence.
  if (!SeqStepNum || !SeqStepDenom || !SeqAddend)
    return None;
 
  return VIDSequence{*SeqStepNum, *SeqStepDenom, *SeqAddend};
}
 
static SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
                                 const RISCVSubtarget &Subtarget) {
  MVT VT = Op.getSimpleValueType();
  assert(VT.isFixedLengthVector() && "Unexpected vector!");
 
  MVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
 
  SDLoc DL(Op);
  SDValue Mask, VL;
  std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
 
  MVT XLenVT = Subtarget.getXLenVT();
  unsigned NumElts = Op.getNumOperands();
 
  if (VT.getVectorElementType() == MVT::i1) {
    if (ISD::isBuildVectorAllZeros(Op.getNode())) {
      SDValue VMClr = DAG.getNode(RISCVISD::VMCLR_VL, DL, ContainerVT, VL);
      return convertFromScalableVector(VT, VMClr, DAG, Subtarget);
    }
 
    if (ISD::isBuildVectorAllOnes(Op.getNode())) {
      SDValue VMSet = DAG.getNode(RISCVISD::VMSET_VL, DL, ContainerVT, VL);
      return convertFromScalableVector(VT, VMSet, DAG, Subtarget);
    }
 
    // Lower constant mask BUILD_VECTORs via an integer vector type, in
    // scalar integer chunks whose bit-width depends on the number of mask
    // bits and XLEN.
    // First, determine the most appropriate scalar integer type to use. This
    // is at most XLenVT, but may be shrunk to a smaller vector element type
    // according to the size of the final vector - use i8 chunks rather than
    // XLenVT if we're producing a v8i1. This results in more consistent
    // codegen across RV32 and RV64.
    unsigned NumViaIntegerBits =
        std::min(std::max(NumElts, 8u), Subtarget.getXLen());
    NumViaIntegerBits = std::min(NumViaIntegerBits,
                                 Subtarget.getMaxELENForFixedLengthVectors());
    if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
      // If we have to use more than one INSERT_VECTOR_ELT then this
      // optimization is likely to increase code size; avoid peforming it in
      // such a case. We can use a load from a constant pool in this case.
      if (DAG.shouldOptForSize() && NumElts > NumViaIntegerBits)
        return SDValue();
      // Now we can create our integer vector type. Note that it may be larger
      // than the resulting mask type: v4i1 would use v1i8 as its integer type.
      MVT IntegerViaVecVT =
          MVT::getVectorVT(MVT::getIntegerVT(NumViaIntegerBits),
                           divideCeil(NumElts, NumViaIntegerBits));
 
      uint64_t Bits = 0;
      unsigned BitPos = 0, IntegerEltIdx = 0;
      SDValue Vec = DAG.getUNDEF(IntegerViaVecVT);
 
      for (unsigned I = 0; I < NumElts; I++, BitPos++) {
        // Once we accumulate enough bits to fill our scalar type, insert into
        // our vector and clear our accumulated data.
        if (I != 0 && I % NumViaIntegerBits == 0) {
          if (NumViaIntegerBits <= 32)
            Bits = SignExtend64(Bits, 32);
          SDValue Elt = DAG.getConstant(Bits, DL, XLenVT);
          Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, IntegerViaVecVT, Vec,
                            Elt, DAG.getConstant(IntegerEltIdx, DL, XLenVT));
          Bits = 0;
          BitPos = 0;
          IntegerEltIdx++;
        }
        SDValue V = Op.getOperand(I);
        bool BitValue = !V.isUndef() && cast<ConstantSDNode>(V)->getZExtValue();
        Bits |= ((uint64_t)BitValue << BitPos);
      }
 
      // Insert the (remaining) scalar value into position in our integer
      // vector type.
      if (NumViaIntegerBits <= 32)
        Bits = SignExtend64(Bits, 32);
      SDValue Elt = DAG.getConstant(Bits, DL, XLenVT);
      Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, IntegerViaVecVT, Vec, Elt,
                        DAG.getConstant(IntegerEltIdx, DL, XLenVT));
 
      if (NumElts < NumViaIntegerBits) {
        // If we're producing a smaller vector than our minimum legal integer
        // type, bitcast to the equivalent (known-legal) mask type, and extract
        // our final mask.
        assert(IntegerViaVecVT == MVT::v1i8 && "Unexpected mask vector type");
        Vec = DAG.getBitcast(MVT::v8i1, Vec);
        Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Vec,
                          DAG.getConstant(0, DL, XLenVT));
      } else {
        // Else we must have produced an integer type with the same size as the
        // mask type; bitcast for the final result.
        assert(VT.getSizeInBits() == IntegerViaVecVT.getSizeInBits());
        Vec = DAG.getBitcast(VT, Vec);
      }
 
      return Vec;
    }
 
    // A BUILD_VECTOR can be lowered as a SETCC. For each fixed-length mask
    // vector type, we have a legal equivalently-sized i8 type, so we can use
    // that.
    MVT WideVecVT = VT.changeVectorElementType(MVT::i8);
    SDValue VecZero = DAG.getConstant(0, DL, WideVecVT);
 
    SDValue WideVec;
    if (SDValue Splat = cast<BuildVectorSDNode>(Op)->getSplatValue()) {
      // For a splat, perform a scalar truncate before creating the wider
      // vector.
      assert(Splat.getValueType() == XLenVT &&
             "Unexpected type for i1 splat value");
      Splat = DAG.getNode(ISD::AND, DL, XLenVT, Splat,
                          DAG.getConstant(1, DL, XLenVT));
      WideVec = DAG.getSplatBuildVector(WideVecVT, DL, Splat);
    } else {
      SmallVector<SDValue, 8> Ops(Op->op_values());
      WideVec = DAG.getBuildVector(WideVecVT, DL, Ops);
      SDValue VecOne = DAG.getConstant(1, DL, WideVecVT);
      WideVec = DAG.getNode(ISD::AND, DL, WideVecVT, WideVec, VecOne);
    }
 
    return DAG.getSetCC(DL, VT, WideVec, VecZero, ISD::SETNE);
  }
 
  if (SDValue Splat = cast<BuildVectorSDNode>(Op)->getSplatValue()) {
    unsigned Opc = VT.isFloatingPoint() ? RISCVISD::VFMV_V_F_VL
                                        : RISCVISD::VMV_V_X_VL;
    Splat = DAG.getNode(Opc, DL, ContainerVT, Splat, VL);
    return convertFromScalableVector(VT, Splat, DAG, Subtarget);
  }
 
  // Try and match index sequences, which we can lower to the vid instruction
  // with optional modifications. An all-undef vector is matched by
  // getSplatValue, above.
  if (auto SimpleVID = isSimpleVIDSequence(Op)) {
    int64_t StepNumerator = SimpleVID->StepNumerator;
    unsigned StepDenominator = SimpleVID->StepDenominator;
    int64_t Addend = SimpleVID->Addend;
 
    assert(StepNumerator != 0 && "Invalid step");
    bool Negate = false;
    int64_t SplatStepVal = StepNumerator;
    unsigned StepOpcode = ISD::MUL;
    if (StepNumerator != 1) {
      if (isPowerOf2_64(std::abs(StepNumerator))) {
        Negate = StepNumerator < 0;
        StepOpcode = ISD::SHL;
        SplatStepVal = Log2_64(std::abs(StepNumerator));
      }
    }
 
    // Only emit VIDs with suitably-small steps/addends. We use imm5 is a
    // threshold since it's the immediate value many RVV instructions accept.
    // There is no vmul.vi instruction so ensure multiply constant can fit in
    // a single addi instruction.
    if (((StepOpcode == ISD::MUL && isInt<12>(SplatStepVal)) ||
         (StepOpcode == ISD::SHL && isUInt<5>(SplatStepVal))) &&
        isPowerOf2_32(StepDenominator) && isInt<5>(Addend)) {
      SDValue VID = DAG.getNode(RISCVISD::VID_VL, DL, ContainerVT, Mask, VL);
      // Convert right out of the scalable type so we can use standard ISD
      // nodes for the rest of the computation. If we used scalable types with
      // these, we'd lose the fixed-length vector info and generate worse
      // vsetvli code.
      VID = convertFromScalableVector(VT, VID, DAG, Subtarget);
      if ((StepOpcode == ISD::MUL && SplatStepVal != 1) ||
          (StepOpcode == ISD::SHL && SplatStepVal != 0)) {
        SDValue SplatStep = DAG.getSplatVector(
            VT, DL, DAG.getConstant(SplatStepVal, DL, XLenVT));
        VID = DAG.getNode(StepOpcode, DL, VT, VID, SplatStep);
      }
      if (StepDenominator != 1) {
        SDValue SplatStep = DAG.getSplatVector(
            VT, DL, DAG.getConstant(Log2_64(StepDenominator), DL, XLenVT));
        VID = DAG.getNode(ISD::SRL, DL, VT, VID, SplatStep);
      }
      if (Addend != 0 || Negate) {
        SDValue SplatAddend =
            DAG.getSplatVector(VT, DL, DAG.getConstant(Addend, DL, XLenVT));
        VID = DAG.getNode(Negate ? ISD::SUB : ISD::ADD, DL, VT, SplatAddend, VID);
      }
      return VID;
    }
  }
 
  // Attempt to detect "hidden" splats, which only reveal themselves as splats
  // when re-interpreted as a vector with a larger element type. For example,
  //   v4i16 = build_vector i16 0, i16 1, i16 0, i16 1
  // could be instead splat as
  //   v2i32 = build_vector i32 0x00010000, i32 0x00010000
  // TODO: This optimization could also work on non-constant splats, but it
  // would require bit-manipulation instructions to construct the splat value.
  SmallVector<SDValue> Sequence;
  unsigned EltBitSize = VT.getScalarSizeInBits();
  const auto *BV = cast<BuildVectorSDNode>(Op);
  if (VT.isInteger() && EltBitSize < 64 &&
      ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
      BV->getRepeatedSequence(Sequence) &&
      (Sequence.size() * EltBitSize) <= 64) {
    unsigned SeqLen = Sequence.size();
    MVT ViaIntVT = MVT::getIntegerVT(EltBitSize * SeqLen);
    MVT ViaVecVT = MVT::getVectorVT(ViaIntVT, NumElts / SeqLen);
    assert((ViaIntVT == MVT::i16 || ViaIntVT == MVT::i32 ||
            ViaIntVT == MVT::i64) &&
           "Unexpected sequence type");
 
    unsigned EltIdx = 0;
    uint64_t EltMask = maskTrailingOnes<uint64_t>(EltBitSize);
    uint64_t SplatValue = 0;
    // Construct the amalgamated value which can be splatted as this larger
    // vector type.
    for (const auto &SeqV : Sequence) {
      if (!SeqV.isUndef())
        SplatValue |= ((cast<ConstantSDNode>(SeqV)->getZExtValue() & EltMask)
                       << (EltIdx * EltBitSize));
      EltIdx++;
    }
 
    // On RV64, sign-extend from 32 to 64 bits where possible in order to
    // achieve better constant materializion.
    if (Subtarget.is64Bit() && ViaIntVT == MVT::i32)
      SplatValue = SignExtend64(SplatValue, 32);
 
