RISCVISelLowering.cpp

Go to the documentation of this file.

//===-- RISCVISelLowering.cpp - RISC-V 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 RISC-V 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/Analysis/VectorUtils.h"
#include "llvm/CodeGen/Analysis.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/SelectionDAGAddressAnalysis.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/Instructions.h"
#include "llvm/IR/IntrinsicsRISCV.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/InstructionCost.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include <optional>
 
using namespace llvm;
 
#define DEBUG_TYPE "riscv-lower"
 
STATISTIC(NumTailCalls, "Number of tail calls");
 
static cl::opt<unsigned> ExtensionMaxWebSize(
    DEBUG_TYPE "-ext-max-web-size", cl::Hidden,
    cl::desc("Give the maximum size (in number of nodes) of the web of "
             "instructions that we will consider for VW expansion"),
    cl::init(18));
 
static cl::opt<bool>
    AllowSplatInVW_W(DEBUG_TYPE "-form-vw-w-with-splat", cl::Hidden,
                     cl::desc("Allow the formation of VW_W operations (e.g., "
                              "VWADD_W) with splat constants"),
                     cl::init(false));
 
static cl::opt<unsigned> NumRepeatedDivisors(
    DEBUG_TYPE "-fp-repeated-divisors", cl::Hidden,
    cl::desc("Set the minimum number of repetitions of a divisor to allow "
             "transformation to multiplications by the reciprocal"),
    cl::init(2));
 
static cl::opt<int>
    FPImmCost(DEBUG_TYPE "-fpimm-cost", cl::Hidden,
              cl::desc("Give the maximum number of instructions that we will "
                       "use for creating a floating-point immediate value"),
              cl::init(2));
 
RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM,
                                         const RISCVSubtarget &STI)
    : TargetLowering(TM), Subtarget(STI) {
 
  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_ILP32E:
  case RISCVABI::ABI_LP64E:
  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.hasStdExtZfhmin())
    addRegisterClass(MVT::f16, &RISCV::FPR16RegClass);
  if (Subtarget.hasStdExtZfbfmin())
    addRegisterClass(MVT::bf16, &RISCV::FPR16RegClass);
  if (Subtarget.hasStdExtF())
    addRegisterClass(MVT::f32, &RISCV::FPR32RegClass);
  if (Subtarget.hasStdExtD())
    addRegisterClass(MVT::f64, &RISCV::FPR64RegClass);
  if (Subtarget.hasStdExtZhinxmin())
    addRegisterClass(MVT::f16, &RISCV::GPRF16RegClass);
  if (Subtarget.hasStdExtZfinx())
    addRegisterClass(MVT::f32, &RISCV::GPRF32RegClass);
  if (Subtarget.hasStdExtZdinx()) {
    if (Subtarget.is64Bit())
      addRegisterClass(MVT::f64, &RISCV::GPRRegClass);
    else
      addRegisterClass(MVT::f64, &RISCV::GPRPairRegClass);
  }
 
  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 BF16VecVTs[] = {
      MVT::nxv1bf16, MVT::nxv2bf16,  MVT::nxv4bf16,
      MVT::nxv8bf16, MVT::nxv16bf16, MVT::nxv32bf16};
  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) {
      // Disable the smallest fractional LMUL types if ELEN is less than
      // RVVBitsPerBlock.
      unsigned MinElts = RISCV::RVVBitsPerBlock / Subtarget.getELen();
      if (VT.getVectorMinNumElements() < MinElts)
        return;
 
      unsigned Size = VT.getSizeInBits().getKnownMinValue();
      const TargetRegisterClass *RC;
      if (Size <= RISCV::RVVBitsPerBlock)
        RC = &RISCV::VRRegClass;
      else if (Size == 2 * RISCV::RVVBitsPerBlock)
        RC = &RISCV::VRM2RegClass;
      else if (Size == 4 * RISCV::RVVBitsPerBlock)
        RC = &RISCV::VRM4RegClass;
      else if (Size == 8 * RISCV::RVVBitsPerBlock)
        RC = &RISCV::VRM8RegClass;
      else
        llvm_unreachable("Unexpected size");
 
      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.hasVInstructionsF16Minimal())
      for (MVT VT : F16VecVTs)
        addRegClassForRVV(VT);
 
    if (Subtarget.hasVInstructionsBF16Minimal())
      for (MVT VT : BF16VecVTs)
        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);
 
  setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, XLenVT,
                   MVT::i1, Promote);
  // DAGCombiner can call isLoadExtLegal for types that aren't legal.
  setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, MVT::i32,
                   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);
 
  setCondCodeAction(ISD::SETGT, XLenVT, Custom);
  setCondCodeAction(ISD::SETGE, XLenVT, Expand);
  setCondCodeAction(ISD::SETUGT, XLenVT, Custom);
  setCondCodeAction(ISD::SETUGE, XLenVT, Expand);
  if (!(Subtarget.hasVendorXCValu() && !Subtarget.is64Bit())) {
    setCondCodeAction(ISD::SETULE, XLenVT, Expand);
    setCondCodeAction(ISD::SETLE, XLenVT, Expand);
  }
 
  setOperationAction({ISD::STACKSAVE, ISD::STACKRESTORE}, MVT::Other, Expand);
 
  setOperationAction(ISD::VASTART, MVT::Other, Custom);
  setOperationAction({ISD::VAARG, ISD::VACOPY, ISD::VAEND}, MVT::Other, Expand);
 
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
 
  setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom);
 
  if (!Subtarget.hasStdExtZbb() && !Subtarget.hasVendorXTHeadBb() &&
      !(Subtarget.hasVendorXCValu() && !Subtarget.is64Bit()))
    setOperationAction(ISD::SIGN_EXTEND_INREG, {MVT::i8, MVT::i16}, Expand);
 
  if (Subtarget.is64Bit()) {
    setOperationAction(ISD::EH_DWARF_CFA, MVT::i64, Custom);
 
    setOperationAction(ISD::LOAD, MVT::i32, Custom);
    setOperationAction({ISD::ADD, ISD::SUB, ISD::SHL, ISD::SRA, ISD::SRL},
                       MVT::i32, Custom);
    setOperationAction({ISD::UADDO, ISD::USUBO, ISD::UADDSAT}, MVT::i32,
                       Custom);
    if (!Subtarget.hasStdExtZbb())
      setOperationAction({ISD::SADDSAT, ISD::SSUBSAT, ISD::USUBSAT}, MVT::i32,
                         Custom);
    setOperationAction(ISD::SADDO, MVT::i32, Custom);
  }
  if (!Subtarget.hasStdExtZmmul()) {
    setOperationAction({ISD::MUL, ISD::MULHS, ISD::MULHU}, XLenVT, Expand);
  } else if (Subtarget.is64Bit()) {
    setOperationAction(ISD::MUL, MVT::i128, Custom);
    setOperationAction(ISD::MUL, MVT::i32, Custom);
  } else {
    setOperationAction(ISD::MUL, MVT::i64, Custom);
  }
 
  if (!Subtarget.hasStdExtM()) {
    setOperationAction({ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM}, XLenVT,
                       Expand);
  } else if (Subtarget.is64Bit()) {
    setOperationAction({ISD::SDIV, ISD::UDIV, ISD::UREM},
                       {MVT::i8, MVT::i16, MVT::i32}, Custom);
  }
 
  setOperationAction(
      {ISD::SDIVREM, ISD::UDIVREM, ISD::SMUL_LOHI, ISD::UMUL_LOHI}, XLenVT,
      Expand);
 
  setOperationAction({ISD::SHL_PARTS, ISD::SRL_PARTS, ISD::SRA_PARTS}, XLenVT,
                     Custom);
 
  if (Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbkb()) {
    if (Subtarget.is64Bit())
      setOperationAction({ISD::ROTL, ISD::ROTR}, MVT::i32, Custom);
  } else if (Subtarget.hasVendorXTHeadBb()) {
    if (Subtarget.is64Bit())
      setOperationAction({ISD::ROTL, ISD::ROTR}, MVT::i32, Custom);
    setOperationAction({ISD::ROTL, ISD::ROTR}, XLenVT, Custom);
  } else if (Subtarget.hasVendorXCVbitmanip() && !Subtarget.is64Bit()) {
    setOperationAction(ISD::ROTL, XLenVT, Expand);
  } else {
    setOperationAction({ISD::ROTL, ISD::ROTR}, XLenVT, Expand);
  }
 
  // 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() ||
                      Subtarget.hasVendorXTHeadBb())
                         ? Legal
                         : Expand);
 
  if (Subtarget.hasVendorXCVbitmanip() && !Subtarget.is64Bit()) {
    setOperationAction(ISD::BITREVERSE, XLenVT, Legal);
  } else {
    // Zbkb can use rev8+brev8 to implement bitreverse.
    setOperationAction(ISD::BITREVERSE, XLenVT,
                       Subtarget.hasStdExtZbkb() ? Custom : Expand);
  }
 
  if (Subtarget.hasStdExtZbb() ||
      (Subtarget.hasVendorXCValu() && !Subtarget.is64Bit())) {
    setOperationAction({ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX}, XLenVT,
                       Legal);
  }
 
  if (Subtarget.hasStdExtZbb() ||
      (Subtarget.hasVendorXCVbitmanip() && !Subtarget.is64Bit())) {
    if (Subtarget.is64Bit())
      setOperationAction({ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF}, MVT::i32, Custom);
  } else {
    setOperationAction({ISD::CTTZ, ISD::CTPOP}, XLenVT, Expand);
  }
 
  if (Subtarget.hasStdExtZbb() || Subtarget.hasVendorXTHeadBb() ||
      (Subtarget.hasVendorXCVbitmanip() && !Subtarget.is64Bit())) {
    // We need the custom lowering to make sure that the resulting sequence
    // for the 32bit case is efficient on 64bit targets.
    if (Subtarget.is64Bit())
      setOperationAction({ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF}, MVT::i32, Custom);
  } else {
    setOperationAction(ISD::CTLZ, XLenVT, Expand);
  }
 
  if (Subtarget.hasVendorXCValu() && !Subtarget.is64Bit()) {
    setOperationAction(ISD::ABS, XLenVT, Legal);
  } else if (Subtarget.hasShortForwardBranchOpt()) {
    // We can use PseudoCCSUB to implement ABS.
    setOperationAction(ISD::ABS, XLenVT, Legal);
  } else if (Subtarget.is64Bit()) {
    setOperationAction(ISD::ABS, MVT::i32, Custom);
  }
 
  if (!Subtarget.hasVendorXTHeadCondMov())
    setOperationAction(ISD::SELECT, XLenVT, Custom);
 
  static const unsigned FPLegalNodeTypes[] = {
      ISD::FMINNUM,        ISD::FMAXNUM,       ISD::LRINT,
      ISD::LLRINT,         ISD::LROUND,        ISD::LLROUND,
      ISD::STRICT_LRINT,   ISD::STRICT_LLRINT, ISD::STRICT_LROUND,
      ISD::STRICT_LLROUND, ISD::STRICT_FMA,    ISD::STRICT_FADD,
      ISD::STRICT_FSUB,    ISD::STRICT_FMUL,   ISD::STRICT_FDIV,
      ISD::STRICT_FSQRT,   ISD::STRICT_FSETCC, ISD::STRICT_FSETCCS};
 
  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 unsigned FPOpToExpand[] = {
      ISD::FSIN, ISD::FCOS,       ISD::FSINCOS,   ISD::FPOW,
      ISD::FREM};
 
  static const unsigned FPRndMode[] = {
      ISD::FCEIL, ISD::FFLOOR, ISD::FTRUNC, ISD::FRINT, ISD::FROUND,
      ISD::FROUNDEVEN};
 
  if (Subtarget.hasStdExtZfhminOrZhinxmin())
    setOperationAction(ISD::BITCAST, MVT::i16, Custom);
 
  static const unsigned ZfhminZfbfminPromoteOps[] = {
      ISD::FMINNUM,      ISD::FMAXNUM,       ISD::FADD,
      ISD::FSUB,         ISD::FMUL,          ISD::FMA,
      ISD::FDIV,         ISD::FSQRT,         ISD::FABS,
      ISD::FNEG,         ISD::STRICT_FMA,    ISD::STRICT_FADD,
      ISD::STRICT_FSUB,  ISD::STRICT_FMUL,   ISD::STRICT_FDIV,
      ISD::STRICT_FSQRT, ISD::STRICT_FSETCC, ISD::STRICT_FSETCCS,
      ISD::SETCC,        ISD::FCEIL,         ISD::FFLOOR,
      ISD::FTRUNC,       ISD::FRINT,         ISD::FROUND,
      ISD::FROUNDEVEN,   ISD::SELECT};
 
  if (Subtarget.hasStdExtZfbfmin()) {
    setOperationAction(ISD::BITCAST, MVT::i16, Custom);
    setOperationAction(ISD::BITCAST, MVT::bf16, Custom);
    setOperationAction(ISD::FP_ROUND, MVT::bf16, Custom);
    setOperationAction(ISD::FP_EXTEND, MVT::f32, Custom);
    setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
    setOperationAction(ISD::ConstantFP, MVT::bf16, Expand);
    setOperationAction(ISD::SELECT_CC, MVT::bf16, Expand);
    setOperationAction(ISD::BR_CC, MVT::bf16, Expand);
    setOperationAction(ZfhminZfbfminPromoteOps, MVT::bf16, Promote);
    setOperationAction(ISD::FREM, MVT::bf16, Promote);
    // FIXME: Need to promote bf16 FCOPYSIGN to f32, but the
    // DAGCombiner::visitFP_ROUND probably needs improvements first.
    setOperationAction(ISD::FCOPYSIGN, MVT::bf16, Expand);
  }
 
  if (Subtarget.hasStdExtZfhminOrZhinxmin()) {
    if (Subtarget.hasStdExtZfhOrZhinx()) {
      setOperationAction(FPLegalNodeTypes, MVT::f16, Legal);
      setOperationAction(FPRndMode, MVT::f16,
                         Subtarget.hasStdExtZfa() ? Legal : Custom);
      setOperationAction(ISD::SELECT, MVT::f16, Custom);
      setOperationAction(ISD::IS_FPCLASS, MVT::f16, Custom);
    } else {
      setOperationAction(ZfhminZfbfminPromoteOps, MVT::f16, Promote);
      setOperationAction({ISD::STRICT_LRINT, ISD::STRICT_LLRINT,
                          ISD::STRICT_LROUND, ISD::STRICT_LLROUND},
                         MVT::f16, Legal);
      // FIXME: Need to promote f16 FCOPYSIGN to f32, but the
      // DAGCombiner::visitFP_ROUND probably needs improvements first.
      setOperationAction(ISD::FCOPYSIGN, MVT::f16, Expand);
    }
 
    setOperationAction(ISD::STRICT_FP_ROUND, MVT::f16, Legal);
    setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f32, Legal);
    setCondCodeAction(FPCCToExpand, MVT::f16, Expand);
    setOperationAction(ISD::SELECT_CC, MVT::f16, Expand);
    setOperationAction(ISD::BR_CC, MVT::f16, Expand);
 
    setOperationAction(ISD::FNEARBYINT, MVT::f16,
                       Subtarget.hasStdExtZfa() ? Legal : Promote);
    setOperationAction({ISD::FREM, ISD::FPOW, ISD::FPOWI,
                        ISD::FCOS, ISD::FSIN, ISD::FSINCOS, ISD::FEXP,
                        ISD::FEXP2, ISD::FEXP10, ISD::FLOG, ISD::FLOG2,
                        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.
    setOperationAction({ISD::STRICT_FCEIL, ISD::STRICT_FFLOOR,
                        ISD::STRICT_FNEARBYINT, ISD::STRICT_FRINT,
                        ISD::STRICT_FROUND, ISD::STRICT_FROUNDEVEN,
                        ISD::STRICT_FTRUNC},
                       MVT::f16, Promote);
 