    // Since we can't introduce illegal i64 types at this stage, we can only
    // perform an i64 splat on RV32 if it is its own sign-extended value. That
    // way we can use RVV instructions to splat.
    assert((ViaIntVT.bitsLE(XLenVT) ||
            (!Subtarget.is64Bit() && ViaIntVT == MVT::i64)) &&
           "Unexpected bitcast sequence");
    if (ViaIntVT.bitsLE(XLenVT) || isInt<32>(SplatValue)) {
      SDValue ViaVL =
          DAG.getConstant(ViaVecVT.getVectorNumElements(), DL, XLenVT);
      MVT ViaContainerVT =
          getContainerForFixedLengthVector(DAG, ViaVecVT, Subtarget);
      SDValue Splat =
          DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ViaContainerVT,
                      DAG.getConstant(SplatValue, DL, XLenVT), ViaVL);
      Splat = convertFromScalableVector(ViaVecVT, Splat, DAG, Subtarget);
      return DAG.getBitcast(VT, Splat);
    }
  }
 
  // Try and optimize BUILD_VECTORs with "dominant values" - these are values
  // which constitute a large proportion of the elements. In such cases we can
  // splat a vector with the dominant element and make up the shortfall with
  // INSERT_VECTOR_ELTs.
  // Note that this includes vectors of 2 elements by association. The
  // upper-most element is the "dominant" one, allowing us to use a splat to
  // "insert" the upper element, and an insert of the lower element at position
  // 0, which improves codegen.
  SDValue DominantValue;
  unsigned MostCommonCount = 0;
  DenseMap<SDValue, unsigned> ValueCounts;
  unsigned NumUndefElts =
      count_if(Op->op_values(), [](const SDValue &V) { return V.isUndef(); });
 
  // Track the number of scalar loads we know we'd be inserting, estimated as
  // any non-zero floating-point constant. Other kinds of element are either
  // already in registers or are materialized on demand. The threshold at which
  // a vector load is more desirable than several scalar materializion and
  // vector-insertion instructions is not known.
  unsigned NumScalarLoads = 0;
 
  for (SDValue V : Op->op_values()) {
    if (V.isUndef())
      continue;
 
    ValueCounts.insert(std::make_pair(V, 0));
    unsigned &Count = ValueCounts[V];
 
    if (auto *CFP = dyn_cast<ConstantFPSDNode>(V))
      NumScalarLoads += !CFP->isExactlyValue(+0.0);
 
    // Is this value dominant? In case of a tie, prefer the highest element as
    // it's cheaper to insert near the beginning of a vector than it is at the
    // end.
    if (++Count >= MostCommonCount) {
      DominantValue = V;
      MostCommonCount = Count;
    }
  }
 
  assert(DominantValue && "Not expecting an all-undef BUILD_VECTOR");
  unsigned NumDefElts = NumElts - NumUndefElts;
  unsigned DominantValueCountThreshold = NumDefElts <= 2 ? 0 : NumDefElts - 2;
 
  // Don't perform this optimization when optimizing for size, since
  // materializing elements and inserting them tends to cause code bloat.
  if (!DAG.shouldOptForSize() && NumScalarLoads < NumElts &&
      ((MostCommonCount > DominantValueCountThreshold) ||
       (ValueCounts.size() <= Log2_32(NumDefElts)))) {
    // Start by splatting the most common element.
    SDValue Vec = DAG.getSplatBuildVector(VT, DL, DominantValue);
 
    DenseSet<SDValue> Processed{DominantValue};
    MVT SelMaskTy = VT.changeVectorElementType(MVT::i1);
    for (const auto &OpIdx : enumerate(Op->ops())) {
      const SDValue &V = OpIdx.value();
      if (V.isUndef() || !Processed.insert(V).second)
        continue;
      if (ValueCounts[V] == 1) {
        Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, Vec, V,
                          DAG.getConstant(OpIdx.index(), DL, XLenVT));
      } else {
        // Blend in all instances of this value using a VSELECT, using a
        // mask where each bit signals whether that element is the one
        // we're after.
        SmallVector<SDValue> Ops;
        transform(Op->op_values(), std::back_inserter(Ops), [&](SDValue V1) {
          return DAG.getConstant(V == V1, DL, XLenVT);
        });
        Vec = DAG.getNode(ISD::VSELECT, DL, VT,
                          DAG.getBuildVector(SelMaskTy, DL, Ops),
                          DAG.getSplatBuildVector(VT, DL, V), Vec);
      }
    }
 
    return Vec;
  }
 
  return SDValue();
}
 
static SDValue splatPartsI64WithVL(const SDLoc &DL, MVT VT, SDValue Lo,
                                   SDValue Hi, SDValue VL, SelectionDAG &DAG) {
  if (isa<ConstantSDNode>(Lo) && isa<ConstantSDNode>(Hi)) {
    int32_t LoC = cast<ConstantSDNode>(Lo)->getSExtValue();
    int32_t HiC = cast<ConstantSDNode>(Hi)->getSExtValue();
    // If Hi constant is all the same sign bit as Lo, lower this as a custom
    // node in order to try and match RVV vector/scalar instructions.
    if ((LoC >> 31) == HiC)
      return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, Lo, VL);
 
    // If vl is equal to VLMax and Hi constant is equal to Lo, we could use
    // vmv.v.x whose EEW = 32 to lower it.
    auto *Const = dyn_cast<ConstantSDNode>(VL);
    if (LoC == HiC && Const && Const->getSExtValue() == RISCV::VLMaxSentinel) {
      MVT InterVT = MVT::getVectorVT(MVT::i32, VT.getVectorElementCount() * 2);
      // TODO: if vl <= min(VLMAX), we can also do this. But we could not
      // access the subtarget here now.
      auto InterVec = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, InterVT, Lo, VL);
      return DAG.getNode(ISD::BITCAST, DL, VT, InterVec);
    }
  }
 
  // Fall back to a stack store and stride x0 vector load.
  return DAG.getNode(RISCVISD::SPLAT_VECTOR_SPLIT_I64_VL, DL, VT, Lo, Hi, VL);
}
 
// Called by type legalization to handle splat of i64 on RV32.
// FIXME: We can optimize this when the type has sign or zero bits in one
// of the halves.
static SDValue splatSplitI64WithVL(const SDLoc &DL, MVT VT, SDValue Scalar,
                                   SDValue VL, SelectionDAG &DAG) {
  assert(Scalar.getValueType() == MVT::i64 && "Unexpected VT!");
  SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Scalar,
                           DAG.getConstant(0, DL, MVT::i32));
  SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, Scalar,
                           DAG.getConstant(1, DL, MVT::i32));
  return splatPartsI64WithVL(DL, VT, Lo, Hi, VL, DAG);
}
 
// This function lowers a splat of a scalar operand Splat with the vector
// length VL. It ensures the final sequence is type legal, which is useful when
// lowering a splat after type legalization.
static SDValue lowerScalarSplat(SDValue Scalar, SDValue VL, MVT VT, SDLoc DL,
                                SelectionDAG &DAG,
                                const RISCVSubtarget &Subtarget) {
  if (VT.isFloatingPoint()) {
    // If VL is 1, we could use vfmv.s.f.
    if (isOneConstant(VL))
      return DAG.getNode(RISCVISD::VFMV_S_F_VL, DL, VT, DAG.getUNDEF(VT),
                         Scalar, VL);
    return DAG.getNode(RISCVISD::VFMV_V_F_VL, DL, VT, Scalar, VL);
  }
 
  MVT XLenVT = Subtarget.getXLenVT();
 
  // Simplest case is that the operand needs to be promoted to XLenVT.
  if (Scalar.getValueType().bitsLE(XLenVT)) {
    // If the operand is a constant, sign extend to increase our chances
    // of being able to use a .vi instruction. ANY_EXTEND would become a
    // a zero extend and the simm5 check in isel would fail.
    // FIXME: Should we ignore the upper bits in isel instead?
    unsigned ExtOpc =
        isa<ConstantSDNode>(Scalar) ? ISD::SIGN_EXTEND : ISD::ANY_EXTEND;
    Scalar = DAG.getNode(ExtOpc, DL, XLenVT, Scalar);
    ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Scalar);
    // If VL is 1 and the scalar value won't benefit from immediate, we could
    // use vmv.s.x.
    if (isOneConstant(VL) &&
        (!Const || isNullConstant(Scalar) || !isInt<5>(Const->getSExtValue())))
      return DAG.getNode(RISCVISD::VMV_S_X_VL, DL, VT, DAG.getUNDEF(VT), Scalar,
                         VL);
    return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, Scalar, VL);
  }
 
  assert(XLenVT == MVT::i32 && Scalar.getValueType() == MVT::i64 &&
         "Unexpected scalar for splat lowering!");
 
  if (isOneConstant(VL) && isNullConstant(Scalar))
    return DAG.getNode(RISCVISD::VMV_S_X_VL, DL, VT, DAG.getUNDEF(VT),
                       DAG.getConstant(0, DL, XLenVT), VL);
 
  // Otherwise use the more complicated splatting algorithm.
  return splatSplitI64WithVL(DL, VT, Scalar, VL, DAG);
}
 
// Is the mask a slidedown that shifts in undefs.
static int matchShuffleAsSlideDown(ArrayRef<int> Mask) {
  int Size = Mask.size();
 
  // Elements shifted in should be undef.
  auto CheckUndefs = [&](int Shift) {
    for (int i = Size - Shift; i != Size; ++i)
      if (Mask[i] >= 0)
        return false;
    return true;
  };
 
  // Elements should be shifted or undef.
  auto MatchShift = [&](int Shift) {
    for (int i = 0; i != Size - Shift; ++i)
       if (Mask[i] >= 0 && Mask[i] != Shift + i)
         return false;
    return true;
  };
 
  // Try all possible shifts.
  for (int Shift = 1; Shift != Size; ++Shift)
    if (CheckUndefs(Shift) && MatchShift(Shift))
      return Shift;
 
  // No match.
  return -1;
}
 
static bool isInterleaveShuffle(ArrayRef<int> Mask, MVT VT, bool &SwapSources,
                                const RISCVSubtarget &Subtarget) {
  // We need to be able to widen elements to the next larger integer type.
  if (VT.getScalarSizeInBits() >= Subtarget.getMaxELENForFixedLengthVectors())
    return false;
 
  int Size = Mask.size();
  assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
 
  int Srcs[] = {-1, -1};
  for (int i = 0; i != Size; ++i) {
    // Ignore undef elements.
    if (Mask[i] < 0)
      continue;
 
    // Is this an even or odd element.
    int Pol = i % 2;
 
    // Ensure we consistently use the same source for this element polarity.
    int Src = Mask[i] / Size;
    if (Srcs[Pol] < 0)
      Srcs[Pol] = Src;
    if (Srcs[Pol] != Src)
      return false;
 
    // Make sure the element within the source is appropriate for this element
    // in the destination.
    int Elt = Mask[i] % Size;
    if (Elt != i / 2)
      return false;
  }
 
  // We need to find a source for each polarity and they can't be the same.
  if (Srcs[0] < 0 || Srcs[1] < 0 || Srcs[0] == Srcs[1])
    return false;
 
  // Swap the sources if the second source was in the even polarity.
  SwapSources = Srcs[0] > Srcs[1];
 
  return true;
}
 
static SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG,
                                   const RISCVSubtarget &Subtarget) {
  SDValue V1 = Op.getOperand(0);
  SDValue V2 = Op.getOperand(1);
  SDLoc DL(Op);
  MVT XLenVT = Subtarget.getXLenVT();
  MVT VT = Op.getSimpleValueType();
  unsigned NumElts = VT.getVectorNumElements();
  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
 
  MVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
 
  SDValue TrueMask, VL;
  std::tie(TrueMask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
 
  if (SVN->isSplat()) {
    const int Lane = SVN->getSplatIndex();
    if (Lane >= 0) {
      MVT SVT = VT.getVectorElementType();
 
      // Turn splatted vector load into a strided load with an X0 stride.
      SDValue V = V1;
      // Peek through CONCAT_VECTORS as VectorCombine can concat a vector
      // with undef.
      // FIXME: Peek through INSERT_SUBVECTOR, EXTRACT_SUBVECTOR, bitcasts?
      int Offset = Lane;
      if (V.getOpcode() == ISD::CONCAT_VECTORS) {
        int OpElements =
            V.getOperand(0).getSimpleValueType().getVectorNumElements();
        V = V.getOperand(Offset / OpElements);
        Offset %= OpElements;
      }
 
      // We need to ensure the load isn't atomic or volatile.
      if (ISD::isNormalLoad(V.getNode()) && cast<LoadSDNode>(V)->isSimple()) {
        auto *Ld = cast<LoadSDNode>(V);
        Offset *= SVT.getStoreSize();
        SDValue NewAddr = DAG.getMemBasePlusOffset(Ld->getBasePtr(),
                                                   TypeSize::Fixed(Offset), DL);
 
        // If this is SEW=64 on RV32, use a strided load with a stride of x0.
        if (SVT.isInteger() && SVT.bitsGT(XLenVT)) {
          SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other});
          SDValue IntID =
              DAG.getTargetConstant(Intrinsic::riscv_vlse, DL, XLenVT);
          SDValue Ops[] = {Ld->getChain(),
                           IntID,
                           DAG.getUNDEF(ContainerVT),
                           NewAddr,
                           DAG.getRegister(RISCV::X0, XLenVT),
                           VL};
          SDValue NewLoad = DAG.getMemIntrinsicNode(
              ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops, SVT,
              DAG.getMachineFunction().getMachineMemOperand(
                  Ld->getMemOperand(), Offset, SVT.getStoreSize()));
          DAG.makeEquivalentMemoryOrdering(Ld, NewLoad);
          return convertFromScalableVector(VT, NewLoad, DAG, Subtarget);
        }
 
        // Otherwise use a scalar load and splat. This will give the best
        // opportunity to fold a splat into the operation. ISel can turn it into
        // the x0 strided load if we aren't able to fold away the select.
        if (SVT.isFloatingPoint())
          V = DAG.getLoad(SVT, DL, Ld->getChain(), NewAddr,
                          Ld->getPointerInfo().getWithOffset(Offset),
                          Ld->getOriginalAlign(),
                          Ld->getMemOperand()->getFlags());
        else
          V = DAG.getExtLoad(ISD::SEXTLOAD, DL, XLenVT, Ld->getChain(), NewAddr,
                             Ld->getPointerInfo().getWithOffset(Offset), SVT,
                             Ld->getOriginalAlign(),
                             Ld->getMemOperand()->getFlags());
        DAG.makeEquivalentMemoryOrdering(Ld, V);
 
        unsigned Opc =
            VT.isFloatingPoint() ? RISCVISD::VFMV_V_F_VL : RISCVISD::VMV_V_X_VL;
        SDValue Splat = DAG.getNode(Opc, DL, ContainerVT, V, VL);
        return convertFromScalableVector(VT, Splat, DAG, Subtarget);
      }
 
      V1 = convertToScalableVector(ContainerVT, V1, DAG, Subtarget);
      assert(Lane < (int)NumElts && "Unexpected lane!");
      SDValue Gather =
          DAG.getNode(RISCVISD::VRGATHER_VX_VL, DL, ContainerVT, V1,
                      DAG.getConstant(Lane, DL, XLenVT), TrueMask, VL);
      return convertFromScalableVector(VT, Gather, DAG, Subtarget);
    }
  }
 
  ArrayRef<int> Mask = SVN->getMask();
 
  // Try to match as a slidedown.
  int SlideAmt = matchShuffleAsSlideDown(Mask);
  if (SlideAmt >= 0) {
    // TODO: Should we reduce the VL to account for the upper undef elements?
    // Requires additional vsetvlis, but might be faster to execute.
    V1 = convertToScalableVector(ContainerVT, V1, DAG, Subtarget);
    SDValue SlideDown =
        DAG.getNode(RISCVISD::VSLIDEDOWN_VL, DL, ContainerVT,
                    DAG.getUNDEF(ContainerVT), V1,
                    DAG.getConstant(SlideAmt, DL, XLenVT),
                    TrueMask, VL);
    return convertFromScalableVector(VT, SlideDown, DAG, Subtarget);
  }
 
  // Detect an interleave shuffle and lower to
  // (vmaccu.vx (vwaddu.vx lohalf(V1), lohalf(V2)), lohalf(V2), (2^eltbits - 1))
  bool SwapSources;
  if (isInterleaveShuffle(Mask, VT, SwapSources, Subtarget)) {
    // Swap sources if needed.
    if (SwapSources)
      std::swap(V1, V2);
 
    // Extract the lower half of the vectors.
    MVT HalfVT = VT.getHalfNumVectorElementsVT();
    V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1,
                     DAG.getConstant(0, DL, XLenVT));
    V2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V2,
                     DAG.getConstant(0, DL, XLenVT));
 
    // Double the element width and halve the number of elements in an int type.
    unsigned EltBits = VT.getScalarSizeInBits();
    MVT WideIntEltVT = MVT::getIntegerVT(EltBits * 2);
    MVT WideIntVT =
        MVT::getVectorVT(WideIntEltVT, VT.getVectorNumElements() / 2);
    // Convert this to a scalable vector. We need to base this on the
    // destination size to ensure there's always a type with a smaller LMUL.
    MVT WideIntContainerVT =
        getContainerForFixedLengthVector(DAG, WideIntVT, Subtarget);
 
    // Convert sources to scalable vectors with the same element count as the
    // larger type.
    MVT HalfContainerVT = MVT::getVectorVT(
        VT.getVectorElementType(), WideIntContainerVT.getVectorElementCount());
    V1 = convertToScalableVector(HalfContainerVT, V1, DAG, Subtarget);
    V2 = convertToScalableVector(HalfContainerVT, V2, DAG, Subtarget);
 
    // Cast sources to integer.
    MVT IntEltVT = MVT::getIntegerVT(EltBits);
    MVT IntHalfVT =
        MVT::getVectorVT(IntEltVT, HalfContainerVT.getVectorElementCount());
    V1 = DAG.getBitcast(IntHalfVT, V1);
    V2 = DAG.getBitcast(IntHalfVT, V2);
 
    // Freeze V2 since we use it twice and we need to be sure that the add and
    // multiply see the same value.
    V2 = DAG.getNode(ISD::FREEZE, DL, IntHalfVT, V2);
 
    // Recreate TrueMask using the widened type's element count.
    MVT MaskVT =
        MVT::getVectorVT(MVT::i1, HalfContainerVT.getVectorElementCount());
    TrueMask = DAG.getNode(RISCVISD::VMSET_VL, DL, MaskVT, VL);
 
    // Widen V1 and V2 with 0s and add one copy of V2 to V1.
    SDValue Add = DAG.getNode(RISCVISD::VWADDU_VL, DL, WideIntContainerVT, V1,
                              V2, TrueMask, VL);
    // Create 2^eltbits - 1 copies of V2 by multiplying by the largest integer.
    SDValue Multiplier = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, IntHalfVT,
                                     DAG.getAllOnesConstant(DL, XLenVT));
    SDValue WidenMul = DAG.getNode(RISCVISD::VWMULU_VL, DL, WideIntContainerVT,
                                   V2, Multiplier, TrueMask, VL);
    // Add the new copies to our previous addition giving us 2^eltbits copies of
    // V2. This is equivalent to shifting V2 left by eltbits. This should
    // combine with the vwmulu.vv above to form vwmaccu.vv.
    Add = DAG.getNode(RISCVISD::ADD_VL, DL, WideIntContainerVT, Add, WidenMul,
                      TrueMask, VL);
    // Cast back to ContainerVT. We need to re-create a new ContainerVT in case
    // WideIntContainerVT is a larger fractional LMUL than implied by the fixed
    // vector VT.
    ContainerVT =
        MVT::getVectorVT(VT.getVectorElementType(),
                         WideIntContainerVT.getVectorElementCount() * 2);
    Add = DAG.getBitcast(ContainerVT, Add);
    return convertFromScalableVector(VT, Add, DAG, Subtarget);
  }
 
  // Detect shuffles which can be re-expressed as vector selects; these are
  // shuffles in which each element in the destination is taken from an element
  // at the corresponding index in either source vectors.
  bool IsSelect = all_of(enumerate(Mask), [&](const auto &MaskIdx) {
    int MaskIndex = MaskIdx.value();
    return MaskIndex < 0 || MaskIdx.index() == (unsigned)MaskIndex % NumElts;
  });
 
  assert(!V1.isUndef() && "Unexpected shuffle canonicalization");
 
  SmallVector<SDValue> MaskVals;
  // As a backup, shuffles can be lowered via a vrgather instruction, possibly
  // merged with a second vrgather.
  SmallVector<SDValue> GatherIndicesLHS, GatherIndicesRHS;
 
  // By default we preserve the original operand order, and use a mask to
  // select LHS as true and RHS as false. However, since RVV vector selects may
  // feature splats but only on the LHS, we may choose to invert our mask and
  // instead select between RHS and LHS.
  bool SwapOps = DAG.isSplatValue(V2) && !DAG.isSplatValue(V1);
  bool InvertMask = IsSelect == SwapOps;
 
  // Keep a track of which non-undef indices are used by each LHS/RHS shuffle
  // half.
  DenseMap<int, unsigned> LHSIndexCounts, RHSIndexCounts;
 