    // We need to custom promote this.
    if (Subtarget.is64Bit())
      setOperationAction(ISD::FPOWI, MVT::i32, Custom);
 
    setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, MVT::f16,
                       Subtarget.hasStdExtZfa() ? Legal : Custom);
  }
 
  if (Subtarget.hasStdExtFOrZfinx()) {
    setOperationAction(FPLegalNodeTypes, MVT::f32, Legal);
    setOperationAction(FPRndMode, MVT::f32,
                       Subtarget.hasStdExtZfa() ? Legal : Custom);
    setCondCodeAction(FPCCToExpand, MVT::f32, Expand);
    setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
    setOperationAction(ISD::SELECT, MVT::f32, Custom);
    setOperationAction(ISD::BR_CC, MVT::f32, Expand);
    setOperationAction(FPOpToExpand, MVT::f32, Expand);
    setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
    setTruncStoreAction(MVT::f32, MVT::f16, Expand);
    setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::bf16, Expand);
    setTruncStoreAction(MVT::f32, MVT::bf16, Expand);
    setOperationAction(ISD::IS_FPCLASS, MVT::f32, Custom);
    setOperationAction(ISD::BF16_TO_FP, MVT::f32, Custom);
    setOperationAction(ISD::FP_TO_BF16, MVT::f32,
                       Subtarget.isSoftFPABI() ? LibCall : Custom);
    setOperationAction(ISD::FP_TO_FP16, MVT::f32, Custom);
    setOperationAction(ISD::FP16_TO_FP, MVT::f32, Custom);
 
    if (Subtarget.hasStdExtZfa()) {
      setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
      setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, MVT::f32, Legal);
    } else {
      setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, MVT::f32, Custom);
    }
  }
 
  if (Subtarget.hasStdExtFOrZfinx() && Subtarget.is64Bit())
    setOperationAction(ISD::BITCAST, MVT::i32, Custom);
 
  if (Subtarget.hasStdExtDOrZdinx()) {
    setOperationAction(FPLegalNodeTypes, MVT::f64, Legal);
 
    if (!Subtarget.is64Bit())
      setOperationAction(ISD::BITCAST, MVT::i64, Custom);
 
    if (Subtarget.hasStdExtZfa()) {
      setOperationAction(FPRndMode, MVT::f64, Legal);
      setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
      setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, MVT::f64, Legal);
    } else {
      if (Subtarget.is64Bit())
        setOperationAction(FPRndMode, MVT::f64, Custom);
 
      setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, MVT::f64, Custom);
    }
 
    setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f64, Legal);
    setCondCodeAction(FPCCToExpand, 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);
    setOperationAction(FPOpToExpand, MVT::f64, Expand);
    setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
    setTruncStoreAction(MVT::f64, MVT::f16, Expand);
    setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::bf16, Expand);
    setTruncStoreAction(MVT::f64, MVT::bf16, Expand);
    setOperationAction(ISD::IS_FPCLASS, MVT::f64, Custom);
    setOperationAction(ISD::BF16_TO_FP, MVT::f64, Custom);
    setOperationAction(ISD::FP_TO_BF16, MVT::f64,
                       Subtarget.isSoftFPABI() ? LibCall : Custom);
    setOperationAction(ISD::FP_TO_FP16, MVT::f64, Custom);
    setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
  }
 
  if (Subtarget.is64Bit()) {
    setOperationAction({ISD::FP_TO_UINT, ISD::FP_TO_SINT,
                        ISD::STRICT_FP_TO_UINT, ISD::STRICT_FP_TO_SINT},
                       MVT::i32, Custom);
    setOperationAction(ISD::LROUND, MVT::i32, Custom);
  }
 
  if (Subtarget.hasStdExtFOrZfinx()) {
    setOperationAction({ISD::FP_TO_UINT_SAT, ISD::FP_TO_SINT_SAT}, XLenVT,
                       Custom);
 
    setOperationAction({ISD::STRICT_FP_TO_UINT, ISD::STRICT_FP_TO_SINT,
                        ISD::STRICT_UINT_TO_FP, ISD::STRICT_SINT_TO_FP},
                       XLenVT, Legal);
 
    setOperationAction(ISD::GET_ROUNDING, XLenVT, Custom);
    setOperationAction(ISD::SET_ROUNDING, MVT::Other, Custom);
  }
 
  setOperationAction({ISD::GlobalAddress, ISD::BlockAddress, ISD::ConstantPool,
                      ISD::JumpTable},
                     XLenVT, Custom);
 
  setOperationAction(ISD::GlobalTLSAddress, XLenVT, Custom);
 
  if (Subtarget.is64Bit())
    setOperationAction(ISD::Constant, MVT::i64, Custom);
 
  // TODO: On M-mode only targets, the cycle[h]/time[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::READSTEADYCOUNTER, MVT::i64,
                     Subtarget.is64Bit() ? Legal : Custom);
 
  setOperationAction({ISD::TRAP, 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.hasStdExtZicbop()) {
    setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
  }
 
  if (Subtarget.hasStdExtA()) {
    setMaxAtomicSizeInBitsSupported(Subtarget.getXLen());
    if (Subtarget.hasStdExtZabha() && Subtarget.hasStdExtZacas())
      setMinCmpXchgSizeInBits(8);
    else
      setMinCmpXchgSizeInBits(32);
  } else if (Subtarget.hasForcedAtomics()) {
    setMaxAtomicSizeInBitsSupported(Subtarget.getXLen());
  } else {
    setMaxAtomicSizeInBitsSupported(0);
  }
 
  setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
 
  setBooleanContents(ZeroOrOneBooleanContent);
 
  if (getTargetMachine().getTargetTriple().isOSLinux()) {
    // Custom lowering of llvm.clear_cache.
    setOperationAction(ISD::CLEAR_CACHE, MVT::Other, Custom);
  }
 
  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, ISD::INTRINSIC_W_CHAIN,
                        ISD::INTRINSIC_VOID},
                       {MVT::i8, MVT::i16}, Custom);
    if (Subtarget.is64Bit())
      setOperationAction({ISD::INTRINSIC_W_CHAIN, ISD::INTRINSIC_VOID},
                         MVT::i32, Custom);
    else
      setOperationAction({ISD::INTRINSIC_WO_CHAIN, ISD::INTRINSIC_W_CHAIN},
                         MVT::i64, Custom);
 
    setOperationAction({ISD::INTRINSIC_W_CHAIN, 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_SRA,         ISD::VP_SRL,
        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,      ISD::VP_FP_TO_SINT,
        ISD::VP_FP_TO_UINT,  ISD::VP_SETCC,       ISD::VP_SIGN_EXTEND,
        ISD::VP_ZERO_EXTEND, ISD::VP_TRUNCATE,    ISD::VP_SMIN,
        ISD::VP_SMAX,        ISD::VP_UMIN,        ISD::VP_UMAX,
        ISD::VP_ABS, ISD::EXPERIMENTAL_VP_REVERSE, ISD::EXPERIMENTAL_VP_SPLICE,
        ISD::VP_SADDSAT,     ISD::VP_UADDSAT,     ISD::VP_SSUBSAT,
        ISD::VP_USUBSAT,     ISD::VP_CTTZ_ELTS,   ISD::VP_CTTZ_ELTS_ZERO_UNDEF,
        ISD::EXPERIMENTAL_VP_SPLAT};
 
    static const unsigned FloatingPointVPOps[] = {
        ISD::VP_FADD,        ISD::VP_FSUB,        ISD::VP_FMUL,
        ISD::VP_FDIV,        ISD::VP_FNEG,        ISD::VP_FABS,
        ISD::VP_FMA,         ISD::VP_REDUCE_FADD, ISD::VP_REDUCE_SEQ_FADD,
        ISD::VP_REDUCE_FMIN, ISD::VP_REDUCE_FMAX, ISD::VP_MERGE,
        ISD::VP_SELECT,      ISD::VP_SINT_TO_FP,  ISD::VP_UINT_TO_FP,
        ISD::VP_SETCC,       ISD::VP_FP_ROUND,    ISD::VP_FP_EXTEND,
        ISD::VP_SQRT,        ISD::VP_FMINNUM,     ISD::VP_FMAXNUM,
        ISD::VP_FCEIL,       ISD::VP_FFLOOR,      ISD::VP_FROUND,
        ISD::VP_FROUNDEVEN,  ISD::VP_FCOPYSIGN,   ISD::VP_FROUNDTOZERO,
        ISD::VP_FRINT,       ISD::VP_FNEARBYINT,  ISD::VP_IS_FPCLASS,
        ISD::VP_FMINIMUM,    ISD::VP_FMAXIMUM,    ISD::VP_LRINT,
        ISD::VP_LLRINT,      ISD::EXPERIMENTAL_VP_REVERSE,
        ISD::EXPERIMENTAL_VP_SPLICE, ISD::VP_REDUCE_FMINIMUM,
        ISD::VP_REDUCE_FMAXIMUM, ISD::EXPERIMENTAL_VP_SPLAT};
 
    static const unsigned IntegerVecReduceOps[] = {
        ISD::VECREDUCE_ADD,  ISD::VECREDUCE_AND,  ISD::VECREDUCE_OR,
        ISD::VECREDUCE_XOR,  ISD::VECREDUCE_SMAX, ISD::VECREDUCE_SMIN,
        ISD::VECREDUCE_UMAX, ISD::VECREDUCE_UMIN};
 
    static const unsigned FloatingPointVecReduceOps[] = {
        ISD::VECREDUCE_FADD, ISD::VECREDUCE_SEQ_FADD, ISD::VECREDUCE_FMIN,
        ISD::VECREDUCE_FMAX, ISD::VECREDUCE_FMINIMUM, ISD::VECREDUCE_FMAXIMUM};
 
    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, ISD::EXTRACT_VECTOR_ELT},
                         MVT::i64, Custom);
 
      setOperationAction(IntegerVecReduceOps, MVT::i64, Custom);
 
      setOperationAction({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},
                         MVT::i64, Custom);
    }
 
    for (MVT VT : BoolVecVTs) {
      if (!isTypeLegal(VT))
        continue;
 
      setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
 
      // Mask VTs are custom-expanded into a series of standard nodes
      setOperationAction({ISD::TRUNCATE, ISD::CONCAT_VECTORS,
                          ISD::INSERT_SUBVECTOR, ISD::EXTRACT_SUBVECTOR,
                          ISD::SCALAR_TO_VECTOR},
                         VT, Custom);
 
      setOperationAction({ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT}, VT,
                         Custom);
 
      setOperationAction(ISD::SELECT, VT, Custom);
      setOperationAction(
          {ISD::SELECT_CC, ISD::VSELECT, ISD::VP_MERGE, ISD::VP_SELECT}, VT,
          Expand);
 
      setOperationAction({ISD::VP_CTTZ_ELTS, ISD::VP_CTTZ_ELTS_ZERO_UNDEF}, VT,
                         Custom);
 
      setOperationAction({ISD::VP_AND, ISD::VP_OR, ISD::VP_XOR}, VT, Custom);
 
      setOperationAction(
          {ISD::VECREDUCE_AND, ISD::VECREDUCE_OR, ISD::VECREDUCE_XOR}, VT,
          Custom);
 
      setOperationAction(
          {ISD::VP_REDUCE_AND, ISD::VP_REDUCE_OR, 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, ISD::UINT_TO_FP, ISD::FP_TO_SINT,
                          ISD::FP_TO_UINT, ISD::STRICT_SINT_TO_FP,
                          ISD::STRICT_UINT_TO_FP, ISD::STRICT_FP_TO_SINT,
                          ISD::STRICT_FP_TO_UINT},
                         VT, Custom);
      setOperationAction({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}, 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(VT, OtherVT, Expand);
        setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, VT,
                         OtherVT, Expand);
      }
 
      setOperationAction({ISD::VP_FP_TO_SINT, ISD::VP_FP_TO_UINT,
                          ISD::VP_TRUNCATE, ISD::VP_SETCC},
                         VT, Custom);
 
      setOperationAction(ISD::VECTOR_DEINTERLEAVE, VT, Custom);
      setOperationAction(ISD::VECTOR_INTERLEAVE, VT, Custom);
 
      setOperationAction(ISD::VECTOR_REVERSE, VT, Custom);
 
      setOperationAction(ISD::EXPERIMENTAL_VP_SPLICE, VT, Custom);
      setOperationAction(ISD::EXPERIMENTAL_VP_REVERSE, VT, Custom);
 
      setOperationPromotedToType(
          ISD::VECTOR_SPLICE, VT,
          MVT::getVectorVT(MVT::i8, VT.getVectorElementCount()));
    }
 
    for (MVT VT : IntVecVTs) {
      if (!isTypeLegal(VT))
        continue;
 
      setOperationAction(ISD::SPLAT_VECTOR, VT, Legal);
      setOperationAction(ISD::SPLAT_VECTOR_PARTS, VT, Custom);
 
      // Vectors implement MULHS/MULHU.
      setOperationAction({ISD::SMUL_LOHI, 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, ISD::MULHS}, VT, Expand);
 
      setOperationAction({ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX}, VT,
                         Legal);
 
      setOperationAction({ISD::ABDS, ISD::ABDU}, VT, Custom);
 
      // Custom-lower extensions and truncations from/to mask types.
      setOperationAction({ISD::ANY_EXTEND, ISD::SIGN_EXTEND, 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, ISD::UINT_TO_FP, ISD::FP_TO_SINT,
                          ISD::FP_TO_UINT, ISD::STRICT_SINT_TO_FP,
                          ISD::STRICT_UINT_TO_FP, ISD::STRICT_FP_TO_SINT,
                          ISD::STRICT_FP_TO_UINT},
                         VT, Custom);
      setOperationAction({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}, VT,
                         Custom);
      setOperationAction({ISD::AVGFLOORS, ISD::AVGFLOORU, ISD::AVGCEILS,
                          ISD::AVGCEILU, ISD::SADDSAT, ISD::UADDSAT,
                          ISD::SSUBSAT, 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, ISD::EXTRACT_VECTOR_ELT}, VT,
                         Custom);
 
      // Custom-lower reduction operations to set up the corresponding custom
      // nodes' operands.
      setOperationAction(IntegerVecReduceOps, VT, Custom);
 
      setOperationAction(IntegerVPOps, VT, Custom);
 
      setOperationAction({ISD::LOAD, ISD::STORE}, VT, Custom);
 
      setOperationAction({ISD::MLOAD, ISD::MSTORE, ISD::MGATHER, ISD::MSCATTER},
                         VT, Custom);
 
      setOperationAction(
          {ISD::VP_LOAD, ISD::VP_STORE, ISD::EXPERIMENTAL_VP_STRIDED_LOAD,
           ISD::EXPERIMENTAL_VP_STRIDED_STORE, ISD::VP_GATHER, ISD::VP_SCATTER},
          VT, Custom);
 
      setOperationAction({ISD::CONCAT_VECTORS, ISD::INSERT_SUBVECTOR,
                          ISD::EXTRACT_SUBVECTOR, ISD::SCALAR_TO_VECTOR},
                         VT, Custom);
 
      setOperationAction(ISD::SELECT, VT, Custom);
      setOperationAction(ISD::SELECT_CC, VT, Expand);
 
      setOperationAction({ISD::STEP_VECTOR, ISD::VECTOR_REVERSE}, VT, Custom);
 
      for (MVT OtherVT : MVT::integer_scalable_vector_valuetypes()) {
        setTruncStoreAction(VT, OtherVT, Expand);
        setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, VT,
                         OtherVT, Expand);
      }
 
      setOperationAction(ISD::VECTOR_DEINTERLEAVE, VT, Custom);
      setOperationAction(ISD::VECTOR_INTERLEAVE, VT, Custom);
 
      // Splice
      setOperationAction(ISD::VECTOR_SPLICE, VT, Custom);
 
      if (Subtarget.hasStdExtZvkb()) {
        setOperationAction(ISD::BSWAP, VT, Legal);
        setOperationAction(ISD::VP_BSWAP, VT, Custom);
      } else {
        setOperationAction({ISD::BSWAP, ISD::VP_BSWAP}, VT, Expand);
        setOperationAction({ISD::ROTL, ISD::ROTR}, VT, Expand);
      }
 
      if (Subtarget.hasStdExtZvbb()) {
        setOperationAction(ISD::BITREVERSE, VT, Legal);
        setOperationAction(ISD::VP_BITREVERSE, VT, Custom);
        setOperationAction({ISD::VP_CTLZ, ISD::VP_CTLZ_ZERO_UNDEF, ISD::VP_CTTZ,
                            ISD::VP_CTTZ_ZERO_UNDEF, ISD::VP_CTPOP},
                           VT, Custom);
      } else {
        setOperationAction({ISD::BITREVERSE, ISD::VP_BITREVERSE}, VT, Expand);
        setOperationAction({ISD::CTLZ, ISD::CTTZ, ISD::CTPOP}, VT, Expand);
        setOperationAction({ISD::VP_CTLZ, ISD::VP_CTLZ_ZERO_UNDEF, ISD::VP_CTTZ,
                            ISD::VP_CTTZ_ZERO_UNDEF, ISD::VP_CTPOP},
                           VT, Expand);
 
        // Lower CTLZ_ZERO_UNDEF and CTTZ_ZERO_UNDEF if element of VT in the
        // range of f32.
        EVT FloatVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
        if (isTypeLegal(FloatVT)) {
          setOperationAction({ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF,
                              ISD::CTTZ_ZERO_UNDEF, ISD::VP_CTLZ,
                              ISD::VP_CTLZ_ZERO_UNDEF, ISD::VP_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,
    };
 
    // TODO: support more ops.
    static const unsigned ZvfhminPromoteOps[] = {
        ISD::FMINNUM,     ISD::FMAXNUM,      ISD::FADD,        ISD::FSUB,
        ISD::FMUL,        ISD::FMA,          ISD::FDIV,        ISD::FSQRT,
        ISD::FABS,        ISD::FNEG,         ISD::FCOPYSIGN,   ISD::FCEIL,
        ISD::FFLOOR,      ISD::FROUND,       ISD::FROUNDEVEN,  ISD::FRINT,
        ISD::FNEARBYINT,  ISD::IS_FPCLASS,   ISD::SETCC,       ISD::FMAXIMUM,
        ISD::FMINIMUM,    ISD::STRICT_FADD,  ISD::STRICT_FSUB, ISD::STRICT_FMUL,
        ISD::STRICT_FDIV, ISD::STRICT_FSQRT, ISD::STRICT_FMA};
 
    // TODO: support more vp ops.
    static const unsigned ZvfhminPromoteVPOps[] = {
        ISD::VP_FADD,        ISD::VP_FSUB,         ISD::VP_FMUL,
        ISD::VP_FDIV,        ISD::VP_FNEG,         ISD::VP_FABS,
        ISD::VP_FMA,         ISD::VP_REDUCE_FADD,  ISD::VP_REDUCE_SEQ_FADD,
        ISD::VP_REDUCE_FMIN, ISD::VP_REDUCE_FMAX,  ISD::VP_SQRT,
        ISD::VP_FMINNUM,     ISD::VP_FMAXNUM,      ISD::VP_FCEIL,
        ISD::VP_FFLOOR,      ISD::VP_FROUND,       ISD::VP_FROUNDEVEN,
        ISD::VP_FCOPYSIGN,   ISD::VP_FROUNDTOZERO, ISD::VP_FRINT,
        ISD::VP_FNEARBYINT,  ISD::VP_SETCC,        ISD::VP_FMINIMUM,
        ISD::VP_FMAXIMUM,    ISD::VP_REDUCE_FMINIMUM, ISD::VP_REDUCE_FMAXIMUM};
 
    // 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, ISD::FP_EXTEND}, VT, Custom);
      setOperationAction({ISD::LRINT, ISD::LLRINT}, VT, Custom);
      // Custom-lower insert/extract operations to simplify patterns.
      setOperationAction({ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT}, VT,
                         Custom);
      // Expand various condition codes (explained above).
      setCondCodeAction(VFPCCToExpand, VT, Expand);
 
      setOperationAction({ISD::FMINNUM, ISD::FMAXNUM}, VT, Legal);
      setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, VT, Custom);
 
      setOperationAction({ISD::FTRUNC, ISD::FCEIL, ISD::FFLOOR, ISD::FROUND,
                          ISD::FROUNDEVEN, ISD::FRINT, ISD::FNEARBYINT,
                          ISD::IS_FPCLASS},
                         VT, Custom);
 
      setOperationAction(FloatingPointVecReduceOps, VT, Custom);
 
      // Expand FP operations that need libcalls.
      setOperationAction(ISD::FREM, VT, Expand);
      setOperationAction(ISD::FPOW, VT, Expand);
      setOperationAction(ISD::FCOS, VT, Expand);
      setOperationAction(ISD::FSIN, VT, Expand);
      setOperationAction(ISD::FSINCOS, VT, Expand);
      setOperationAction(ISD::FEXP, VT, Expand);
      setOperationAction(ISD::FEXP2, VT, Expand);
      setOperationAction(ISD::FEXP10, VT, Expand);
      setOperationAction(ISD::FLOG, VT, Expand);
      setOperationAction(ISD::FLOG2, VT, Expand);
      setOperationAction(ISD::FLOG10, VT, Expand);
 
      setOperationAction(ISD::FCOPYSIGN, VT, Legal);
 
      setOperationAction({ISD::LOAD, ISD::STORE}, VT, Custom);
 
      setOperationAction({ISD::MLOAD, ISD::MSTORE, ISD::MGATHER, ISD::MSCATTER},
                         VT, Custom);
 
      setOperationAction(
          {ISD::VP_LOAD, ISD::VP_STORE, ISD::EXPERIMENTAL_VP_STRIDED_LOAD,
           ISD::EXPERIMENTAL_VP_STRIDED_STORE, ISD::VP_GATHER, ISD::VP_SCATTER},
          VT, Custom);
 
      setOperationAction(ISD::SELECT, VT, Custom);
      setOperationAction(ISD::SELECT_CC, VT, Expand);
 
      setOperationAction({ISD::CONCAT_VECTORS, ISD::INSERT_SUBVECTOR,
                          ISD::EXTRACT_SUBVECTOR, ISD::SCALAR_TO_VECTOR},
                         VT, Custom);
 
      setOperationAction(ISD::VECTOR_DEINTERLEAVE, VT, Custom);
      setOperationAction(ISD::VECTOR_INTERLEAVE, VT, Custom);
 
      setOperationAction({ISD::VECTOR_REVERSE, ISD::VECTOR_SPLICE}, VT, Custom);
 
      setOperationAction(FloatingPointVPOps, VT, Custom);
 
      setOperationAction({ISD::STRICT_FP_EXTEND, ISD::STRICT_FP_ROUND}, VT,
                         Custom);
      setOperationAction({ISD::STRICT_FADD, ISD::STRICT_FSUB, ISD::STRICT_FMUL,
                          ISD::STRICT_FDIV, ISD::STRICT_FSQRT, ISD::STRICT_FMA},
                         VT, Legal);
      setOperationAction({ISD::STRICT_FSETCC, ISD::STRICT_FSETCCS,
                          ISD::STRICT_FTRUNC, ISD::STRICT_FCEIL,
                          ISD::STRICT_FFLOOR, ISD::STRICT_FROUND,
                          ISD::STRICT_FROUNDEVEN, ISD::STRICT_FNEARBYINT},
                         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) {
        if (!isTypeLegal(VT))
          continue;
        SetCommonVFPActions(VT);
      }
    } else if (Subtarget.hasVInstructionsF16Minimal()) {
      for (MVT VT : F16VecVTs) {
        if (!isTypeLegal(VT))
          continue;
        setOperationAction({ISD::FP_ROUND, ISD::FP_EXTEND}, VT, Custom);
        setOperationAction({ISD::STRICT_FP_ROUND, ISD::STRICT_FP_EXTEND}, VT,
                           Custom);
        setOperationAction({ISD::VP_FP_ROUND, ISD::VP_FP_EXTEND}, VT, Custom);
        setOperationAction({ISD::VP_MERGE, ISD::VP_SELECT, ISD::SELECT}, VT,
                           Custom);
        setOperationAction(ISD::SELECT_CC, VT, Expand);
        setOperationAction({ISD::SINT_TO_FP, ISD::UINT_TO_FP,
                            ISD::VP_SINT_TO_FP, ISD::VP_UINT_TO_FP},
                           VT, Custom);
        setOperationAction({ISD::CONCAT_VECTORS, ISD::INSERT_SUBVECTOR,
                            ISD::EXTRACT_SUBVECTOR},
                           VT, Custom);
        if (Subtarget.hasStdExtZfhmin())
          setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
        // load/store
        setOperationAction({ISD::LOAD, ISD::STORE}, VT, Custom);
 
        // Custom split nxv32f16 since nxv32f32 if not legal.
        if (VT == MVT::nxv32f16) {
          setOperationAction(ZvfhminPromoteOps, VT, Custom);
          setOperationAction(ZvfhminPromoteVPOps, VT, Custom);
          continue;
        }
        // Add more promote ops.
        MVT F32VecVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
        setOperationPromotedToType(ZvfhminPromoteOps, VT, F32VecVT);
        setOperationPromotedToType(ZvfhminPromoteVPOps, VT, F32VecVT);
      }
    }
 
    // TODO: Could we merge some code with zvfhmin?
    if (Subtarget.hasVInstructionsBF16Minimal()) {
      for (MVT VT : BF16VecVTs) {
        if (!isTypeLegal(VT))
          continue;
        setOperationAction({ISD::FP_ROUND, ISD::FP_EXTEND}, VT, Custom);
        setOperationAction({ISD::VP_FP_ROUND, ISD::VP_FP_EXTEND}, VT, Custom);
        setOperationAction({ISD::STRICT_FP_ROUND, ISD::STRICT_FP_EXTEND}, VT,
                           Custom);
        setOperationAction({ISD::CONCAT_VECTORS, ISD::INSERT_SUBVECTOR,
                            ISD::EXTRACT_SUBVECTOR},
                           VT, Custom);
        setOperationAction({ISD::LOAD, ISD::STORE}, VT, Custom);
        if (Subtarget.hasStdExtZfbfmin())
          setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
        setOperationAction({ISD::VP_MERGE, ISD::VP_SELECT, ISD::SELECT}, VT,
                           Custom);
        setOperationAction(ISD::SELECT_CC, VT, Expand);
        // TODO: Promote to fp32.
      }
    }
 
    if (Subtarget.hasVInstructionsF32()) {
      for (MVT VT : F32VecVTs) {
        if (!isTypeLegal(VT))
          continue;
        SetCommonVFPActions(VT);
        SetCommonVFPExtLoadTruncStoreActions(VT, F16VecVTs);
      }
    }
 
    if (Subtarget.hasVInstructionsF64()) {
      for (MVT VT : F64VecVTs) {
        if (!isTypeLegal(VT))
          continue;
        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, ISD::SEXTLOAD, ISD::ZEXTLOAD}, VT,
                           OtherVT, Expand);
        }
 
        // Custom lower fixed vector undefs to scalable vector undefs to avoid
        // expansion to a build_vector of 0s.
        setOperationAction(ISD::UNDEF, VT, Custom);
 