  // Now construct the mask that will be used by the vselect or blended
  // vrgather operation. For vrgathers, construct the appropriate indices into
  // each vector.
  for (int MaskIndex : Mask) {
    bool SelectMaskVal = (MaskIndex < (int)NumElts) ^ InvertMask;
    MaskVals.push_back(DAG.getConstant(SelectMaskVal, DL, XLenVT));
    if (!IsSelect) {
      bool IsLHSOrUndefIndex = MaskIndex < (int)NumElts;
      GatherIndicesLHS.push_back(IsLHSOrUndefIndex && MaskIndex >= 0
                                     ? DAG.getConstant(MaskIndex, DL, XLenVT)
                                     : DAG.getUNDEF(XLenVT));
      GatherIndicesRHS.push_back(
          IsLHSOrUndefIndex ? DAG.getUNDEF(XLenVT)
                            : DAG.getConstant(MaskIndex - NumElts, DL, XLenVT));
      if (IsLHSOrUndefIndex && MaskIndex >= 0)
        ++LHSIndexCounts[MaskIndex];
      if (!IsLHSOrUndefIndex)
        ++RHSIndexCounts[MaskIndex - NumElts];
    }
  }
 
  if (SwapOps) {
    std::swap(V1, V2);
    std::swap(GatherIndicesLHS, GatherIndicesRHS);
  }
 
  assert(MaskVals.size() == NumElts && "Unexpected select-like shuffle");
  MVT MaskVT = MVT::getVectorVT(MVT::i1, NumElts);
  SDValue SelectMask = DAG.getBuildVector(MaskVT, DL, MaskVals);
 
  if (IsSelect)
    return DAG.getNode(ISD::VSELECT, DL, VT, SelectMask, V1, V2);
 
  if (VT.getScalarSizeInBits() == 8 && VT.getVectorNumElements() > 256) {
    // On such a large vector we're unable to use i8 as the index type.
    // FIXME: We could promote the index to i16 and use vrgatherei16, but that
    // may involve vector splitting if we're already at LMUL=8, or our
    // user-supplied maximum fixed-length LMUL.
    return SDValue();
  }
 
  unsigned GatherVXOpc = RISCVISD::VRGATHER_VX_VL;
  unsigned GatherVVOpc = RISCVISD::VRGATHER_VV_VL;
  MVT IndexVT = VT.changeTypeToInteger();
  // Since we can't introduce illegal index types at this stage, use i16 and
  // vrgatherei16 if the corresponding index type for plain vrgather is greater
  // than XLenVT.
  if (IndexVT.getScalarType().bitsGT(XLenVT)) {
    GatherVVOpc = RISCVISD::VRGATHEREI16_VV_VL;
    IndexVT = IndexVT.changeVectorElementType(MVT::i16);
  }
 
  MVT IndexContainerVT =
      ContainerVT.changeVectorElementType(IndexVT.getScalarType());
 
  SDValue Gather;
  // TODO: This doesn't trigger for i64 vectors on RV32, since there we
  // encounter a bitcasted BUILD_VECTOR with low/high i32 values.
  if (SDValue SplatValue = DAG.getSplatValue(V1, /*LegalTypes*/ true)) {
    Gather = lowerScalarSplat(SplatValue, VL, ContainerVT, DL, DAG, Subtarget);
  } else {
    V1 = convertToScalableVector(ContainerVT, V1, DAG, Subtarget);
    // If only one index is used, we can use a "splat" vrgather.
    // TODO: We can splat the most-common index and fix-up any stragglers, if
    // that's beneficial.
    if (LHSIndexCounts.size() == 1) {
      int SplatIndex = LHSIndexCounts.begin()->getFirst();
      Gather =
          DAG.getNode(GatherVXOpc, DL, ContainerVT, V1,
                      DAG.getConstant(SplatIndex, DL, XLenVT), TrueMask, VL);
    } else {
      SDValue LHSIndices = DAG.getBuildVector(IndexVT, DL, GatherIndicesLHS);
      LHSIndices =
          convertToScalableVector(IndexContainerVT, LHSIndices, DAG, Subtarget);
 
      Gather = DAG.getNode(GatherVVOpc, DL, ContainerVT, V1, LHSIndices,
                           TrueMask, VL);
    }
  }
 
  // If a second vector operand is used by this shuffle, blend it in with an
  // additional vrgather.
  if (!V2.isUndef()) {
    V2 = convertToScalableVector(ContainerVT, V2, DAG, Subtarget);
    // If only one index is used, we can use a "splat" vrgather.
    // TODO: We can splat the most-common index and fix-up any stragglers, if
    // that's beneficial.
    if (RHSIndexCounts.size() == 1) {
      int SplatIndex = RHSIndexCounts.begin()->getFirst();
      V2 = DAG.getNode(GatherVXOpc, DL, ContainerVT, V2,
                       DAG.getConstant(SplatIndex, DL, XLenVT), TrueMask, VL);
    } else {
      SDValue RHSIndices = DAG.getBuildVector(IndexVT, DL, GatherIndicesRHS);
      RHSIndices =
          convertToScalableVector(IndexContainerVT, RHSIndices, DAG, Subtarget);
      V2 = DAG.getNode(GatherVVOpc, DL, ContainerVT, V2, RHSIndices, TrueMask,
                       VL);
    }
 
    MVT MaskContainerVT = ContainerVT.changeVectorElementType(MVT::i1);
    SelectMask =
        convertToScalableVector(MaskContainerVT, SelectMask, DAG, Subtarget);
 
    Gather = DAG.getNode(RISCVISD::VSELECT_VL, DL, ContainerVT, SelectMask, V2,
                         Gather, VL);
  }
 
  return convertFromScalableVector(VT, Gather, DAG, Subtarget);
}
 
static SDValue getRVVFPExtendOrRound(SDValue Op, MVT VT, MVT ContainerVT,
                                     SDLoc DL, SelectionDAG &DAG,
                                     const RISCVSubtarget &Subtarget) {
  if (VT.isScalableVector())
    return DAG.getFPExtendOrRound(Op, DL, VT);
  assert(VT.isFixedLengthVector() &&
         "Unexpected value type for RVV FP extend/round lowering");
  SDValue Mask, VL;
  std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
  unsigned RVVOpc = ContainerVT.bitsGT(Op.getSimpleValueType())
                        ? RISCVISD::FP_EXTEND_VL
                        : RISCVISD::FP_ROUND_VL;
  return DAG.getNode(RVVOpc, DL, ContainerVT, Op, Mask, VL);
}
 
// Lower CTLZ_ZERO_UNDEF or CTTZ_ZERO_UNDEF by converting to FP and extracting
// the exponent.
static SDValue lowerCTLZ_CTTZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
  MVT VT = Op.getSimpleValueType();
  unsigned EltSize = VT.getScalarSizeInBits();
  SDValue Src = Op.getOperand(0);
  SDLoc DL(Op);
 
  // We need a FP type that can represent the value.
  // TODO: Use f16 for i8 when possible?
  MVT FloatEltVT = EltSize == 32 ? MVT::f64 : MVT::f32;
  MVT FloatVT = MVT::getVectorVT(FloatEltVT, VT.getVectorElementCount());
 
  // Legal types should have been checked in the RISCVTargetLowering
  // constructor.
  // TODO: Splitting may make sense in some cases.
  assert(DAG.getTargetLoweringInfo().isTypeLegal(FloatVT) &&
         "Expected legal float type!");
 
  // For CTTZ_ZERO_UNDEF, we need to extract the lowest set bit using X & -X.
  // The trailing zero count is equal to log2 of this single bit value.
  if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF) {
    SDValue Neg =
        DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Src);
    Src = DAG.getNode(ISD::AND, DL, VT, Src, Neg);
  }
 
  // We have a legal FP type, convert to it.
  SDValue FloatVal = DAG.getNode(ISD::UINT_TO_FP, DL, FloatVT, Src);
  // Bitcast to integer and shift the exponent to the LSB.
  EVT IntVT = FloatVT.changeVectorElementTypeToInteger();
  SDValue Bitcast = DAG.getBitcast(IntVT, FloatVal);
  unsigned ShiftAmt = FloatEltVT == MVT::f64 ? 52 : 23;
  SDValue Shift = DAG.getNode(ISD::SRL, DL, IntVT, Bitcast,
                              DAG.getConstant(ShiftAmt, DL, IntVT));
  // Truncate back to original type to allow vnsrl.
  SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, VT, Shift);
  // The exponent contains log2 of the value in biased form.
  unsigned ExponentBias = FloatEltVT == MVT::f64 ? 1023 : 127;
 
  // For trailing zeros, we just need to subtract the bias.
  if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF)
    return DAG.getNode(ISD::SUB, DL, VT, Trunc,
                       DAG.getConstant(ExponentBias, DL, VT));
 
  // For leading zeros, we need to remove the bias and convert from log2 to
  // leading zeros. We can do this by subtracting from (Bias + (EltSize - 1)).
  unsigned Adjust = ExponentBias + (EltSize - 1);
  return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(Adjust, DL, VT), Trunc);
}
 
// While RVV has alignment restrictions, we should always be able to load as a
// legal equivalently-sized byte-typed vector instead. This method is
// responsible for re-expressing a ISD::LOAD via a correctly-aligned type. If
// the load is already correctly-aligned, it returns SDValue().
SDValue RISCVTargetLowering::expandUnalignedRVVLoad(SDValue Op,
                                                    SelectionDAG &DAG) const {
  auto *Load = cast<LoadSDNode>(Op);
  assert(Load && Load->getMemoryVT().isVector() && "Expected vector load");
 
  if (allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
                                     Load->getMemoryVT(),
                                     *Load->getMemOperand()))
    return SDValue();
 
  SDLoc DL(Op);
  MVT VT = Op.getSimpleValueType();
  unsigned EltSizeBits = VT.getScalarSizeInBits();
  assert((EltSizeBits == 16 || EltSizeBits == 32 || EltSizeBits == 64) &&
         "Unexpected unaligned RVV load type");
  MVT NewVT =
      MVT::getVectorVT(MVT::i8, VT.getVectorElementCount() * (EltSizeBits / 8));
  assert(NewVT.isValid() &&
         "Expecting equally-sized RVV vector types to be legal");
  SDValue L = DAG.getLoad(NewVT, DL, Load->getChain(), Load->getBasePtr(),
                          Load->getPointerInfo(), Load->getOriginalAlign(),
                          Load->getMemOperand()->getFlags());
  return DAG.getMergeValues({DAG.getBitcast(VT, L), L.getValue(1)}, DL);
}
 
// While RVV has alignment restrictions, we should always be able to store as a
// legal equivalently-sized byte-typed vector instead. This method is
// responsible for re-expressing a ISD::STORE via a correctly-aligned type. It
// returns SDValue() if the store is already correctly aligned.
SDValue RISCVTargetLowering::expandUnalignedRVVStore(SDValue Op,
                                                     SelectionDAG &DAG) const {
  auto *Store = cast<StoreSDNode>(Op);
  assert(Store && Store->getValue().getValueType().isVector() &&
         "Expected vector store");
 
  if (allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
                                     Store->getMemoryVT(),
                                     *Store->getMemOperand()))
    return SDValue();
 
  SDLoc DL(Op);
  SDValue StoredVal = Store->getValue();
  MVT VT = StoredVal.getSimpleValueType();
  unsigned EltSizeBits = VT.getScalarSizeInBits();
  assert((EltSizeBits == 16 || EltSizeBits == 32 || EltSizeBits == 64) &&
         "Unexpected unaligned RVV store type");
  MVT NewVT =
      MVT::getVectorVT(MVT::i8, VT.getVectorElementCount() * (EltSizeBits / 8));
  assert(NewVT.isValid() &&
         "Expecting equally-sized RVV vector types to be legal");
  StoredVal = DAG.getBitcast(NewVT, StoredVal);
  return DAG.getStore(Store->getChain(), DL, StoredVal, Store->getBasePtr(),
                      Store->getPointerInfo(), Store->getOriginalAlign(),
                      Store->getMemOperand()->getFlags());
}
 