        // We use EXTRACT_SUBVECTOR as a "cast" from scalable to fixed.
        setOperationAction({ISD::INSERT_SUBVECTOR, ISD::EXTRACT_SUBVECTOR}, VT,
                           Custom);
 
        setOperationAction({ISD::BUILD_VECTOR, ISD::CONCAT_VECTORS}, VT,
                           Custom);
 
        setOperationAction({ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT},
                           VT, Custom);
 
        setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
 
        setOperationAction({ISD::LOAD, 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, ISD::VECREDUCE_OR, ISD::VECREDUCE_XOR}, VT,
            Custom);
 
        setOperationAction(
            {ISD::VP_REDUCE_AND, ISD::VP_REDUCE_OR, ISD::VP_REDUCE_XOR}, VT,
            Custom);
 
        setOperationAction(
            {
                ISD::SINT_TO_FP,
                ISD::UINT_TO_FP,
                ISD::FP_TO_SINT,
                ISD::FP_TO_UINT,
                ISD::STRICT_SINT_TO_FP,
                ISD::STRICT_UINT_TO_FP,
                ISD::STRICT_FP_TO_SINT,
                ISD::STRICT_FP_TO_UINT,
            },
            VT, Custom);
        setOperationAction({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}, VT,
                           Custom);
 
        setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
 
        // Operations below are different for between masks and other vectors.
        if (VT.getVectorElementType() == MVT::i1) {
          setOperationAction({ISD::VP_AND, ISD::VP_OR, ISD::VP_XOR, ISD::AND,
                              ISD::OR, ISD::XOR},
                             VT, Custom);
 
          setOperationAction({ISD::VP_FP_TO_SINT, ISD::VP_FP_TO_UINT,
                              ISD::VP_SETCC, ISD::VP_TRUNCATE},
                             VT, Custom);
 
          setOperationAction(ISD::EXPERIMENTAL_VP_SPLICE, VT, Custom);
          setOperationAction(ISD::EXPERIMENTAL_VP_REVERSE, VT, Custom);
          continue;
        }
 
        // Make SPLAT_VECTOR Legal so DAGCombine will convert splat vectors to
        // it before type legalization for i64 vectors on RV32. It will then be
        // type legalized to SPLAT_VECTOR_PARTS which we need to Custom handle.
        // FIXME: Use SPLAT_VECTOR for all types? DAGCombine probably needs
        // improvements first.
        if (!Subtarget.is64Bit() && VT.getVectorElementType() == MVT::i64) {
          setOperationAction(ISD::SPLAT_VECTOR, VT, Legal);
          setOperationAction(ISD::SPLAT_VECTOR_PARTS, VT, Custom);
        }
 
        setOperationAction(
            {ISD::MLOAD, ISD::MSTORE, ISD::MGATHER, ISD::MSCATTER}, VT, Custom);
 
        setOperationAction({ISD::VP_LOAD, ISD::VP_STORE,
                            ISD::EXPERIMENTAL_VP_STRIDED_LOAD,
                            ISD::EXPERIMENTAL_VP_STRIDED_STORE, ISD::VP_GATHER,
                            ISD::VP_SCATTER},
                           VT, Custom);
 
        setOperationAction({ISD::ADD, ISD::MUL, ISD::SUB, ISD::AND, ISD::OR,
                            ISD::XOR, ISD::SDIV, ISD::SREM, ISD::UDIV,
                            ISD::UREM, ISD::SHL, ISD::SRA, ISD::SRL},
                           VT, Custom);
 
        setOperationAction(
            {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX, ISD::ABS}, VT, Custom);
 
        setOperationAction({ISD::ABDS, ISD::ABDU}, VT, Custom);
 
        // vXi64 MULHS/MULHU requires the V extension instead of Zve64*.
        if (VT.getVectorElementType() != MVT::i64 || Subtarget.hasStdExtV())
          setOperationAction({ISD::MULHS, ISD::MULHU}, VT, Custom);
 
        setOperationAction({ISD::AVGFLOORS, ISD::AVGFLOORU, ISD::AVGCEILS,
                            ISD::AVGCEILU, ISD::SADDSAT, ISD::UADDSAT,
                            ISD::SSUBSAT, ISD::USUBSAT},
                           VT, Custom);
 
        setOperationAction(ISD::VSELECT, VT, Custom);
 
        setOperationAction(
            {ISD::ANY_EXTEND, ISD::SIGN_EXTEND, ISD::ZERO_EXTEND}, VT, Custom);
 
        // Custom-lower reduction operations to set up the corresponding custom
        // nodes' operands.
        setOperationAction({ISD::VECREDUCE_ADD, ISD::VECREDUCE_SMAX,
                            ISD::VECREDUCE_SMIN, ISD::VECREDUCE_UMAX,
                            ISD::VECREDUCE_UMIN},
                           VT, Custom);
 
        setOperationAction(IntegerVPOps, VT, Custom);
 
        if (Subtarget.hasStdExtZvkb())
          setOperationAction({ISD::BSWAP, ISD::ROTL, ISD::ROTR}, VT, Custom);
 
        if (Subtarget.hasStdExtZvbb()) {
          setOperationAction({ISD::BITREVERSE, ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF,
                              ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF, ISD::CTPOP},
                             VT, Custom);
        } else {
          // Lower CTLZ_ZERO_UNDEF and CTTZ_ZERO_UNDEF if element of VT in the
          // range of f32.
          EVT FloatVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
          if (isTypeLegal(FloatVT))
            setOperationAction(
                {ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF, ISD::CTTZ_ZERO_UNDEF}, VT,
                Custom);
        }
      }
 
      for (MVT VT : MVT::fp_fixedlen_vector_valuetypes()) {
        // There are no extending loads or truncating stores.
        for (MVT InnerVT : MVT::fp_fixedlen_vector_valuetypes()) {
          setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
          setTruncStoreAction(VT, InnerVT, Expand);
        }
 
        if (!useRVVForFixedLengthVectorVT(VT))
          continue;
 
        // By default everything must be expanded.
        for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op)
          setOperationAction(Op, VT, Expand);
 
        // Custom lower fixed vector undefs to scalable vector undefs to avoid
        // expansion to a build_vector of 0s.
        setOperationAction(ISD::UNDEF, VT, Custom);
 
        setOperationAction({ISD::CONCAT_VECTORS, ISD::INSERT_SUBVECTOR,
                            ISD::EXTRACT_SUBVECTOR},
                           VT, Custom);
 
        // FIXME: mload, mstore, mgather, mscatter, vp_load/store,
        // vp_stride_load/store, vp_gather/scatter can be hoisted to here.
        setOperationAction({ISD::LOAD, ISD::STORE}, VT, Custom);
 
        setOperationAction({ISD::FP_ROUND, ISD::FP_EXTEND}, VT, Custom);
        setOperationAction({ISD::STRICT_FP_ROUND, ISD::STRICT_FP_EXTEND}, VT,
                           Custom);
 
        if (VT.getVectorElementType() == MVT::f16 &&
            !Subtarget.hasVInstructionsF16()) {
          setOperationAction({ISD::VP_FP_ROUND, ISD::VP_FP_EXTEND}, VT, Custom);
          setOperationAction(
              {ISD::VP_MERGE, ISD::VP_SELECT, ISD::VSELECT, ISD::SELECT}, VT,
              Custom);
          setOperationAction({ISD::SINT_TO_FP, ISD::UINT_TO_FP,
                              ISD::VP_SINT_TO_FP, ISD::VP_UINT_TO_FP},
                             VT, Custom);
          setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
          if (Subtarget.hasStdExtZfhmin()) {
            // FIXME: We should prefer BUILD_VECTOR over SPLAT_VECTOR.
            setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
          } else {
            // We need to custom legalize f16 build vectors if Zfhmin isn't
            // available.
            setOperationAction(ISD::BUILD_VECTOR, MVT::f16, Custom);
          }
          MVT F32VecVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount());
          // Don't promote f16 vector operations to f32 if f32 vector type is
          // not legal.
          // TODO: could split the f16 vector into two vectors and do promotion.
          if (!isTypeLegal(F32VecVT))
            continue;
          setOperationPromotedToType(ZvfhminPromoteOps, VT, F32VecVT);
          setOperationPromotedToType(ZvfhminPromoteVPOps, VT, F32VecVT);
          continue;
        }
 
        if (VT.getVectorElementType() == MVT::bf16) {
          setOperationAction({ISD::VP_FP_ROUND, ISD::VP_FP_EXTEND}, VT, Custom);
          // FIXME: We should prefer BUILD_VECTOR over SPLAT_VECTOR.
          setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
          setOperationAction(
              {ISD::VP_MERGE, ISD::VP_SELECT, ISD::VSELECT, ISD::SELECT}, VT,
              Custom);
          // TODO: Promote to fp32.
          continue;
        }
 
        setOperationAction({ISD::BUILD_VECTOR, ISD::VECTOR_SHUFFLE,
                            ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT},
                           VT, Custom);
 
        setOperationAction(
            {ISD::MLOAD, ISD::MSTORE, ISD::MGATHER, ISD::MSCATTER}, VT, Custom);
 
        setOperationAction({ISD::VP_LOAD, ISD::VP_STORE,
                            ISD::EXPERIMENTAL_VP_STRIDED_LOAD,
                            ISD::EXPERIMENTAL_VP_STRIDED_STORE, ISD::VP_GATHER,
                            ISD::VP_SCATTER},
                           VT, Custom);
 
        setOperationAction({ISD::FADD, ISD::FSUB, ISD::FMUL, ISD::FDIV,
                            ISD::FNEG, ISD::FABS, ISD::FCOPYSIGN, ISD::FSQRT,
                            ISD::FMA, ISD::FMINNUM, ISD::FMAXNUM,
                            ISD::IS_FPCLASS, ISD::FMAXIMUM, ISD::FMINIMUM},
                           VT, Custom);
 
        setOperationAction({ISD::FTRUNC, ISD::FCEIL, ISD::FFLOOR, ISD::FROUND,
                            ISD::FROUNDEVEN, ISD::FRINT, ISD::FNEARBYINT},
                           VT, Custom);
 
        setCondCodeAction(VFPCCToExpand, VT, Expand);
 
        setOperationAction(ISD::SETCC, VT, Custom);
        setOperationAction({ISD::VSELECT, ISD::SELECT}, VT, Custom);
 
        setOperationAction(ISD::BITCAST, VT, Custom);
 
        setOperationAction(FloatingPointVecReduceOps, VT, Custom);
 
        setOperationAction(FloatingPointVPOps, VT, Custom);
 
        setOperationAction(
            {ISD::STRICT_FADD, ISD::STRICT_FSUB, ISD::STRICT_FMUL,
             ISD::STRICT_FDIV, ISD::STRICT_FSQRT, ISD::STRICT_FMA,
             ISD::STRICT_FSETCC, ISD::STRICT_FSETCCS, ISD::STRICT_FTRUNC,
             ISD::STRICT_FCEIL, ISD::STRICT_FFLOOR, ISD::STRICT_FROUND,
             ISD::STRICT_FROUNDEVEN, ISD::STRICT_FNEARBYINT},
            VT, Custom);
      }
 
      // Custom-legalize bitcasts from fixed-length vectors to scalar types.
      setOperationAction(ISD::BITCAST, {MVT::i8, MVT::i16, MVT::i32, MVT::i64},
                         Custom);
      if (Subtarget.hasStdExtZfhminOrZhinxmin())
        setOperationAction(ISD::BITCAST, MVT::f16, Custom);
      if (Subtarget.hasStdExtFOrZfinx())
        setOperationAction(ISD::BITCAST, MVT::f32, Custom);
      if (Subtarget.hasStdExtDOrZdinx())
        setOperationAction(ISD::BITCAST, MVT::f64, Custom);
    }
  }
 
  if (Subtarget.hasStdExtA())
    setOperationAction(ISD::ATOMIC_LOAD_SUB, XLenVT, Expand);
 
  if (Subtarget.hasForcedAtomics()) {
    // Force __sync libcalls to be emitted for atomic rmw/cas operations.
    setOperationAction(
        {ISD::ATOMIC_CMP_SWAP, ISD::ATOMIC_SWAP, ISD::ATOMIC_LOAD_ADD,
         ISD::ATOMIC_LOAD_SUB, ISD::ATOMIC_LOAD_AND, ISD::ATOMIC_LOAD_OR,
         ISD::ATOMIC_LOAD_XOR, ISD::ATOMIC_LOAD_NAND, ISD::ATOMIC_LOAD_MIN,
         ISD::ATOMIC_LOAD_MAX, ISD::ATOMIC_LOAD_UMIN, ISD::ATOMIC_LOAD_UMAX},
        XLenVT, LibCall);
  }
 
  if (Subtarget.hasVendorXTHeadMemIdx()) {
    for (unsigned im : {ISD::PRE_INC, ISD::POST_INC}) {
      setIndexedLoadAction(im, MVT::i8, Legal);
      setIndexedStoreAction(im, MVT::i8, Legal);
      setIndexedLoadAction(im, MVT::i16, Legal);
      setIndexedStoreAction(im, MVT::i16, Legal);
      setIndexedLoadAction(im, MVT::i32, Legal);
      setIndexedStoreAction(im, MVT::i32, Legal);
 
      if (Subtarget.is64Bit()) {
        setIndexedLoadAction(im, MVT::i64, Legal);
        setIndexedStoreAction(im, MVT::i64, Legal);
      }
    }
  }
 
  if (Subtarget.hasVendorXCVmem() && !Subtarget.is64Bit()) {
    setIndexedLoadAction(ISD::POST_INC, MVT::i8, Legal);
    setIndexedLoadAction(ISD::POST_INC, MVT::i16, Legal);
    setIndexedLoadAction(ISD::POST_INC, MVT::i32, Legal);
 
    setIndexedStoreAction(ISD::POST_INC, MVT::i8, Legal);
    setIndexedStoreAction(ISD::POST_INC, MVT::i16, Legal);
    setIndexedStoreAction(ISD::POST_INC, MVT::i32, Legal);
  }
 
  // Function alignments.
  const Align FunctionAlignment(Subtarget.hasStdExtCOrZca() ? 2 : 4);
  setMinFunctionAlignment(FunctionAlignment);
  // Set preferred alignments.
  setPrefFunctionAlignment(Subtarget.getPrefFunctionAlignment());
  setPrefLoopAlignment(Subtarget.getPrefLoopAlignment());
 
  setTargetDAGCombine({ISD::INTRINSIC_VOID, ISD::INTRINSIC_W_CHAIN,
                       ISD::INTRINSIC_WO_CHAIN, ISD::ADD, ISD::SUB, ISD::MUL,
                       ISD::AND, ISD::OR, ISD::XOR, ISD::SETCC, ISD::SELECT});
  setTargetDAGCombine(ISD::SRA);
 
  if (Subtarget.hasStdExtFOrZfinx())
    setTargetDAGCombine({ISD::FADD, ISD::FMAXNUM, ISD::FMINNUM, ISD::FMUL});
 
  if (Subtarget.hasStdExtZbb())
    setTargetDAGCombine({ISD::UMAX, ISD::UMIN, ISD::SMAX, ISD::SMIN});
 
  if ((Subtarget.hasStdExtZbs() && Subtarget.is64Bit()) ||
      Subtarget.hasVInstructions())
    setTargetDAGCombine(ISD::TRUNCATE);
 
  if (Subtarget.hasStdExtZbkb())
    setTargetDAGCombine(ISD::BITREVERSE);
  if (Subtarget.hasStdExtZfhminOrZhinxmin())
    setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
  if (Subtarget.hasStdExtFOrZfinx())
    setTargetDAGCombine({ISD::ZERO_EXTEND, ISD::FP_TO_SINT, ISD::FP_TO_UINT,
                         ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT});
  if (Subtarget.hasVInstructions())
    setTargetDAGCombine({ISD::FCOPYSIGN, ISD::MGATHER, ISD::MSCATTER,
                         ISD::VP_GATHER, ISD::VP_SCATTER, ISD::SRA, ISD::SRL,
                         ISD::SHL, ISD::STORE, ISD::SPLAT_VECTOR,
                         ISD::BUILD_VECTOR, ISD::CONCAT_VECTORS,
                         ISD::EXPERIMENTAL_VP_REVERSE, ISD::MUL,
                         ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM,
                         ISD::INSERT_VECTOR_ELT, ISD::ABS});
  if (Subtarget.hasVendorXTHeadMemPair())
    setTargetDAGCombine({ISD::LOAD, ISD::STORE});
  if (Subtarget.useRVVForFixedLengthVectors())
    setTargetDAGCombine(ISD::BITCAST);
 
  setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
  setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
 
  // Disable strict node mutation.
  IsStrictFPEnabled = true;
 
  // Let the subtarget decide if a predictable select is more expensive than the
  // corresponding branch. This information is used in CGP/SelectOpt to decide
  // when to convert selects into branches.
  PredictableSelectIsExpensive = Subtarget.predictableSelectIsExpensive();
}
 
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();
}
 
// Return false if we can lower get_vector_length to a vsetvli intrinsic.
bool RISCVTargetLowering::shouldExpandGetVectorLength(EVT TripCountVT,
                                                      unsigned VF,
                                                      bool IsScalable) const {
  if (!Subtarget.hasVInstructions())
    return true;
 
  if (!IsScalable)
    return true;
 
  if (TripCountVT != MVT::i32 && TripCountVT != Subtarget.getXLenVT())
    return true;
 
  // Don't allow VF=1 if those types are't legal.
  if (VF < RISCV::RVVBitsPerBlock / Subtarget.getELen())
    return true;
 
  // VLEN=32 support is incomplete.
  if (Subtarget.getRealMinVLen() < RISCV::RVVBitsPerBlock)
    return true;
 
  // The maximum VF is for the smallest element width with LMUL=8.
  // VF must be a power of 2.
  unsigned MaxVF = (RISCV::RVVBitsPerBlock / 8) * 8;
  return VF > MaxVF || !isPowerOf2_32(VF);
}
 
bool RISCVTargetLowering::shouldExpandCttzElements(EVT VT) const {
  return !Subtarget.hasVInstructions() ||
         VT.getVectorElementType() != MVT::i1 || !isTypeLegal(VT);
}
 
bool RISCVTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
                                             const CallInst &I,
                                             MachineFunction &MF,
                                             unsigned Intrinsic) const {
  auto &DL = I.getDataLayout();
 
  auto SetRVVLoadStoreInfo = [&](unsigned PtrOp, bool IsStore,
                                 bool IsUnitStrided, bool UsePtrVal = false) {
    Info.opc = IsStore ? ISD::INTRINSIC_VOID : ISD::INTRINSIC_W_CHAIN;
    // We can't use ptrVal if the intrinsic can access memory before the
    // pointer. This means we can't use it for strided or indexed intrinsics.
    if (UsePtrVal)
      Info.ptrVal = I.getArgOperand(PtrOp);
    else
      Info.fallbackAddressSpace =
          I.getArgOperand(PtrOp)->getType()->getPointerAddressSpace();
    Type *MemTy;
    if (IsStore) {
      // Store value is the first operand.
      MemTy = I.getArgOperand(0)->getType();
    } else {
      // Use return type. If it's segment load, return type is a struct.
      MemTy = I.getType();
      if (MemTy->isStructTy())
        MemTy = MemTy->getStructElementType(0);
    }
    if (!IsUnitStrided)
      MemTy = MemTy->getScalarType();
 
    Info.memVT = getValueType(DL, MemTy);
    Info.align = Align(DL.getTypeSizeInBits(MemTy->getScalarType()) / 8);
    Info.size = MemoryLocation::UnknownSize;
    Info.flags |=
        IsStore ? MachineMemOperand::MOStore : MachineMemOperand::MOLoad;
    return true;
  };
 
  if (I.hasMetadata(LLVMContext::MD_nontemporal))
    Info.flags |= MachineMemOperand::MONonTemporal;
 
  Info.flags |= RISCVTargetLowering::getTargetMMOFlags(I);
  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:
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    Info.memVT = MVT::i32;
    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_seg2_load:
  case Intrinsic::riscv_seg3_load:
  case Intrinsic::riscv_seg4_load:
  case Intrinsic::riscv_seg5_load:
  case Intrinsic::riscv_seg6_load:
  case Intrinsic::riscv_seg7_load:
  case Intrinsic::riscv_seg8_load:
    return SetRVVLoadStoreInfo(/*PtrOp*/ 0, /*IsStore*/ false,
                               /*IsUnitStrided*/ false, /*UsePtrVal*/ true);
  case Intrinsic::riscv_seg2_store:
  case Intrinsic::riscv_seg3_store:
  case Intrinsic::riscv_seg4_store:
  case Intrinsic::riscv_seg5_store:
  case Intrinsic::riscv_seg6_store:
  case Intrinsic::riscv_seg7_store:
  case Intrinsic::riscv_seg8_store:
    // Operands are (vec, ..., vec, ptr, vl)
    return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 2,
                               /*IsStore*/ true,
                               /*IsUnitStrided*/ false, /*UsePtrVal*/ true);
  case Intrinsic::riscv_vle:
  case Intrinsic::riscv_vle_mask:
  case Intrinsic::riscv_vleff:
  case Intrinsic::riscv_vleff_mask:
    return SetRVVLoadStoreInfo(/*PtrOp*/ 1,
                               /*IsStore*/ false,
                               /*IsUnitStrided*/ true,
                               /*UsePtrVal*/ true);
  case Intrinsic::riscv_vse:
  case Intrinsic::riscv_vse_mask:
    return SetRVVLoadStoreInfo(/*PtrOp*/ 1,
                               /*IsStore*/ true,
                               /*IsUnitStrided*/ true,
                               /*UsePtrVal*/ true);
  case Intrinsic::riscv_vlse:
  case Intrinsic::riscv_vlse_mask:
  case Intrinsic::riscv_vloxei:
  case Intrinsic::riscv_vloxei_mask:
  case Intrinsic::riscv_vluxei:
  case Intrinsic::riscv_vluxei_mask:
    return SetRVVLoadStoreInfo(/*PtrOp*/ 1,
                               /*IsStore*/ false,
                               /*IsUnitStrided*/ false);
  case Intrinsic::riscv_vsse:
  case Intrinsic::riscv_vsse_mask:
  case Intrinsic::riscv_vsoxei:
  case Intrinsic::riscv_vsoxei_mask:
  case Intrinsic::riscv_vsuxei:
  case Intrinsic::riscv_vsuxei_mask:
    return SetRVVLoadStoreInfo(/*PtrOp*/ 1,
                               /*IsStore*/ true,
                               /*IsUnitStrided*/ false);
  case Intrinsic::riscv_vlseg2:
  case Intrinsic::riscv_vlseg3:
  case Intrinsic::riscv_vlseg4:
  case Intrinsic::riscv_vlseg5:
  case Intrinsic::riscv_vlseg6:
  case Intrinsic::riscv_vlseg7:
  case Intrinsic::riscv_vlseg8:
  case Intrinsic::riscv_vlseg2ff:
  case Intrinsic::riscv_vlseg3ff:
  case Intrinsic::riscv_vlseg4ff:
  case Intrinsic::riscv_vlseg5ff:
  case Intrinsic::riscv_vlseg6ff:
  case Intrinsic::riscv_vlseg7ff:
  case Intrinsic::riscv_vlseg8ff:
    return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 2,
                               /*IsStore*/ false,
                               /*IsUnitStrided*/ false, /*UsePtrVal*/ true);
  case Intrinsic::riscv_vlseg2_mask:
  case Intrinsic::riscv_vlseg3_mask:
  case Intrinsic::riscv_vlseg4_mask:
  case Intrinsic::riscv_vlseg5_mask:
  case Intrinsic::riscv_vlseg6_mask:
  case Intrinsic::riscv_vlseg7_mask:
  case Intrinsic::riscv_vlseg8_mask:
  case Intrinsic::riscv_vlseg2ff_mask:
  case Intrinsic::riscv_vlseg3ff_mask:
  case Intrinsic::riscv_vlseg4ff_mask:
  case Intrinsic::riscv_vlseg5ff_mask:
  case Intrinsic::riscv_vlseg6ff_mask:
  case Intrinsic::riscv_vlseg7ff_mask:
  case Intrinsic::riscv_vlseg8ff_mask:
    return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 4,
                               /*IsStore*/ false,
                               /*IsUnitStrided*/ false, /*UsePtrVal*/ true);
  case Intrinsic::riscv_vlsseg2:
  case Intrinsic::riscv_vlsseg3:
  case Intrinsic::riscv_vlsseg4:
  case Intrinsic::riscv_vlsseg5:
  case Intrinsic::riscv_vlsseg6:
  case Intrinsic::riscv_vlsseg7:
  case Intrinsic::riscv_vlsseg8:
  case Intrinsic::riscv_vloxseg2:
  case Intrinsic::riscv_vloxseg3:
  case Intrinsic::riscv_vloxseg4:
  case Intrinsic::riscv_vloxseg5:
  case Intrinsic::riscv_vloxseg6:
  case Intrinsic::riscv_vloxseg7:
  case Intrinsic::riscv_vloxseg8:
  case Intrinsic::riscv_vluxseg2:
  case Intrinsic::riscv_vluxseg3:
  case Intrinsic::riscv_vluxseg4:
  case Intrinsic::riscv_vluxseg5:
  case Intrinsic::riscv_vluxseg6:
  case Intrinsic::riscv_vluxseg7:
  case Intrinsic::riscv_vluxseg8:
    return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 3,
                               /*IsStore*/ false,
                               /*IsUnitStrided*/ false);
  case Intrinsic::riscv_vlsseg2_mask:
  case Intrinsic::riscv_vlsseg3_mask:
  case Intrinsic::riscv_vlsseg4_mask:
  case Intrinsic::riscv_vlsseg5_mask:
  case Intrinsic::riscv_vlsseg6_mask:
  case Intrinsic::riscv_vlsseg7_mask:
  case Intrinsic::riscv_vlsseg8_mask:
  case Intrinsic::riscv_vloxseg2_mask:
  case Intrinsic::riscv_vloxseg3_mask:
  case Intrinsic::riscv_vloxseg4_mask:
  case Intrinsic::riscv_vloxseg5_mask:
  case Intrinsic::riscv_vloxseg6_mask:
  case Intrinsic::riscv_vloxseg7_mask:
  case Intrinsic::riscv_vloxseg8_mask:
  case Intrinsic::riscv_vluxseg2_mask:
  case Intrinsic::riscv_vluxseg3_mask:
  case Intrinsic::riscv_vluxseg4_mask:
  case Intrinsic::riscv_vluxseg5_mask:
  case Intrinsic::riscv_vluxseg6_mask:
  case Intrinsic::riscv_vluxseg7_mask:
  case Intrinsic::riscv_vluxseg8_mask:
    return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 5,
                               /*IsStore*/ false,
                               /*IsUnitStrided*/ false);
  case Intrinsic::riscv_vsseg2:
  case Intrinsic::riscv_vsseg3:
  case Intrinsic::riscv_vsseg4:
  case Intrinsic::riscv_vsseg5:
  case Intrinsic::riscv_vsseg6:
  case Intrinsic::riscv_vsseg7:
  case Intrinsic::riscv_vsseg8:
    return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 2,
                               /*IsStore*/ true,
                               /*IsUnitStrided*/ false);
  case Intrinsic::riscv_vsseg2_mask:
  case Intrinsic::riscv_vsseg3_mask:
  case Intrinsic::riscv_vsseg4_mask:
  case Intrinsic::riscv_vsseg5_mask:
  case Intrinsic::riscv_vsseg6_mask:
  case Intrinsic::riscv_vsseg7_mask:
  case Intrinsic::riscv_vsseg8_mask:
    return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 3,
                               /*IsStore*/ true,
                               /*IsUnitStrided*/ false);
  case Intrinsic::riscv_vssseg2:
  case Intrinsic::riscv_vssseg3:
  case Intrinsic::riscv_vssseg4:
  case Intrinsic::riscv_vssseg5:
  case Intrinsic::riscv_vssseg6:
  case Intrinsic::riscv_vssseg7:
  case Intrinsic::riscv_vssseg8:
  case Intrinsic::riscv_vsoxseg2:
  case Intrinsic::riscv_vsoxseg3:
  case Intrinsic::riscv_vsoxseg4:
  case Intrinsic::riscv_vsoxseg5:
  case Intrinsic::riscv_vsoxseg6:
  case Intrinsic::riscv_vsoxseg7:
  case Intrinsic::riscv_vsoxseg8:
  case Intrinsic::riscv_vsuxseg2:
  case Intrinsic::riscv_vsuxseg3:
  case Intrinsic::riscv_vsuxseg4:
  case Intrinsic::riscv_vsuxseg5:
  case Intrinsic::riscv_vsuxseg6:
  case Intrinsic::riscv_vsuxseg7:
  case Intrinsic::riscv_vsuxseg8:
    return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 3,
                               /*IsStore*/ true,
                               /*IsUnitStrided*/ false);
  case Intrinsic::riscv_vssseg2_mask:
  case Intrinsic::riscv_vssseg3_mask:
  case Intrinsic::riscv_vssseg4_mask:
  case Intrinsic::riscv_vssseg5_mask:
  case Intrinsic::riscv_vssseg6_mask:
  case Intrinsic::riscv_vssseg7_mask:
  case Intrinsic::riscv_vssseg8_mask:
  case Intrinsic::riscv_vsoxseg2_mask:
  case Intrinsic::riscv_vsoxseg3_mask:
  case Intrinsic::riscv_vsoxseg4_mask:
  case Intrinsic::riscv_vsoxseg5_mask:
  case Intrinsic::riscv_vsoxseg6_mask:
  case Intrinsic::riscv_vsoxseg7_mask:
  case Intrinsic::riscv_vsoxseg8_mask:
  case Intrinsic::riscv_vsuxseg2_mask:
  case Intrinsic::riscv_vsuxseg3_mask:
  case Intrinsic::riscv_vsuxseg4_mask:
  case Intrinsic::riscv_vsuxseg5_mask:
  case Intrinsic::riscv_vsuxseg6_mask:
  case Intrinsic::riscv_vsuxseg7_mask:
  case Intrinsic::riscv_vsuxseg8_mask:
    return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 4,
                               /*IsStore*/ true,
                               /*IsUnitStrided*/ false);
  }
}
 