SDValue RISCVTargetLowering::LowerOperation(SDValue Op,
                                            SelectionDAG &DAG) const {
  switch (Op.getOpcode()) {
  default:
    report_fatal_error("unimplemented operand");
  case ISD::GlobalAddress:
    return lowerGlobalAddress(Op, DAG);
  case ISD::BlockAddress:
    return lowerBlockAddress(Op, DAG);
  case ISD::ConstantPool:
    return lowerConstantPool(Op, DAG);
  case ISD::JumpTable:
    return lowerJumpTable(Op, DAG);
  case ISD::GlobalTLSAddress:
    return lowerGlobalTLSAddress(Op, DAG);
  case ISD::SELECT:
    return lowerSELECT(Op, DAG);
  case ISD::BRCOND:
    return lowerBRCOND(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::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::BITCAST: {
    SDLoc DL(Op);
    EVT VT = Op.getValueType();
    SDValue Op0 = Op.getOperand(0);
    EVT Op0VT = Op0.getValueType();
    MVT XLenVT = Subtarget.getXLenVT();
    if (VT.isFixedLengthVector()) {
      // We can handle fixed length vector bitcasts with a simple replacement
      // in isel.
      if (Op0VT.isFixedLengthVector())
        return Op;
      // When bitcasting from scalar to fixed-length vector, insert the scalar
      // into a one-element vector of the result type, and perform a vector
      // bitcast.
      if (!Op0VT.isVector()) {
        EVT BVT = EVT::getVectorVT(*DAG.getContext(), Op0VT, 1);
        if (!isTypeLegal(BVT))
          return SDValue();
        return DAG.getBitcast(VT, DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, BVT,
                                              DAG.getUNDEF(BVT), Op0,
                                              DAG.getConstant(0, DL, XLenVT)));
      }
      return SDValue();
    }
    // Custom-legalize bitcasts from fixed-length vector types to scalar types
    // thus: bitcast the vector to a one-element vector type whose element type
    // is the same as the result type, and extract the first element.
    if (!VT.isVector() && Op0VT.isFixedLengthVector()) {
      EVT BVT = EVT::getVectorVT(*DAG.getContext(), VT, 1);
      if (!isTypeLegal(BVT))
        return SDValue();
      SDValue BVec = DAG.getBitcast(BVT, Op0);
      return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, BVec,
                         DAG.getConstant(0, DL, XLenVT));
    }
    if (VT == MVT::f16 && Op0VT == MVT::i16 && Subtarget.hasStdExtZfh()) {
      SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, Op0);
      SDValue FPConv = DAG.getNode(RISCVISD::FMV_H_X, DL, MVT::f16, NewOp0);
      return FPConv;
    }
    if (VT == MVT::f32 && Op0VT == MVT::i32 && Subtarget.is64Bit() &&
        Subtarget.hasStdExtF()) {
      SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op0);
      SDValue FPConv =
          DAG.getNode(RISCVISD::FMV_W_X_RV64, DL, MVT::f32, NewOp0);
      return FPConv;
    }
    return SDValue();
  }
  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::BSWAP:
  case ISD::BITREVERSE: {
    // Convert BSWAP/BITREVERSE to GREVI to enable GREVI combinining.
    assert(Subtarget.hasStdExtZbp() && "Unexpected custom legalisation");
    MVT VT = Op.getSimpleValueType();
    SDLoc DL(Op);
    // Start with the maximum immediate value which is the bitwidth - 1.
    unsigned Imm = VT.getSizeInBits() - 1;
    // If this is BSWAP rather than BITREVERSE, clear the lower 3 bits.
    if (Op.getOpcode() == ISD::BSWAP)
      Imm &= ~0x7U;
    return DAG.getNode(RISCVISD::GREV, DL, VT, Op.getOperand(0),
                       DAG.getConstant(Imm, DL, VT));
  }
  case ISD::FSHL:
  case ISD::FSHR: {
    MVT VT = Op.getSimpleValueType();
    assert(VT == Subtarget.getXLenVT() && "Unexpected custom legalization");
    SDLoc DL(Op);
    // FSL/FSR take a log2(XLen)+1 bit shift amount but XLenVT FSHL/FSHR only
    // use log(XLen) bits. Mask the shift amount accordingly to prevent
    // accidentally setting the extra bit.
    unsigned ShAmtWidth = Subtarget.getXLen() - 1;
    SDValue ShAmt = DAG.getNode(ISD::AND, DL, VT, Op.getOperand(2),
                                DAG.getConstant(ShAmtWidth, DL, VT));
    // fshl and fshr concatenate their operands in the same order. fsr and fsl
    // instruction use different orders. fshl will return its first operand for
    // shift of zero, fshr will return its second operand. fsl and fsr both
    // return rs1 so the ISD nodes need to have different operand orders.
    // Shift amount is in rs2.
    SDValue Op0 = Op.getOperand(0);
    SDValue Op1 = Op.getOperand(1);
    unsigned Opc = RISCVISD::FSL;
    if (Op.getOpcode() == ISD::FSHR) {
      std::swap(Op0, Op1);
      Opc = RISCVISD::FSR;
    }
    return DAG.getNode(Opc, DL, VT, Op0, Op1, ShAmt);
  }
  case ISD::TRUNCATE: {
    SDLoc DL(Op);
    MVT VT = Op.getSimpleValueType();
    // Only custom-lower vector truncates
    if (!VT.isVector())
      return Op;
 
    // Truncates to mask types are handled differently
    if (VT.getVectorElementType() == MVT::i1)
      return lowerVectorMaskTrunc(Op, DAG);
 
    // RVV only has truncates which operate from SEW*2->SEW, so lower arbitrary
    // truncates as a series of "RISCVISD::TRUNCATE_VECTOR_VL" nodes which
    // truncate by one power of two at a time.
    MVT DstEltVT = VT.getVectorElementType();
 
    SDValue Src = Op.getOperand(0);
    MVT SrcVT = Src.getSimpleValueType();
    MVT SrcEltVT = SrcVT.getVectorElementType();
 
    assert(DstEltVT.bitsLT(SrcEltVT) &&
           isPowerOf2_64(DstEltVT.getSizeInBits()) &&
           isPowerOf2_64(SrcEltVT.getSizeInBits()) &&
           "Unexpected vector truncate lowering");
 
    MVT ContainerVT = SrcVT;
    if (SrcVT.isFixedLengthVector()) {
      ContainerVT = getContainerForFixedLengthVector(SrcVT);
      Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
    }
 
    SDValue Result = Src;
    SDValue Mask, VL;
    std::tie(Mask, VL) =
        getDefaultVLOps(SrcVT, ContainerVT, DL, DAG, Subtarget);
    LLVMContext &Context = *DAG.getContext();
    const ElementCount Count = ContainerVT.getVectorElementCount();
    do {
      SrcEltVT = MVT::getIntegerVT(SrcEltVT.getSizeInBits() / 2);
      EVT ResultVT = EVT::getVectorVT(Context, SrcEltVT, Count);
      Result = DAG.getNode(RISCVISD::TRUNCATE_VECTOR_VL, DL, ResultVT, Result,
                           Mask, VL);
    } while (SrcEltVT != DstEltVT);
 
    if (SrcVT.isFixedLengthVector())
      Result = convertFromScalableVector(VT, Result, DAG, Subtarget);
 
    return Result;
  }
  case ISD::ANY_EXTEND:
  case ISD::ZERO_EXTEND:
    if (Op.getOperand(0).getValueType().isVector() &&
        Op.getOperand(0).getValueType().getVectorElementType() == MVT::i1)
      return lowerVectorMaskExt(Op, DAG, /*ExtVal*/ 1);
    return lowerFixedLengthVectorExtendToRVV(Op, DAG, RISCVISD::VZEXT_VL);
  case ISD::SIGN_EXTEND:
    if (Op.getOperand(0).getValueType().isVector() &&
        Op.getOperand(0).getValueType().getVectorElementType() == MVT::i1)
      return lowerVectorMaskExt(Op, DAG, /*ExtVal*/ -1);
    return lowerFixedLengthVectorExtendToRVV(Op, DAG, RISCVISD::VSEXT_VL);
  case ISD::SPLAT_VECTOR_PARTS:
    return lowerSPLAT_VECTOR_PARTS(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::VSCALE: {
    MVT VT = Op.getSimpleValueType();
    SDLoc DL(Op);
    SDValue VLENB = DAG.getNode(RISCVISD::READ_VLENB, DL, VT);
    // We define our scalable vector types for lmul=1 to use a 64 bit known
    // minimum size. e.g. <vscale x 2 x i32>. VLENB is in bytes so we calculate
    // vscale as VLENB / 8.
    static_assert(RISCV::RVVBitsPerBlock == 64, "Unexpected bits per block!");
    if (isa<ConstantSDNode>(Op.getOperand(0))) {
      // We assume VLENB is a multiple of 8. We manually choose the best shift
      // here because SimplifyDemandedBits isn't always able to simplify it.
      uint64_t Val = Op.getConstantOperandVal(0);
      if (isPowerOf2_64(Val)) {
        uint64_t Log2 = Log2_64(Val);
        if (Log2 < 3)
          return DAG.getNode(ISD::SRL, DL, VT, VLENB,
                             DAG.getConstant(3 - Log2, DL, VT));
        if (Log2 > 3)
          return DAG.getNode(ISD::SHL, DL, VT, VLENB,
                             DAG.getConstant(Log2 - 3, DL, VT));
        return VLENB;
      }
      // If the multiplier is a multiple of 8, scale it down to avoid needing
      // to shift the VLENB value.
      if ((Val % 8) == 0)
        return DAG.getNode(ISD::MUL, DL, VT, VLENB,
                           DAG.getConstant(Val / 8, DL, VT));
    }
 
    SDValue VScale = DAG.getNode(ISD::SRL, DL, VT, VLENB,
                                 DAG.getConstant(3, DL, VT));
    return DAG.getNode(ISD::MUL, DL, VT, VScale, Op.getOperand(0));
  }
  case ISD::FPOWI: {
    // Custom promote f16 powi with illegal i32 integer type on RV64. Once
    // promoted this will be legalized into a libcall by LegalizeIntegerTypes.
    if (Op.getValueType() == MVT::f16 && Subtarget.is64Bit() &&
        Op.getOperand(1).getValueType() == MVT::i32) {
      SDLoc DL(Op);
      SDValue Op0 = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Op.getOperand(0));
      SDValue Powi =
          DAG.getNode(ISD::FPOWI, DL, MVT::f32, Op0, Op.getOperand(1));
      return DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, Powi,
                         DAG.getIntPtrConstant(0, DL));
    }
    return SDValue();
  }
  case ISD::FP_EXTEND: {
    // RVV can only do fp_extend to types double the size as the source. We
    // custom-lower f16->f64 extensions to two hops of ISD::FP_EXTEND, going
    // via f32.
    SDLoc DL(Op);
    MVT VT = Op.getSimpleValueType();
    SDValue Src = Op.getOperand(0);
    MVT SrcVT = Src.getSimpleValueType();
 
    // Prepare any fixed-length vector operands.
    MVT ContainerVT = VT;
    if (SrcVT.isFixedLengthVector()) {
      ContainerVT = getContainerForFixedLengthVector(VT);
      MVT SrcContainerVT =
          ContainerVT.changeVectorElementType(SrcVT.getVectorElementType());
      Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget);
    }
 
    if (!VT.isVector() || VT.getVectorElementType() != MVT::f64 ||
        SrcVT.getVectorElementType() != MVT::f16) {
      // For scalable vectors, we only need to close the gap between
      // vXf16->vXf64.
      if (!VT.isFixedLengthVector())
        return Op;
      // For fixed-length vectors, lower the FP_EXTEND to a custom "VL" version.
      Src = getRVVFPExtendOrRound(Src, VT, ContainerVT, DL, DAG, Subtarget);
      return convertFromScalableVector(VT, Src, DAG, Subtarget);
    }
 