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;
 
  // None of our addressing modes allows a scalable offset
  if (AM.ScalableOffset)
    return false;
 
  // RVV instructions only support register addressing.
  if (Subtarget.hasVInstructions() && isa<VectorType>(Ty))
    return AM.HasBaseReg && AM.Scale == 0 && !AM.BaseOffs;
 
  // 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.
// FIXME: Should we consider i64->i32 free on RV64 to match the EVT version of
// isTruncateFree?
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 {
  // We consider i64->i32 free on RV64 since we have good selection of W
  // instructions that make promoting operations back to i64 free in many cases.
  if (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::isTruncateFree(SDValue Val, EVT VT2) const {
  EVT SrcVT = Val.getValueType();
  // free truncate from vnsrl and vnsra
  if (Subtarget.hasVInstructions() &&
      (Val.getOpcode() == ISD::SRL || Val.getOpcode() == ISD::SRA) &&
      SrcVT.isVector() && VT2.isVector()) {
    unsigned SrcBits = SrcVT.getVectorElementType().getSizeInBits();
    unsigned DestBits = VT2.getVectorElementType().getSizeInBits();
    if (SrcBits == DestBits * 2) {
      return true;
    }
  }
  return TargetLowering::isTruncateFree(Val, VT2);
}
 
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::signExtendConstant(const ConstantInt *CI) const {
  return Subtarget.is64Bit() && CI->getType()->isIntegerTy(32);
}
 
bool RISCVTargetLowering::isCheapToSpeculateCttz(Type *Ty) const {
  return Subtarget.hasStdExtZbb() ||
         (Subtarget.hasVendorXCVbitmanip() && !Subtarget.is64Bit());
}
 
bool RISCVTargetLowering::isCheapToSpeculateCtlz(Type *Ty) const {
  return Subtarget.hasStdExtZbb() || Subtarget.hasVendorXTHeadBb() ||
         (Subtarget.hasVendorXCVbitmanip() && !Subtarget.is64Bit());
}
 
bool RISCVTargetLowering::isMaskAndCmp0FoldingBeneficial(
    const Instruction &AndI) const {
  // We expect to be able to match a bit extraction instruction if the Zbs
  // extension is supported and the mask is a power of two. However, we
  // conservatively return false if the mask would fit in an ANDI instruction,
  // on the basis that it's possible the sinking+duplication of the AND in
  // CodeGenPrepare triggered by this hook wouldn't decrease the instruction
  // count and would increase code size (e.g. ANDI+BNEZ => BEXTI+BNEZ).
  if (!Subtarget.hasStdExtZbs() && !Subtarget.hasVendorXTHeadBs())
    return false;
  ConstantInt *Mask = dyn_cast<ConstantInt>(AndI.getOperand(1));
  if (!Mask)
    return false;
  return !Mask->getValue().isSignedIntN(12) && Mask->getValue().isPowerOf2();
}
 
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.hasStdExtZbkb()) &&
         (!isa<ConstantSDNode>(Y) || cast<ConstantSDNode>(Y)->isOpaque());
}
 
bool RISCVTargetLowering::hasBitTest(SDValue X, SDValue Y) const {
  // Zbs provides BEXT[_I], which can be used with SEQZ/SNEZ as a bit test.
  if (Subtarget.hasStdExtZbs())
    return X.getValueType().isScalarInteger();
  auto *C = dyn_cast<ConstantSDNode>(Y);
  // XTheadBs provides th.tst (similar to bexti), if Y is a constant
  if (Subtarget.hasVendorXTHeadBs())
    return C != nullptr;
  // We can use ANDI+SEQZ/SNEZ as a bit test. Y contains the bit position.
  return C && C->getAPIntValue().ule(10);
}
 
bool RISCVTargetLowering::shouldFoldSelectWithIdentityConstant(unsigned Opcode,
                                                               EVT VT) const {
  // Only enable for rvv.
  if (!VT.isVector() || !Subtarget.hasVInstructions())
    return false;
 
  if (VT.isFixedLengthVector() && !isTypeLegal(VT))
    return false;
 
  return true;
}
 
bool RISCVTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
                                                            Type *Ty) const {
  assert(Ty->isIntegerTy());
 
  unsigned BitSize = Ty->getIntegerBitWidth();
  if (BitSize > Subtarget.getXLen())
    return false;
 
  // Fast path, assume 32-bit immediates are cheap.
  int64_t Val = Imm.getSExtValue();
  if (isInt<32>(Val))
    return true;
 
  // A constant pool entry may be more aligned thant he load we're trying to
  // replace. If we don't support unaligned scalar mem, prefer the constant
  // pool.
  // TODO: Can the caller pass down the alignment?
  if (!Subtarget.enableUnalignedScalarMem())
    return true;
 
  // Prefer to keep the load if it would require many instructions.
  // This uses the same threshold we use for constant pools but doesn't
  // check useConstantPoolForLargeInts.
  // TODO: Should we keep the load only when we're definitely going to emit a
  // constant pool?
 
  RISCVMatInt::InstSeq Seq = RISCVMatInt::generateInstSeq(Val, Subtarget);
  return Seq.size() <= Subtarget.getMaxBuildIntsCost();
}
 
bool RISCVTargetLowering::
    shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd(
        SDValue X, ConstantSDNode *XC, ConstantSDNode *CC, SDValue Y,
        unsigned OldShiftOpcode, unsigned NewShiftOpcode,
        SelectionDAG &DAG) const {
  // One interesting pattern that we'd want to form is 'bit extract':
  //   ((1 >> Y) & 1) ==/!= 0
  // But we also need to be careful not to try to reverse that fold.
 
  // Is this '((1 >> Y) & 1)'?
  if (XC && OldShiftOpcode == ISD::SRL && XC->isOne())
    return false; // Keep the 'bit extract' pattern.
 
  // Will this be '((1 >> Y) & 1)' after the transform?
  if (NewShiftOpcode == ISD::SRL && CC->isOne())
    return true; // Do form the 'bit extract' pattern.
 
  // If 'X' is a constant, and we transform, then we will immediately
  // try to undo the fold, thus causing endless combine loop.
  // So only do the transform if X is not a constant. This matches the default
  // implementation of this function.
  return !XC;
}
 
bool RISCVTargetLowering::canSplatOperand(unsigned Opcode, int Operand) const {
  switch (Opcode) {
  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:
  case Instruction::Select:
    return Operand == 1;
  default:
    return false;
  }
}
 
 
bool RISCVTargetLowering::canSplatOperand(Instruction *I, int Operand) const {
  if (!I->getType()->isVectorTy() || !Subtarget.hasVInstructions())
    return false;
 
  if (canSplatOperand(I->getOpcode(), Operand))
    return true;
 
  auto *II = dyn_cast<IntrinsicInst>(I);
  if (!II)
    return false;
 
  switch (II->getIntrinsicID()) {
  case Intrinsic::fma:
  case Intrinsic::vp_fma:
    return Operand == 0 || Operand == 1;
  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:
  case Intrinsic::ssub_sat:
  case Intrinsic::vp_ssub_sat:
  case Intrinsic::usub_sat:
  case Intrinsic::vp_usub_sat:
    return Operand == 1;
    // These intrinsics are 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_icmp:
  case Intrinsic::vp_fcmp:
  case Intrinsic::smin:
  case Intrinsic::vp_smin:
  case Intrinsic::umin:
  case Intrinsic::vp_umin:
  case Intrinsic::smax:
  case Intrinsic::vp_smax:
  case Intrinsic::umax:
  case Intrinsic::vp_umax:
  case Intrinsic::sadd_sat:
  case Intrinsic::vp_sadd_sat:
  case Intrinsic::uadd_sat:
  case Intrinsic::vp_uadd_sat:
    // These intrinsics have 'vr' versions.
  case Intrinsic::vp_sub:
  case Intrinsic::vp_fsub:
  case Intrinsic::vp_fdiv:
    return Operand == 0 || Operand == 1;
  default:
    return false;
  }
}
 
/// 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;
 
  // Don't sink splat operands if the target prefers it. Some targets requires
  // S2V transfer buffers and we can run out of them copying the same value
  // repeatedly.
  // FIXME: It could still be worth doing if it would improve vector register
  // pressure and prevent a vector spill.
  if (!Subtarget.sinkSplatOperands())
    return false;
 
  for (auto OpIdx : enumerate(I->operands())) {
    if (!canSplatOperand(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;
 
    // Don't sink i1 splats.
    if (cast<VectorType>(Op->getType())->getElementType()->isIntegerTy(1))
      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 (!canSplatOperand(Insn, U.getOperandNo()))
        return false;
    }
 
    Ops.push_back(&Op->getOperandUse(0));
    Ops.push_back(&OpIdx.value());
  }
  return true;
}
 
bool RISCVTargetLowering::shouldScalarizeBinop(SDValue VecOp) const {
  unsigned Opc = VecOp.getOpcode();
 
  // Assume target opcodes can't be scalarized.
  // TODO - do we have any exceptions?
  if (Opc >= ISD::BUILTIN_OP_END)
    return false;
 
  // If the vector op is not supported, try to convert to scalar.
  EVT VecVT = VecOp.getValueType();
  if (!isOperationLegalOrCustomOrPromote(Opc, VecVT))
    return true;
 
  // If the vector op is supported, but the scalar op is not, the transform may
  // not be worthwhile.
  // Permit a vector binary operation can be converted to scalar binary
  // operation which is custom lowered with illegal type.
  EVT ScalarVT = VecVT.getScalarType();
  return isOperationLegalOrCustomOrPromote(Opc, ScalarVT) ||
         isOperationCustom(Opc, ScalarVT);
}
 
bool RISCVTargetLowering::isOffsetFoldingLegal(
    const GlobalAddressSDNode *GA) const {
  // In order to maximise the opportunity for common subexpression elimination,
  // keep a separate ADD node for the global address offset instead of folding
  // it in the global address node. Later peephole optimisations may choose to
  // fold it back in when profitable.
  return false;
}
 
// Return one of the followings:
// (1) `{0-31 value, false}` if FLI is available for Imm's type and FP value.
// (2) `{0-31 value, true}` if Imm is negative and FLI is available for its
// positive counterpart, which will be materialized from the first returned
// element. The second returned element indicated that there should be a FNEG
// followed.
// (3) `{-1, _}` if there is no way FLI can be used to materialize Imm.
std::pair<int, bool> RISCVTargetLowering::getLegalZfaFPImm(const APFloat &Imm,
                                                           EVT VT) const {
  if (!Subtarget.hasStdExtZfa())
    return std::make_pair(-1, false);
 
  bool IsSupportedVT = false;
  if (VT == MVT::f16) {
    IsSupportedVT = Subtarget.hasStdExtZfh() || Subtarget.hasStdExtZvfh();
  } else if (VT == MVT::f32) {
    IsSupportedVT = true;
  } else if (VT == MVT::f64) {
    assert(Subtarget.hasStdExtD() && "Expect D extension");
    IsSupportedVT = true;
  }
 
  if (!IsSupportedVT)
    return std::make_pair(-1, false);
 
  int Index = RISCVLoadFPImm::getLoadFPImm(Imm);
  if (Index < 0 && Imm.isNegative())
    // Try the combination of its positive counterpart + FNEG.
    return std::make_pair(RISCVLoadFPImm::getLoadFPImm(-Imm), true);
  else
    return std::make_pair(Index, false);
}
 
bool RISCVTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
                                       bool ForCodeSize) const {
  bool IsLegalVT = false;
  if (VT == MVT::f16)
    IsLegalVT = Subtarget.hasStdExtZfhminOrZhinxmin();
  else if (VT == MVT::f32)
    IsLegalVT = Subtarget.hasStdExtFOrZfinx();
  else if (VT == MVT::f64)
    IsLegalVT = Subtarget.hasStdExtDOrZdinx();
  else if (VT == MVT::bf16)
    IsLegalVT = Subtarget.hasStdExtZfbfmin();
 
  if (!IsLegalVT)
    return false;
 
  if (getLegalZfaFPImm(Imm, VT).first >= 0)
    return true;
 
  // Cannot create a 64 bit floating-point immediate value for rv32.
  if (Subtarget.getXLen() < VT.getScalarSizeInBits()) {
    // td can handle +0.0 or -0.0 already.
    // -0.0 can be created by fmv + fneg.
    return Imm.isZero();
  }
 
  // Special case: fmv + fneg
  if (Imm.isNegZero())
    return true;
 
  // Building an integer and then converting requires a fmv at the end of
  // the integer sequence.
  const int Cost =
      1 + RISCVMatInt::getIntMatCost(Imm.bitcastToAPInt(), Subtarget.getXLen(),
                                     Subtarget);
  return Cost <= FPImmCost;
}
 
// TODO: This is very conservative.
bool RISCVTargetLowering::isExtractSubvectorCheap(EVT ResVT, EVT SrcVT,
                                                  unsigned Index) const {
  if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
    return false;
 
  // Only support extracting a fixed from a fixed vector for now.
  if (ResVT.isScalableVector() || SrcVT.isScalableVector())
    return false;
 
  EVT EltVT = ResVT.getVectorElementType();
  assert(EltVT == SrcVT.getVectorElementType() && "Should hold for node");
 
  // The smallest type we can slide is i8.
  // TODO: We can extract index 0 from a mask vector without a slide.
  if (EltVT == MVT::i1)
    return false;
 
  unsigned ResElts = ResVT.getVectorNumElements();
  unsigned SrcElts = SrcVT.getVectorNumElements();
 
  unsigned MinVLen = Subtarget.getRealMinVLen();
  unsigned MinVLMAX = MinVLen / EltVT.getSizeInBits();
 
  // If we're extracting only data from the first VLEN bits of the source
  // then we can always do this with an m1 vslidedown.vx.  Restricting the
  // Index ensures we can use a vslidedown.vi.
  // TODO: We can generalize this when the exact VLEN is known.
  if (Index + ResElts <= MinVLMAX && Index < 31)
    return true;
 