    MVT InterVT = VT.changeVectorElementType(MVT::f32);
    MVT InterContainerVT = ContainerVT.changeVectorElementType(MVT::f32);
    SDValue IntermediateExtend = getRVVFPExtendOrRound(
        Src, InterVT, InterContainerVT, DL, DAG, Subtarget);
 
    SDValue Extend = getRVVFPExtendOrRound(IntermediateExtend, VT, ContainerVT,
                                           DL, DAG, Subtarget);
    if (VT.isFixedLengthVector())
      return convertFromScalableVector(VT, Extend, DAG, Subtarget);
    return Extend;
  }
  case ISD::FP_ROUND: {
    // RVV can only do fp_round to types half the size as the source. We
    // custom-lower f64->f16 rounds via RVV's round-to-odd float
    // conversion instruction.
    SDLoc DL(Op);
    MVT VT = Op.getSimpleValueType();
    SDValue Src = Op.getOperand(0);
    MVT SrcVT = Src.getSimpleValueType();
 
    // Prepare any fixed-length vector operands.
    MVT ContainerVT = VT;
    if (VT.isFixedLengthVector()) {
      MVT SrcContainerVT = getContainerForFixedLengthVector(SrcVT);
      ContainerVT =
          SrcContainerVT.changeVectorElementType(VT.getVectorElementType());
      Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget);
    }
 
    if (!VT.isVector() || VT.getVectorElementType() != MVT::f16 ||
        SrcVT.getVectorElementType() != MVT::f64) {
      // For scalable vectors, we only need to close the gap between
      // vXf64<->vXf16.
      if (!VT.isFixedLengthVector())
        return Op;
      // For fixed-length vectors, lower the FP_ROUND to a custom "VL" version.
      Src = getRVVFPExtendOrRound(Src, VT, ContainerVT, DL, DAG, Subtarget);
      return convertFromScalableVector(VT, Src, DAG, Subtarget);
    }
 
    SDValue Mask, VL;
    std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
 
    MVT InterVT = ContainerVT.changeVectorElementType(MVT::f32);
    SDValue IntermediateRound =
        DAG.getNode(RISCVISD::VFNCVT_ROD_VL, DL, InterVT, Src, Mask, VL);
    SDValue Round = getRVVFPExtendOrRound(IntermediateRound, VT, ContainerVT,
                                          DL, DAG, Subtarget);
 
    if (VT.isFixedLengthVector())
      return convertFromScalableVector(VT, Round, DAG, Subtarget);
    return Round;
  }
  case ISD::FP_TO_SINT:
  case ISD::FP_TO_UINT:
  case ISD::SINT_TO_FP:
  case ISD::UINT_TO_FP: {
    // RVV can only do fp<->int conversions to types half/double the size as
    // the source. We custom-lower any conversions that do two hops into
    // sequences.
    MVT VT = Op.getSimpleValueType();
    if (!VT.isVector())
      return Op;
    SDLoc DL(Op);
    SDValue Src = Op.getOperand(0);
    MVT EltVT = VT.getVectorElementType();
    MVT SrcVT = Src.getSimpleValueType();
    MVT SrcEltVT = SrcVT.getVectorElementType();
    unsigned EltSize = EltVT.getSizeInBits();
    unsigned SrcEltSize = SrcEltVT.getSizeInBits();
    assert(isPowerOf2_32(EltSize) && isPowerOf2_32(SrcEltSize) &&
           "Unexpected vector element types");
 
    bool IsInt2FP = SrcEltVT.isInteger();
    // Widening conversions
    if (EltSize > SrcEltSize && (EltSize / SrcEltSize >= 4)) {
      if (IsInt2FP) {
        // Do a regular integer sign/zero extension then convert to float.
        MVT IVecVT = MVT::getVectorVT(MVT::getIntegerVT(EltVT.getSizeInBits()),
                                      VT.getVectorElementCount());
        unsigned ExtOpcode = Op.getOpcode() == ISD::UINT_TO_FP
                                 ? ISD::ZERO_EXTEND
                                 : ISD::SIGN_EXTEND;
        SDValue Ext = DAG.getNode(ExtOpcode, DL, IVecVT, Src);
        return DAG.getNode(Op.getOpcode(), DL, VT, Ext);
      }
      // FP2Int
      assert(SrcEltVT == MVT::f16 && "Unexpected FP_TO_[US]INT lowering");
      // Do one doubling fp_extend then complete the operation by converting
      // to int.
      MVT InterimFVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
      SDValue FExt = DAG.getFPExtendOrRound(Src, DL, InterimFVT);
      return DAG.getNode(Op.getOpcode(), DL, VT, FExt);
    }
 
    // Narrowing conversions
    if (SrcEltSize > EltSize && (SrcEltSize / EltSize >= 4)) {
      if (IsInt2FP) {
        // One narrowing int_to_fp, then an fp_round.
        assert(EltVT == MVT::f16 && "Unexpected [US]_TO_FP lowering");
        MVT InterimFVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
        SDValue Int2FP = DAG.getNode(Op.getOpcode(), DL, InterimFVT, Src);
        return DAG.getFPExtendOrRound(Int2FP, DL, VT);
      }
      // FP2Int
      // One narrowing fp_to_int, then truncate the integer. If the float isn't
      // representable by the integer, the result is poison.
      MVT IVecVT =
          MVT::getVectorVT(MVT::getIntegerVT(SrcEltVT.getSizeInBits() / 2),
                           VT.getVectorElementCount());
      SDValue FP2Int = DAG.getNode(Op.getOpcode(), DL, IVecVT, Src);
      return DAG.getNode(ISD::TRUNCATE, DL, VT, FP2Int);
    }
 
    // Scalable vectors can exit here. Patterns will handle equally-sized
    // conversions halving/doubling ones.
    if (!VT.isFixedLengthVector())
      return Op;
 
    // For fixed-length vectors we lower to a custom "VL" node.
    unsigned RVVOpc = 0;
    switch (Op.getOpcode()) {
    default:
      llvm_unreachable("Impossible opcode");
    case ISD::FP_TO_SINT:
      RVVOpc = RISCVISD::FP_TO_SINT_VL;
      break;
    case ISD::FP_TO_UINT:
      RVVOpc = RISCVISD::FP_TO_UINT_VL;
      break;
    case ISD::SINT_TO_FP:
      RVVOpc = RISCVISD::SINT_TO_FP_VL;
      break;
    case ISD::UINT_TO_FP:
      RVVOpc = RISCVISD::UINT_TO_FP_VL;
      break;
    }
 
    MVT ContainerVT, SrcContainerVT;
    // Derive the reference container type from the larger vector type.
    if (SrcEltSize > EltSize) {
      SrcContainerVT = getContainerForFixedLengthVector(SrcVT);
      ContainerVT =
          SrcContainerVT.changeVectorElementType(VT.getVectorElementType());
    } else {
      ContainerVT = getContainerForFixedLengthVector(VT);
      SrcContainerVT = ContainerVT.changeVectorElementType(SrcEltVT);
    }
 