  // Convervatively only handle extracting half of a vector.
  // TODO: For sizes which aren't multiples of VLEN sizes, this may not be
  // a cheap extract.  However, this case is important in practice for
  // shuffled extracts of longer vectors.  How resolve?
  if ((ResElts * 2) != SrcElts)
    return false;
 
  // Slide can support arbitrary index, but we only treat vslidedown.vi as
  // cheap.
  if (Index >= 32)
    return false;
 
  // TODO: We can do arbitrary slidedowns, but for now only support extracting
  // the upper half of a vector until we have more test coverage.
  return Index == 0 || Index == ResElts;
}
 
MVT RISCVTargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
                                                      CallingConv::ID CC,
                                                      EVT VT) const {
  // Use f32 to pass f16 if it is legal and Zfh/Zfhmin is not enabled.
  // We might still end up using a GPR but that will be decided based on ABI.
  if (VT == MVT::f16 && Subtarget.hasStdExtFOrZfinx() &&
      !Subtarget.hasStdExtZfhminOrZhinxmin())
    return MVT::f32;
 
  MVT PartVT = TargetLowering::getRegisterTypeForCallingConv(Context, CC, VT);
 
  return PartVT;
}
 
unsigned RISCVTargetLowering::getNumRegistersForCallingConv(LLVMContext &Context,
                                                           CallingConv::ID CC,
                                                           EVT VT) const {
  // Use f32 to pass f16 if it is legal and Zfh/Zfhmin is not enabled.
  // We might still end up using a GPR but that will be decided based on ABI.
  if (VT == MVT::f16 && Subtarget.hasStdExtFOrZfinx() &&
      !Subtarget.hasStdExtZfhminOrZhinxmin())
    return 1;
 
  return TargetLowering::getNumRegistersForCallingConv(Context, CC, VT);
}
 
unsigned RISCVTargetLowering::getVectorTypeBreakdownForCallingConv(
    LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT,
    unsigned &NumIntermediates, MVT &RegisterVT) const {
  unsigned NumRegs = TargetLowering::getVectorTypeBreakdownForCallingConv(
      Context, CC, VT, IntermediateVT, NumIntermediates, RegisterVT);
 
  return NumRegs;
}
 
// 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) {
  // If this is a single bit test that can't be handled by ANDI, shift the
  // bit to be tested to the MSB and perform a signed compare with 0.
  if (isIntEqualitySetCC(CC) && isNullConstant(RHS) &&
      LHS.getOpcode() == ISD::AND && LHS.hasOneUse() &&
      isa<ConstantSDNode>(LHS.getOperand(1))) {
    uint64_t Mask = LHS.getConstantOperandVal(1);
    if ((isPowerOf2_64(Mask) || isMask_64(Mask)) && !isInt<12>(Mask)) {
      unsigned ShAmt = 0;
      if (isPowerOf2_64(Mask)) {
        CC = CC == ISD::SETEQ ? ISD::SETGE : ISD::SETLT;
        ShAmt = LHS.getValueSizeInBits() - 1 - Log2_64(Mask);
      } else {
        ShAmt = LHS.getValueSizeInBits() - llvm::bit_width(Mask);
      }
 
      LHS = LHS.getOperand(0);
      if (ShAmt != 0)
        LHS = DAG.getNode(ISD::SHL, DL, LHS.getValueType(), LHS,
                          DAG.getConstant(ShAmt, DL, LHS.getValueType()));
      return;
    }
  }
 
  if (auto *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
    int64_t C = RHSC->getSExtValue();
    switch (CC) {
    default: break;
    case ISD::SETGT:
      // Convert X > -1 to X >= 0.
      if (C == -1) {
        RHS = DAG.getConstant(0, DL, RHS.getValueType());
        CC = ISD::SETGE;
        return;
      }
      break;
    case ISD::SETLT:
      // Convert X < 1 to 0 >= X.
      if (C == 1) {
        RHS = LHS;
        LHS = DAG.getConstant(0, DL, RHS.getValueType());
        CC = ISD::SETGE;
        return;
      }
      break;
    }
  }
 
  switch (CC) {
  default:
    break;
  case ISD::SETGT:
  case ISD::SETLE:
  case ISD::SETUGT:
  case ISD::SETULE:
    CC = ISD::getSetCCSwappedOperands(CC);
    std::swap(LHS, RHS);
    break;
  }
}
 
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(EVT ScalarTy) const {
  if (!ScalarTy.isSimple())
    return false;
  switch (ScalarTy.getSimpleVT().SimpleTy) {
  case MVT::iPTR:
    return Subtarget.is64Bit() ? Subtarget.hasVInstructionsI64() : true;
  case MVT::i8:
  case MVT::i16:
  case MVT::i32:
    return true;
  case MVT::i64:
    return Subtarget.hasVInstructionsI64();
  case MVT::f16:
    return Subtarget.hasVInstructionsF16();
  case MVT::f32:
    return Subtarget.hasVInstructionsF32();
  case MVT::f64:
    return Subtarget.hasVInstructionsF64();
  default:
    return false;
  }
}
 
 
unsigned RISCVTargetLowering::combineRepeatedFPDivisors() const {
  return NumRepeatedDivisors;
}
 
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.getRealMinVLen();
 
  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.hasVInstructionsF16Minimal())
      return false;
    break;
  case MVT::bf16:
    if (!Subtarget.hasVInstructionsBF16Minimal())
      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.getELen())
    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.getRealMinVLen();
  unsigned MaxELen = Subtarget.getELen();
 
  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::bf16:
  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.getVectorIdxConstant(0, DL);
  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);
}
 
/// Return the type of the mask type suitable for masking the provided
/// vector type.  This is simply an i1 element type vector of the same
/// (possibly scalable) length.
static MVT getMaskTypeFor(MVT VecVT) {
  assert(VecVT.isVector());
  ElementCount EC = VecVT.getVectorElementCount();
  return MVT::getVectorVT(MVT::i1, EC);
}
 
/// Creates an all ones mask suitable for masking a vector of type VecTy with
/// vector length VL.  .
static SDValue getAllOnesMask(MVT VecVT, SDValue VL, const SDLoc &DL,
                              SelectionDAG &DAG) {
  MVT MaskVT = getMaskTypeFor(VecVT);
  return DAG.getNode(RISCVISD::VMSET_VL, DL, MaskVT, VL);
}
 
static std::pair<SDValue, SDValue>
getDefaultScalableVLOps(MVT VecVT, const SDLoc &DL, SelectionDAG &DAG,
                        const RISCVSubtarget &Subtarget) {
  assert(VecVT.isScalableVector() && "Expecting a scalable vector");
  SDValue VL = DAG.getRegister(RISCV::X0, Subtarget.getXLenVT());
  SDValue Mask = getAllOnesMask(VecVT, VL, DL, DAG);
  return {Mask, VL};
}
 
static std::pair<SDValue, SDValue>
getDefaultVLOps(uint64_t NumElts, MVT ContainerVT, const SDLoc &DL,
                SelectionDAG &DAG, const RISCVSubtarget &Subtarget) {
  assert(ContainerVT.isScalableVector() && "Expecting scalable container type");
  SDValue VL = DAG.getConstant(NumElts, DL, Subtarget.getXLenVT());
  SDValue Mask = getAllOnesMask(ContainerVT, VL, DL, DAG);
  return {Mask, VL};
}
 
// 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 the fixed-length vector is contained in. Otherwise if
// VecVT is scalable, then ContainerVT should be the same as VecVT.
static std::pair<SDValue, SDValue>
getDefaultVLOps(MVT VecVT, MVT ContainerVT, const SDLoc &DL, SelectionDAG &DAG,
                const RISCVSubtarget &Subtarget) {
  if (VecVT.isFixedLengthVector())
    return getDefaultVLOps(VecVT.getVectorNumElements(), ContainerVT, DL, DAG,
                           Subtarget);
  assert(ContainerVT.isScalableVector() && "Expecting scalable container type");
  return getDefaultScalableVLOps(ContainerVT, DL, DAG, Subtarget);
}
 
SDValue RISCVTargetLowering::computeVLMax(MVT VecVT, const SDLoc &DL,
                                          SelectionDAG &DAG) const {
  assert(VecVT.isScalableVector() && "Expected scalable vector");
  return DAG.getElementCount(DL, Subtarget.getXLenVT(),
                             VecVT.getVectorElementCount());
}
 
std::pair<unsigned, unsigned>
RISCVTargetLowering::computeVLMAXBounds(MVT VecVT,
                                        const RISCVSubtarget &Subtarget) {
  assert(VecVT.isScalableVector() && "Expected scalable vector");
 
  unsigned EltSize = VecVT.getScalarSizeInBits();
  unsigned MinSize = VecVT.getSizeInBits().getKnownMinValue();
 
  unsigned VectorBitsMax = Subtarget.getRealMaxVLen();
  unsigned MaxVLMAX =
      RISCVTargetLowering::computeVLMAX(VectorBitsMax, EltSize, MinSize);
 
  unsigned VectorBitsMin = Subtarget.getRealMinVLen();
  unsigned MinVLMAX =
      RISCVTargetLowering::computeVLMAX(VectorBitsMin, EltSize, MinSize);
 
  return std::make_pair(MinVLMAX, MaxVLMAX);
}
 
// 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;
}
 
InstructionCost RISCVTargetLowering::getLMULCost(MVT VT) const {
  // TODO: Here assume reciprocal throughput is 1 for LMUL_1, it is
  // implementation-defined.
  if (!VT.isVector())
    return InstructionCost::getInvalid();
  unsigned DLenFactor = Subtarget.getDLenFactor();
  unsigned Cost;
  if (VT.isScalableVector()) {
    unsigned LMul;
    bool Fractional;
    std::tie(LMul, Fractional) =
        RISCVVType::decodeVLMUL(RISCVTargetLowering::getLMUL(VT));
    if (Fractional)
      Cost = LMul <= DLenFactor ? (DLenFactor / LMul) : 1;
    else
      Cost = (LMul * DLenFactor);
  } else {
    Cost = divideCeil(VT.getSizeInBits(), Subtarget.getRealMinVLen() / DLenFactor);
  }
  return Cost;
}
 
 
/// Return the cost of a vrgather.vv instruction for the type VT.  vrgather.vv
/// is generally quadratic in the number of vreg implied by LMUL.  Note that
/// operand (index and possibly mask) are handled separately.
InstructionCost RISCVTargetLowering::getVRGatherVVCost(MVT VT) const {
  return getLMULCost(VT) * getLMULCost(VT);
}
 
/// Return the cost of a vrgather.vi (or vx) instruction for the type VT.
/// vrgather.vi/vx may be linear in the number of vregs implied by LMUL,
/// or may track the vrgather.vv cost. It is implementation-dependent.
InstructionCost RISCVTargetLowering::getVRGatherVICost(MVT VT) const {
  return getLMULCost(VT);
}
 
/// Return the cost of a vslidedown.vx or vslideup.vx instruction
/// for the type VT.  (This does not cover the vslide1up or vslide1down
/// variants.)  Slides may be linear in the number of vregs implied by LMUL,
/// or may track the vrgather.vv cost. It is implementation-dependent.
InstructionCost RISCVTargetLowering::getVSlideVXCost(MVT VT) const {
  return getLMULCost(VT);
}
 
/// Return the cost of a vslidedown.vi or vslideup.vi instruction
/// for the type VT.  (This does not cover the vslide1up or vslide1down
/// variants.)  Slides may be linear in the number of vregs implied by LMUL,
/// or may track the vrgather.vv cost. It is implementation-dependent.
InstructionCost RISCVTargetLowering::getVSlideVICost(MVT VT) const {
  return getLMULCost(VT);
}
 
static SDValue lowerFP_TO_INT_SAT(SDValue Op, SelectionDAG &DAG,
                                  const RISCVSubtarget &Subtarget) {
  // RISC-V 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);
 
  MVT DstVT = Op.getSimpleValueType();
  EVT SatVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
 
  bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT_SAT;
 
  if (!DstVT.isVector()) {
    // For bf16 or for f16 in absense of Zfh, promote to f32, then saturate
    // the result.
    if ((Src.getValueType() == MVT::f16 && !Subtarget.hasStdExtZfhOrZhinx()) ||
        Src.getValueType() == MVT::bf16) {
      Src = DAG.getNode(ISD::FP_EXTEND, SDLoc(Op), MVT::f32, Src);
    }
 
    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()));
 
    if (Opc == RISCVISD::FCVT_WU_RV64)
      FpToInt = DAG.getZeroExtendInReg(FpToInt, DL, MVT::i32);
 
    SDValue ZeroInt = DAG.getConstant(0, DL, DstVT);
    return DAG.getSelectCC(DL, Src, Src, ZeroInt, FpToInt,
                           ISD::CondCode::SETUO);
  }
 
  // Vectors.
 