    SDValue Mask, VL;
    std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
 
    Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget);
    Src = DAG.getNode(RVVOpc, DL, ContainerVT, Src, Mask, VL);
    return convertFromScalableVector(VT, Src, DAG, Subtarget);
  }
  case ISD::FP_TO_SINT_SAT:
  case ISD::FP_TO_UINT_SAT:
    return lowerFP_TO_INT_SAT(Op, DAG, Subtarget);
  case ISD::FTRUNC:
  case ISD::FCEIL:
  case ISD::FFLOOR:
    return lowerFTRUNC_FCEIL_FFLOOR(Op, DAG);
  case ISD::VECREDUCE_ADD:
  case ISD::VECREDUCE_UMAX:
  case ISD::VECREDUCE_SMAX:
  case ISD::VECREDUCE_UMIN:
  case ISD::VECREDUCE_SMIN:
    return lowerVECREDUCE(Op, DAG);
  case ISD::VECREDUCE_AND:
  case ISD::VECREDUCE_OR:
  case ISD::VECREDUCE_XOR:
    if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i1)
      return lowerVectorMaskVecReduction(Op, DAG, /*IsVP*/ false);
    return lowerVECREDUCE(Op, DAG);
  case ISD::VECREDUCE_FADD:
  case ISD::VECREDUCE_SEQ_FADD:
  case ISD::VECREDUCE_FMIN:
  case ISD::VECREDUCE_FMAX:
    return lowerFPVECREDUCE(Op, DAG);
  case ISD::VP_REDUCE_ADD:
  case ISD::VP_REDUCE_UMAX:
  case ISD::VP_REDUCE_SMAX:
  case ISD::VP_REDUCE_UMIN:
  case ISD::VP_REDUCE_SMIN:
  case ISD::VP_REDUCE_FADD:
  case ISD::VP_REDUCE_SEQ_FADD:
  case ISD::VP_REDUCE_FMIN:
  case ISD::VP_REDUCE_FMAX:
    return lowerVPREDUCE(Op, DAG);
  case ISD::VP_REDUCE_AND:
  case ISD::VP_REDUCE_OR:
  case ISD::VP_REDUCE_XOR:
    if (Op.getOperand(1).getValueType().getVectorElementType() == MVT::i1)
      return lowerVectorMaskVecReduction(Op, DAG, /*IsVP*/ true);
    return lowerVPREDUCE(Op, DAG);
  case ISD::INSERT_SUBVECTOR:
    return lowerINSERT_SUBVECTOR(Op, DAG);
  case ISD::EXTRACT_SUBVECTOR:
    return lowerEXTRACT_SUBVECTOR(Op, DAG);
  case ISD::STEP_VECTOR:
    return lowerSTEP_VECTOR(Op, DAG);
  case ISD::VECTOR_REVERSE:
    return lowerVECTOR_REVERSE(Op, DAG);
  case ISD::BUILD_VECTOR:
    return lowerBUILD_VECTOR(Op, DAG, Subtarget);
  case ISD::SPLAT_VECTOR:
    if (Op.getValueType().getVectorElementType() == MVT::i1)
      return lowerVectorMaskSplat(Op, DAG);
    return lowerSPLAT_VECTOR(Op, DAG, Subtarget);
  case ISD::VECTOR_SHUFFLE:
    return lowerVECTOR_SHUFFLE(Op, DAG, Subtarget);
  case ISD::CONCAT_VECTORS: {
    // Split CONCAT_VECTORS into a series of INSERT_SUBVECTOR nodes. This is
    // better than going through the stack, as the default expansion does.
    SDLoc DL(Op);
    MVT VT = Op.getSimpleValueType();
    unsigned NumOpElts =
        Op.getOperand(0).getSimpleValueType().getVectorMinNumElements();
    SDValue Vec = DAG.getUNDEF(VT);
    for (const auto &OpIdx : enumerate(Op->ops())) {
      SDValue SubVec = OpIdx.value();
      // Don't insert undef subvectors.
      if (SubVec.isUndef())
        continue;
      Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, Vec, SubVec,
                        DAG.getIntPtrConstant(OpIdx.index() * NumOpElts, DL));
    }
    return Vec;
  }
  case ISD::LOAD:
    if (auto V = expandUnalignedRVVLoad(Op, DAG))
      return V;
    if (Op.getValueType().isFixedLengthVector())
      return lowerFixedLengthVectorLoadToRVV(Op, DAG);
    return Op;
  case ISD::STORE:
    if (auto V = expandUnalignedRVVStore(Op, DAG))
      return V;
    if (Op.getOperand(1).getValueType().isFixedLengthVector())
      return lowerFixedLengthVectorStoreToRVV(Op, DAG);
    return Op;
  case ISD::MLOAD:
  case ISD::VP_LOAD:
    return lowerMaskedLoad(Op, DAG);
  case ISD::MSTORE:
  case ISD::VP_STORE:
    return lowerMaskedStore(Op, DAG);
  case ISD::SETCC:
    return lowerFixedLengthVectorSetccToRVV(Op, DAG);
  case ISD::ADD:
    return lowerToScalableOp(Op, DAG, RISCVISD::ADD_VL);
  case ISD::SUB:
    return lowerToScalableOp(Op, DAG, RISCVISD::SUB_VL);
  case ISD::MUL:
    return lowerToScalableOp(Op, DAG, RISCVISD::MUL_VL);
  case ISD::MULHS:
    return lowerToScalableOp(Op, DAG, RISCVISD::MULHS_VL);
  case ISD::MULHU:
    return lowerToScalableOp(Op, DAG, RISCVISD::MULHU_VL);
  case ISD::AND:
    return lowerFixedLengthVectorLogicOpToRVV(Op, DAG, RISCVISD::VMAND_VL,
                                              RISCVISD::AND_VL);
  case ISD::OR:
    return lowerFixedLengthVectorLogicOpToRVV(Op, DAG, RISCVISD::VMOR_VL,
                                              RISCVISD::OR_VL);
  case ISD::XOR:
    return lowerFixedLengthVectorLogicOpToRVV(Op, DAG, RISCVISD::VMXOR_VL,
                                              RISCVISD::XOR_VL);
  case ISD::SDIV:
    return lowerToScalableOp(Op, DAG, RISCVISD::SDIV_VL);
  case ISD::SREM:
    return lowerToScalableOp(Op, DAG, RISCVISD::SREM_VL);
  case ISD::UDIV:
    return lowerToScalableOp(Op, DAG, RISCVISD::UDIV_VL);
  case ISD::UREM:
    return lowerToScalableOp(Op, DAG, RISCVISD::UREM_VL);
  case ISD::SHL:
  case ISD::SRA:
  case ISD::SRL:
    if (Op.getSimpleValueType().isFixedLengthVector())
      return lowerFixedLengthVectorShiftToRVV(Op, DAG);
    // This can be called for an i32 shift amount that needs to be promoted.
    assert(Op.getOperand(1).getValueType() == MVT::i32 && Subtarget.is64Bit() &&
           "Unexpected custom legalisation");
    return SDValue();
  case ISD::SADDSAT:
    return lowerToScalableOp(Op, DAG, RISCVISD::SADDSAT_VL);
  case ISD::UADDSAT:
    return lowerToScalableOp(Op, DAG, RISCVISD::UADDSAT_VL);
  case ISD::SSUBSAT:
    return lowerToScalableOp(Op, DAG, RISCVISD::SSUBSAT_VL);
  case ISD::USUBSAT:
    return lowerToScalableOp(Op, DAG, RISCVISD::USUBSAT_VL);
  case ISD::FADD:
    return lowerToScalableOp(Op, DAG, RISCVISD::FADD_VL);
  case ISD::FSUB:
    return lowerToScalableOp(Op, DAG, RISCVISD::FSUB_VL);
  case ISD::FMUL:
    return lowerToScalableOp(Op, DAG, RISCVISD::FMUL_VL);
  case ISD::FDIV:
    return lowerToScalableOp(Op, DAG, RISCVISD::FDIV_VL);
  case ISD::FNEG:
    return lowerToScalableOp(Op, DAG, RISCVISD::FNEG_VL);
  case ISD::FABS:
    return lowerToScalableOp(Op, DAG, RISCVISD::FABS_VL);
  case ISD::FSQRT:
    return lowerToScalableOp(Op, DAG, RISCVISD::FSQRT_VL);
  case ISD::FMA:
    return lowerToScalableOp(Op, DAG, RISCVISD::FMA_VL);
  case ISD::SMIN:
    return lowerToScalableOp(Op, DAG, RISCVISD::SMIN_VL);
  case ISD::SMAX:
    return lowerToScalableOp(Op, DAG, RISCVISD::SMAX_VL);
  case ISD::UMIN:
    return lowerToScalableOp(Op, DAG, RISCVISD::UMIN_VL);
  case ISD::UMAX:
    return lowerToScalableOp(Op, DAG, RISCVISD::UMAX_VL);
  case ISD::FMINNUM:
    return lowerToScalableOp(Op, DAG, RISCVISD::FMINNUM_VL);
  case ISD::FMAXNUM:
    return lowerToScalableOp(Op, DAG, RISCVISD::FMAXNUM_VL);
  case ISD::ABS:
    return lowerABS(Op, DAG);
  case ISD::CTLZ_ZERO_UNDEF:
  case ISD::CTTZ_ZERO_UNDEF:
    return lowerCTLZ_CTTZ_ZERO_UNDEF(Op, DAG);
  case ISD::VSELECT:
    return lowerFixedLengthVectorSelectToRVV(Op, DAG);
  case ISD::FCOPYSIGN:
    return lowerFixedLengthVectorFCOPYSIGNToRVV(Op, DAG);
  case ISD::MGATHER:
  case ISD::VP_GATHER:
    return lowerMaskedGather(Op, DAG);
  case ISD::MSCATTER:
  case ISD::VP_SCATTER:
    return lowerMaskedScatter(Op, DAG);
  case ISD::FLT_ROUNDS_:
    return lowerGET_ROUNDING(Op, DAG);
  case ISD::SET_ROUNDING:
    return lowerSET_ROUNDING(Op, DAG);
  case ISD::VP_SELECT:
    return lowerVPOp(Op, DAG, RISCVISD::VSELECT_VL);
  case ISD::VP_MERGE:
    return lowerVPOp(Op, DAG, RISCVISD::VP_MERGE_VL);
  case ISD::VP_ADD:
    return lowerVPOp(Op, DAG, RISCVISD::ADD_VL);
  case ISD::VP_SUB:
    return lowerVPOp(Op, DAG, RISCVISD::SUB_VL);
  case ISD::VP_MUL:
    return lowerVPOp(Op, DAG, RISCVISD::MUL_VL);
  case ISD::VP_SDIV:
    return lowerVPOp(Op, DAG, RISCVISD::SDIV_VL);
  case ISD::VP_UDIV:
    return lowerVPOp(Op, DAG, RISCVISD::UDIV_VL);
  case ISD::VP_SREM:
    return lowerVPOp(Op, DAG, RISCVISD::SREM_VL);
  case ISD::VP_UREM:
    return lowerVPOp(Op, DAG, RISCVISD::UREM_VL);
  case ISD::VP_AND:
    return lowerLogicVPOp(Op, DAG, RISCVISD::VMAND_VL, RISCVISD::AND_VL);
  case ISD::VP_OR:
    return lowerLogicVPOp(Op, DAG, RISCVISD::VMOR_VL, RISCVISD::OR_VL);
  case ISD::VP_XOR:
    return lowerLogicVPOp(Op, DAG, RISCVISD::VMXOR_VL, RISCVISD::XOR_VL);
  case ISD::VP_ASHR:
    return lowerVPOp(Op, DAG, RISCVISD::SRA_VL);
  case ISD::VP_LSHR:
    return lowerVPOp(Op, DAG, RISCVISD::SRL_VL);
  case ISD::VP_SHL:
    return lowerVPOp(Op, DAG, RISCVISD::SHL_VL);
  case ISD::VP_FADD:
    return lowerVPOp(Op, DAG, RISCVISD::FADD_VL);
  case ISD::VP_FSUB:
    return lowerVPOp(Op, DAG, RISCVISD::FSUB_VL);
  case ISD::VP_FMUL:
    return lowerVPOp(Op, DAG, RISCVISD::FMUL_VL);
  case ISD::VP_FDIV:
    return lowerVPOp(Op, DAG, RISCVISD::FDIV_VL);
  }
}
 
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 RISCVTargetLowering::getAddr(NodeTy *N, SelectionDAG &DAG,
                                     bool IsLocal) const {
  SDLoc DL(N);
  EVT Ty = getPointerTy(DAG.getDataLayout());
 
  if (isPositionIndependent()) {
    SDValue Addr = getTargetNode(N, DL, Ty, DAG, 0);
    if (IsLocal)
      // Use PC-relative addressing to access the symbol. This generates the
      // pattern (PseudoLLA sym), which expands to (addi (auipc %pcrel_hi(sym))
      // %pcrel_lo(auipc)).
      return SDValue(DAG.getMachineNode(RISCV::PseudoLLA, DL, Ty, Addr), 0);
 
    // Use PC-relative addressing to access the GOT for this symbol, then load
    // the address from the GOT. This generates the pattern (PseudoLA sym),
    // which expands to (ld (addi (auipc %got_pcrel_hi(sym)) %pcrel_lo(auipc))).
    return SDValue(DAG.getMachineNode(RISCV::PseudoLA, DL, Ty, Addr), 0);
  }
 
  switch (getTargetMachine().getCodeModel()) {
  default:
    report_fatal_error("Unsupported code model for lowering");
  case CodeModel::Small: {
    // Generate a sequence for accessing addresses within the first 2 GiB of
    // address space. This generates the pattern (addi (lui %hi(sym)) %lo(sym)).
    SDValue AddrHi = getTargetNode(N, DL, Ty, DAG, RISCVII::MO_HI);
    SDValue AddrLo = getTargetNode(N, DL, Ty, DAG, RISCVII::MO_LO);
    SDValue MNHi = SDValue(DAG.getMachineNode(RISCV::LUI, DL, Ty, AddrHi), 0);
    return SDValue(DAG.getMachineNode(RISCV::ADDI, DL, Ty, MNHi, AddrLo), 0);
  }
  case CodeModel::Medium: {
    // Generate a sequence for accessing addresses within any 2GiB range within
    // the address space. This generates the pattern (PseudoLLA sym), which
    // expands to (addi (auipc %pcrel_hi(sym)) %pcrel_lo(auipc)).
    SDValue Addr = getTargetNode(N, DL, Ty, DAG, 0);
    return SDValue(DAG.getMachineNode(RISCV::PseudoLLA, DL, Ty, Addr), 0);
  }
  }
}
 
SDValue RISCVTargetLowering::lowerGlobalAddress(SDValue Op,
                                                SelectionDAG &DAG) const {
  SDLoc DL(Op);
  EVT Ty = Op.getValueType();
  GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
  int64_t Offset = N->getOffset();
  MVT XLenVT = Subtarget.getXLenVT();
 
  const GlobalValue *GV = N->getGlobal();
  bool IsLocal = getTargetMachine().shouldAssumeDSOLocal(*GV->getParent(), GV);
  SDValue Addr = getAddr(N, DAG, IsLocal);
 
  // In order to maximise the opportunity for common subexpression elimination,
  // emit 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.
  if (Offset != 0)
    return DAG.getNode(ISD::ADD, DL, Ty, Addr,
                       DAG.getConstant(Offset, DL, XLenVT));
  return Addr;
}
 
SDValue RISCVTargetLowering::lowerBlockAddress(SDValue Op,
                                               SelectionDAG &DAG) const {
  BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op);
 
  return getAddr(N, DAG);
}
 
SDValue RISCVTargetLowering::lowerConstantPool(SDValue Op,
                                               SelectionDAG &DAG) const {
  ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
 
  return getAddr(N, DAG);
}
 
SDValue RISCVTargetLowering::lowerJumpTable(SDValue Op,
                                            SelectionDAG &DAG) const {
  JumpTableSDNode *N = cast<JumpTableSDNode>(Op);
 
  return getAddr(N, DAG);
}
 
SDValue RISCVTargetLowering::getStaticTLSAddr(GlobalAddressSDNode *N,
                                              SelectionDAG &DAG,
                                              bool UseGOT) const {
  SDLoc DL(N);
  EVT Ty = getPointerTy(DAG.getDataLayout());
  const GlobalValue *GV = N->getGlobal();
  MVT XLenVT = Subtarget.getXLenVT();
 
  if (UseGOT) {
    // Use PC-relative addressing to access the GOT for this TLS symbol, then
    // load the address from the GOT and add the thread pointer. This generates
    // the pattern (PseudoLA_TLS_IE sym), which expands to
    // (ld (auipc %tls_ie_pcrel_hi(sym)) %pcrel_lo(auipc)).
    SDValue Addr = DAG.getTargetGlobalAddress(GV, DL, Ty, 0, 0);
    SDValue Load =
        SDValue(DAG.getMachineNode(RISCV::PseudoLA_TLS_IE, DL, Ty, Addr), 0);
 
    // Add the thread pointer.
    SDValue TPReg = DAG.getRegister(RISCV::X4, XLenVT);
    return DAG.getNode(ISD::ADD, DL, Ty, Load, TPReg);
  }
 
  // Generate a sequence for accessing the address relative to the thread
  // pointer, with the appropriate adjustment for the thread pointer offset.
  // This generates the pattern
  // (add (add_tprel (lui %tprel_hi(sym)) tp %tprel_add(sym)) %tprel_lo(sym))
  SDValue AddrHi =
      DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_HI);
  SDValue AddrAdd =
      DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_ADD);
  SDValue AddrLo =
      DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_LO);
 
  SDValue MNHi = SDValue(DAG.getMachineNode(RISCV::LUI, DL, Ty, AddrHi), 0);
  SDValue TPReg = DAG.getRegister(RISCV::X4, XLenVT);
  SDValue MNAdd = SDValue(
      DAG.getMachineNode(RISCV::PseudoAddTPRel, DL, Ty, MNHi, TPReg, AddrAdd),
      0);
  return SDValue(DAG.getMachineNode(RISCV::ADDI, DL, Ty, MNAdd, AddrLo), 0);
}
 
SDValue RISCVTargetLowering::getDynamicTLSAddr(GlobalAddressSDNode *N,
                                               SelectionDAG &DAG) const {
  SDLoc DL(N);
  EVT Ty = getPointerTy(DAG.getDataLayout());
  IntegerType *CallTy = Type::getIntNTy(*DAG.getContext(), Ty.getSizeInBits());
  const GlobalValue *GV = N->getGlobal();
 
  // Use a PC-relative addressing mode to access the global dynamic GOT address.
  // This generates the pattern (PseudoLA_TLS_GD sym), which expands to
  // (addi (auipc %tls_gd_pcrel_hi(sym)) %pcrel_lo(auipc)).
  SDValue Addr = DAG.getTargetGlobalAddress(GV, DL, Ty, 0, 0);
  SDValue Load =
      SDValue(DAG.getMachineNode(RISCV::PseudoLA_TLS_GD, DL, Ty, Addr), 0);
 
  // Prepare argument list to generate call.
  ArgListTy Args;
  ArgListEntry Entry;
  Entry.Node = Load;
  Entry.Ty = CallTy;
  Args.push_back(Entry);
 
  // 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 RISCVTargetLowering::lowerGlobalTLSAddress(SDValue Op,
                                                   SelectionDAG &DAG) const {
  SDLoc DL(Op);
  EVT Ty = Op.getValueType();
  GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
  int64_t Offset = N->getOffset();
  MVT XLenVT = Subtarget.getXLenVT();
 
  TLSModel::Model Model = getTargetMachine().getTLSModel(N->getGlobal());
 
  if (DAG.getMachineFunction().getFunction().getCallingConv() ==
      CallingConv::GHC)
    report_fatal_error("In GHC calling convention TLS is not supported");
 
  SDValue Addr;
  switch (Model) {
  case TLSModel::LocalExec:
    Addr = getStaticTLSAddr(N, DAG, /*UseGOT=*/false);
    break;
  case TLSModel::InitialExec:
    Addr = getStaticTLSAddr(N, DAG, /*UseGOT=*/true);
    break;
  case TLSModel::LocalDynamic:
  case TLSModel::GeneralDynamic:
    Addr = getDynamicTLSAddr(N, DAG);
    break;
  }
 
  // In order to maximise the opportunity for common subexpression elimination,
  // emit 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.
  if (Offset != 0)
    return DAG.getNode(ISD::ADD, DL, Ty, Addr,
                       DAG.getConstant(Offset, DL, XLenVT));
  return Addr;
}
 
SDValue RISCVTargetLowering::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 XLenVT = Subtarget.getXLenVT();
 
  // Lower vector SELECTs to VSELECTs by splatting the condition.
  if (VT.isVector()) {
    MVT SplatCondVT = VT.changeVectorElementType(MVT::i1);
    SDValue CondSplat = VT.isScalableVector()
                            ? DAG.getSplatVector(SplatCondVT, DL, CondV)
                            : DAG.getSplatBuildVector(SplatCondVT, DL, CondV);
    return DAG.getNode(ISD::VSELECT, DL, VT, CondSplat, TrueV, FalseV);
  }
 
  // If the result type is XLenVT and CondV is the output of a SETCC node
  // which also operated on XLenVT inputs, then merge the SETCC node into the
  // lowered RISCVISD::SELECT_CC to take advantage of the integer
  // compare+branch instructions. i.e.:
  // (select (setcc lhs, rhs, cc), truev, falsev)
  // -> (riscvisd::select_cc lhs, rhs, cc, truev, falsev)
  if (VT == XLenVT && CondV.getOpcode() == ISD::SETCC &&
      CondV.getOperand(0).getSimpleValueType() == XLenVT) {
    SDValue LHS = CondV.getOperand(0);
    SDValue RHS = CondV.getOperand(1);
    const auto *CC = cast<CondCodeSDNode>(CondV.getOperand(2));
    ISD::CondCode CCVal = CC->get();
 
    // Special case for a select of 2 constants that have a diffence 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 = cast<ConstantSDNode>(TrueV)->getAPIntValue();
      const APInt &FalseVal = cast<ConstantSDNode>(FalseV)->getAPIntValue();
      if (TrueVal - 1 == FalseVal)
        return DAG.getNode(ISD::ADD, DL, Op.getValueType(), CondV, FalseV);
      if (TrueVal + 1 == FalseVal)
        return DAG.getNode(ISD::SUB, DL, Op.getValueType(), FalseV, CondV);
    }
 
    translateSetCCForBranch(DL, LHS, RHS, CCVal, DAG);
 
    SDValue TargetCC = DAG.getCondCode(CCVal);
    SDValue Ops[] = {LHS, RHS, TargetCC, TrueV, FalseV};
    return DAG.getNode(RISCVISD::SELECT_CC, DL, Op.getValueType(), Ops);
  }
 
  // Otherwise:
  // (select condv, truev, falsev)
  // -> (riscvisd::select_cc condv, zero, setne, truev, falsev)
  SDValue Zero = DAG.getConstant(0, DL, XLenVT);
  SDValue SetNE = DAG.getCondCode(ISD::SETNE);
 
  SDValue Ops[] = {CondV, Zero, SetNE, TrueV, FalseV};
 
  return DAG.getNode(RISCVISD::SELECT_CC, DL, Op.getValueType(), Ops);
}
 
SDValue RISCVTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
  SDValue CondV = Op.getOperand(1);
  SDLoc DL(Op);
  MVT XLenVT = Subtarget.getXLenVT();
 
  if (CondV.getOpcode() == ISD::SETCC &&
      CondV.getOperand(0).getValueType() == XLenVT) {
    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(RISCVISD::BR_CC, DL, Op.getValueType(), Op.getOperand(0),
                       LHS, RHS, TargetCC, Op.getOperand(2));
  }
 
  return DAG.getNode(RISCVISD::BR_CC, DL, Op.getValueType(), Op.getOperand(0),
                     CondV, DAG.getConstant(0, DL, XLenVT),
                     DAG.getCondCode(ISD::SETNE), Op.getOperand(2));
}
 
SDValue RISCVTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  RISCVMachineFunctionInfo *FuncInfo = MF.getInfo<RISCVMachineFunctionInfo>();
 
  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 RISCVTargetLowering::lowerFRAMEADDR(SDValue Op,
                                            SelectionDAG &DAG) const {
  const RISCVRegisterInfo &RI = *Subtarget.getRegisterInfo();
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo &MFI = MF.getFrameInfo();
  MFI.setFrameAddressIsTaken(true);
  Register FrameReg = RI.getFrameRegister(MF);
  int XLenInBytes = Subtarget.getXLen() / 8;
 
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, FrameReg, VT);
  unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
  while (Depth--) {
    int Offset = -(XLenInBytes * 2);
    SDValue Ptr = DAG.getNode(ISD::ADD, DL, VT, FrameAddr,
                              DAG.getIntPtrConstant(Offset, DL));
    FrameAddr =
        DAG.getLoad(VT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo());
  }
  return FrameAddr;
}
 
SDValue RISCVTargetLowering::lowerRETURNADDR(SDValue Op,
                                             SelectionDAG &DAG) const {
  const RISCVRegisterInfo &RI = *Subtarget.getRegisterInfo();
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo &MFI = MF.getFrameInfo();
  MFI.setReturnAddressIsTaken(true);
  MVT XLenVT = Subtarget.getXLenVT();
  int XLenInBytes = Subtarget.getXLen() / 8;
 
  if (verifyReturnAddressArgumentIsConstant(Op, DAG))
    return SDValue();
 
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
  if (Depth) {
    int Off = -XLenInBytes;
    SDValue FrameAddr = lowerFRAMEADDR(Op, DAG);
    SDValue Offset = DAG.getConstant(Off, DL, VT);
    return DAG.getLoad(VT, DL, DAG.getEntryNode(),
                       DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
                       MachinePointerInfo());
  }
 
  // Return the value of the return address register, marking it an implicit
  // live-in.
  Register Reg = MF.addLiveIn(RI.getRARegister(), getRegClassFor(XLenVT));
  return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, XLenVT);
}
 
SDValue RISCVTargetLowering::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-XLEN < 0: // Shamt < XLEN
  //   Lo = Lo << Shamt
  //   Hi = (Hi << Shamt) | ((Lo >>u 1) >>u (XLEN-1 - Shamt))
  // else:
  //   Lo = 0
  //   Hi = Lo << (Shamt-XLEN)
 
  SDValue Zero = DAG.getConstant(0, DL, VT);
  SDValue One = DAG.getConstant(1, DL, VT);
  SDValue MinusXLen = DAG.getConstant(-(int)Subtarget.