  MVT DstEltVT = DstVT.getVectorElementType();
  MVT SrcVT = Src.getSimpleValueType();
  MVT SrcEltVT = SrcVT.getVectorElementType();
  unsigned SrcEltSize = SrcEltVT.getSizeInBits();
  unsigned DstEltSize = DstEltVT.getSizeInBits();
 
  // Only handle saturating to the destination type.
  if (SatVT != DstEltVT)
    return SDValue();
 
  MVT DstContainerVT = DstVT;
  MVT SrcContainerVT = SrcVT;
  if (DstVT.isFixedLengthVector()) {
    DstContainerVT = getContainerForFixedLengthVector(DAG, DstVT, Subtarget);
    SrcContainerVT = getContainerForFixedLengthVector(DAG, SrcVT, Subtarget);
    assert(DstContainerVT.getVectorElementCount() ==
               SrcContainerVT.getVectorElementCount() &&
           "Expected same element count");
    Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget);
  }
 
  SDLoc DL(Op);
 
  auto [Mask, VL] = getDefaultVLOps(DstVT, DstContainerVT, DL, DAG, Subtarget);
 
  SDValue IsNan = DAG.getNode(RISCVISD::SETCC_VL, DL, Mask.getValueType(),
                              {Src, Src, DAG.getCondCode(ISD::SETNE),
                               DAG.getUNDEF(Mask.getValueType()), Mask, VL});
 
  // Need to widen by more than 1 step, promote the FP type, then do a widening
  // convert.
  if (DstEltSize > (2 * SrcEltSize)) {
    assert(SrcContainerVT.getVectorElementType() == MVT::f16 && "Unexpected VT!");
    MVT InterVT = SrcContainerVT.changeVectorElementType(MVT::f32);
    Src = DAG.getNode(RISCVISD::FP_EXTEND_VL, DL, InterVT, Src, Mask, VL);
  }
 
  MVT CvtContainerVT = DstContainerVT;
  MVT CvtEltVT = DstEltVT;
  if (SrcEltSize > (2 * DstEltSize)) {
    CvtEltVT = MVT::getIntegerVT(SrcEltVT.getSizeInBits() / 2);
    CvtContainerVT = CvtContainerVT.changeVectorElementType(CvtEltVT);
  }
 
  unsigned RVVOpc =
      IsSigned ? RISCVISD::VFCVT_RTZ_X_F_VL : RISCVISD::VFCVT_RTZ_XU_F_VL;
  SDValue Res = DAG.getNode(RVVOpc, DL, CvtContainerVT, Src, Mask, VL);
 
  while (CvtContainerVT != DstContainerVT) {
    CvtEltVT = MVT::getIntegerVT(CvtEltVT.getSizeInBits() / 2);
    CvtContainerVT = CvtContainerVT.changeVectorElementType(CvtEltVT);
    // Rounding mode here is arbitrary since we aren't shifting out any bits.
    unsigned ClipOpc = IsSigned ? RISCVISD::TRUNCATE_VECTOR_VL_SSAT
                                : RISCVISD::TRUNCATE_VECTOR_VL_USAT;
    Res = DAG.getNode(ClipOpc, DL, CvtContainerVT, Res, Mask, VL);
  }
 
  SDValue SplatZero = DAG.getNode(
      RISCVISD::VMV_V_X_VL, DL, DstContainerVT, DAG.getUNDEF(DstContainerVT),
      DAG.getConstant(0, DL, Subtarget.getXLenVT()), VL);
  Res = DAG.getNode(RISCVISD::VMERGE_VL, DL, DstContainerVT, IsNan, SplatZero,
                    Res, DAG.getUNDEF(DstContainerVT), VL);
 
  if (DstVT.isFixedLengthVector())
    Res = convertFromScalableVector(DstVT, Res, DAG, Subtarget);
 
  return Res;
}
 
static RISCVFPRndMode::RoundingMode matchRoundingOp(unsigned Opc) {
  switch (Opc) {
  case ISD::FROUNDEVEN:
  case ISD::STRICT_FROUNDEVEN:
  case ISD::VP_FROUNDEVEN:
    return RISCVFPRndMode::RNE;
  case ISD::FTRUNC:
  case ISD::STRICT_FTRUNC:
  case ISD::VP_FROUNDTOZERO:
    return RISCVFPRndMode::RTZ;
  case ISD::FFLOOR:
  case ISD::STRICT_FFLOOR:
  case ISD::VP_FFLOOR:
    return RISCVFPRndMode::RDN;
  case ISD::FCEIL:
  case ISD::STRICT_FCEIL:
  case ISD::VP_FCEIL:
    return RISCVFPRndMode::RUP;
  case ISD::FROUND:
  case ISD::STRICT_FROUND:
  case ISD::VP_FROUND:
    return RISCVFPRndMode::RMM;
  case ISD::FRINT:
    return RISCVFPRndMode::DYN;
  }
 
  return RISCVFPRndMode::Invalid;
}
 
// Expand vector FTRUNC, FCEIL, FFLOOR, FROUND, VP_FCEIL, VP_FFLOOR, VP_FROUND
// VP_FROUNDEVEN, VP_FROUNDTOZERO, VP_FRINT and VP_FNEARBYINT by converting to
// the integer domain and back. Taking care to avoid converting values that are
// nan or already correct.
static SDValue
lowerVectorFTRUNC_FCEIL_FFLOOR_FROUND(SDValue Op, SelectionDAG &DAG,
                                      const RISCVSubtarget &Subtarget) {
  MVT VT = Op.getSimpleValueType();
  assert(VT.isVector() && "Unexpected type");
 
  SDLoc DL(Op);
 
  SDValue Src = Op.getOperand(0);
 
  MVT ContainerVT = VT;
  if (VT.isFixedLengthVector()) {
    ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
    Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
  }
 
  SDValue Mask, VL;
  if (Op->isVPOpcode()) {
    Mask = Op.getOperand(1);
    if (VT.isFixedLengthVector())
      Mask = convertToScalableVector(getMaskTypeFor(ContainerVT), Mask, DAG,
                                     Subtarget);
    VL = Op.getOperand(2);
  } else {
    std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
  }
 
  // Freeze the source since we are increasing the number of uses.
  Src = DAG.getFreeze(Src);
 
  // We do the conversion on the absolute value and fix the sign at the end.
  SDValue Abs = DAG.getNode(RISCVISD::FABS_VL, DL, ContainerVT, Src, Mask, VL);
 
  // 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(ContainerVT);
  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, ContainerVT.getVectorElementType());
  SDValue MaxValSplat = DAG.getNode(RISCVISD::VFMV_V_F_VL, DL, ContainerVT,
                                    DAG.getUNDEF(ContainerVT), MaxValNode, VL);
 
  // If abs(Src) was larger than MaxVal or nan, keep it.
  MVT SetccVT = MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount());
  Mask =
      DAG.getNode(RISCVISD::SETCC_VL, DL, SetccVT,
                  {Abs, MaxValSplat, DAG.getCondCode(ISD::SETOLT),
                   Mask, Mask, VL});
 
  // Truncate to integer and convert back to FP.
  MVT IntVT = ContainerVT.changeVectorElementTypeToInteger();
  MVT XLenVT = Subtarget.getXLenVT();
  SDValue Truncated;
 
  switch (Op.getOpcode()) {
  default:
    llvm_unreachable("Unexpected opcode");
  case ISD::FCEIL:
  case ISD::VP_FCEIL:
  case ISD::FFLOOR:
  case ISD::VP_FFLOOR:
  case ISD::FROUND:
  case ISD::FROUNDEVEN:
  case ISD::VP_FROUND:
  case ISD::VP_FROUNDEVEN:
  case ISD::VP_FROUNDTOZERO: {
    RISCVFPRndMode::RoundingMode FRM = matchRoundingOp(Op.getOpcode());
    assert(FRM != RISCVFPRndMode::Invalid);
    Truncated = DAG.getNode(RISCVISD::VFCVT_RM_X_F_VL, DL, IntVT, Src, Mask,
                            DAG.getTargetConstant(FRM, DL, XLenVT), VL);
    break;
  }
  case ISD::FTRUNC:
    Truncated = DAG.getNode(RISCVISD::VFCVT_RTZ_X_F_VL, DL, IntVT, Src,
                            Mask, VL);
    break;
  case ISD::FRINT:
  case ISD::VP_FRINT:
    Truncated = DAG.getNode(RISCVISD::VFCVT_X_F_VL, DL, IntVT, Src, Mask, VL);
    break;
  case ISD::FNEARBYINT:
  case ISD::VP_FNEARBYINT:
    Truncated = DAG.getNode(RISCVISD::VFROUND_NOEXCEPT_VL, DL, ContainerVT, Src,
                            Mask, VL);
    break;
  }
 
  // VFROUND_NOEXCEPT_VL includes SINT_TO_FP_VL.
  if (Truncated.getOpcode() != RISCVISD::VFROUND_NOEXCEPT_VL)
    Truncated = DAG.getNode(RISCVISD::SINT_TO_FP_VL, DL, ContainerVT, Truncated,
                            Mask, VL);
 
  // Restore the original sign so that -0.0 is preserved.
  Truncated = DAG.getNode(RISCVISD::FCOPYSIGN_VL, DL, ContainerVT, Truncated,
                          Src, Src, Mask, VL);
 
  if (!VT.isFixedLengthVector())
    return Truncated;
 
  return convertFromScalableVector(VT, Truncated, DAG, Subtarget);
}
 
// Expand vector STRICT_FTRUNC, STRICT_FCEIL, STRICT_FFLOOR, STRICT_FROUND
// STRICT_FROUNDEVEN and STRICT_FNEARBYINT by converting sNan of the source to
// qNan and coverting the new source to integer and back to FP.
static SDValue
lowerVectorStrictFTRUNC_FCEIL_FFLOOR_FROUND(SDValue Op, SelectionDAG &DAG,
                                            const RISCVSubtarget &Subtarget) {
  SDLoc DL(Op);
  MVT VT = Op.getSimpleValueType();
  SDValue Chain = Op.getOperand(0);
  SDValue Src = Op.getOperand(1);
 
  MVT ContainerVT = VT;
  if (VT.isFixedLengthVector()) {
    ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
    Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
  }
 
  auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
 
  // Freeze the source since we are increasing the number of uses.
  Src = DAG.getFreeze(Src);
 
  // Covert sNan to qNan by executing x + x for all unordered elemenet x in Src.
  MVT MaskVT = Mask.getSimpleValueType();
  SDValue Unorder = DAG.getNode(RISCVISD::STRICT_FSETCC_VL, DL,
                                DAG.getVTList(MaskVT, MVT::Other),
                                {Chain, Src, Src, DAG.getCondCode(ISD::SETUNE),
                                 DAG.getUNDEF(MaskVT), Mask, VL});
  Chain = Unorder.getValue(1);
  Src = DAG.getNode(RISCVISD::STRICT_FADD_VL, DL,
                    DAG.getVTList(ContainerVT, MVT::Other),
                    {Chain, Src, Src, Src, Unorder, VL});
  Chain = Src.getValue(1);
 
  // We do the conversion on the absolute value and fix the sign at the end.
  SDValue Abs = DAG.getNode(RISCVISD::FABS_VL, DL, ContainerVT, Src, Mask, VL);
 
  // 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(ContainerVT);
  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, ContainerVT.getVectorElementType());
  SDValue MaxValSplat = DAG.getNode(RISCVISD::VFMV_V_F_VL, DL, ContainerVT,
                                    DAG.getUNDEF(ContainerVT), MaxValNode, VL);
 
  // If abs(Src) was larger than MaxVal or nan, keep it.
  Mask = DAG.getNode(
      RISCVISD::SETCC_VL, DL, MaskVT,
      {Abs, MaxValSplat, DAG.getCondCode(ISD::SETOLT), Mask, Mask, VL});
 
  // Truncate to integer and convert back to FP.
  MVT IntVT = ContainerVT.changeVectorElementTypeToInteger();
  MVT XLenVT = Subtarget.getXLenVT();
  SDValue Truncated;
 
  switch (Op.getOpcode()) {
  default:
    llvm_unreachable("Unexpected opcode");
  case ISD::STRICT_FCEIL:
  case ISD::STRICT_FFLOOR:
  case ISD::STRICT_FROUND:
  case ISD::STRICT_FROUNDEVEN: {
    RISCVFPRndMode::RoundingMode FRM = matchRoundingOp(Op.getOpcode());
    assert(FRM != RISCVFPRndMode::Invalid);
    Truncated = DAG.getNode(
        RISCVISD::STRICT_VFCVT_RM_X_F_VL, DL, DAG.getVTList(IntVT, MVT::Other),
        {Chain, Src, Mask, DAG.getTargetConstant(FRM, DL, XLenVT), VL});
    break;
  }
  case ISD::STRICT_FTRUNC:
    Truncated =
        DAG.getNode(RISCVISD::STRICT_VFCVT_RTZ_X_F_VL, DL,
                    DAG.getVTList(IntVT, MVT::Other), Chain, Src, Mask, VL);
    break;
  case ISD::STRICT_FNEARBYINT:
    Truncated = DAG.getNode(RISCVISD::STRICT_VFROUND_NOEXCEPT_VL, DL,
                            DAG.getVTList(ContainerVT, MVT::Other), Chain, Src,
                            Mask, VL);
    break;
  }
  Chain = Truncated.getValue(1);
 
  // VFROUND_NOEXCEPT_VL includes SINT_TO_FP_VL.
  if (Op.getOpcode() != ISD::STRICT_FNEARBYINT) {
    Truncated = DAG.getNode(RISCVISD::STRICT_SINT_TO_FP_VL, DL,
                            DAG.getVTList(ContainerVT, MVT::Other), Chain,
                            Truncated, Mask, VL);
    Chain = Truncated.getValue(1);
  }
 
  // Restore the original sign so that -0.0 is preserved.
  Truncated = DAG.getNode(RISCVISD::FCOPYSIGN_VL, DL, ContainerVT, Truncated,
                          Src, Src, Mask, VL);
 
  if (VT.isFixedLengthVector())
    Truncated = convertFromScalableVector(VT, Truncated, DAG, Subtarget);
  return DAG.getMergeValues({Truncated, Chain}, DL);
}
 
static SDValue
lowerFTRUNC_FCEIL_FFLOOR_FROUND(SDValue Op, SelectionDAG &DAG,
                                const RISCVSubtarget &Subtarget) {
  MVT VT = Op.getSimpleValueType();
  if (VT.isVector())
    return lowerVectorFTRUNC_FCEIL_FFLOOR_FROUND(Op, DAG, Subtarget);
 
  if (DAG.shouldOptForSize())
    return SDValue();
 
  SDLoc DL(Op);
  SDValue Src = Op.getOperand(0);
 
  // Create an integer the size of the mantissa with the MSB set. This and all
  // 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);
 
  RISCVFPRndMode::RoundingMode FRM = matchRoundingOp(Op.getOpcode());
  return DAG.getNode(RISCVISD::FROUND, DL, VT, Src, MaxValNode,
                     DAG.getTargetConstant(FRM, DL, Subtarget.getXLenVT()));
}
 
// Expand vector LRINT and LLRINT by converting to the integer domain.
static SDValue lowerVectorXRINT(SDValue Op, SelectionDAG &DAG,
                                const RISCVSubtarget &Subtarget) {
  MVT VT = Op.getSimpleValueType();
  assert(VT.isVector() && "Unexpected type");
 
  SDLoc DL(Op);
  SDValue Src = Op.getOperand(0);
  MVT ContainerVT = VT;
 
  if (VT.isFixedLengthVector()) {
    ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
    Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget);
  }
 
  auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
  SDValue Truncated =
      DAG.getNode(RISCVISD::VFCVT_X_F_VL, DL, ContainerVT, Src, Mask, VL);
 
  if (!VT.isFixedLengthVector())
    return Truncated;
 
  return convertFromScalableVector(VT, Truncated, DAG, Subtarget);
}
 
static SDValue
getVSlidedown(SelectionDAG &DAG, const RISCVSubtarget &Subtarget,
              const SDLoc &DL, EVT VT, SDValue Passthru, SDValue Op,
              SDValue Offset, SDValue Mask, SDValue VL,
              unsigned Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED) {
  if (Passthru.isUndef())
    Policy = RISCVII::TAIL_AGNOSTIC | RISCVII::MASK_AGNOSTIC;
  SDValue PolicyOp = DAG.getTargetConstant(Policy, DL, Subtarget.getXLenVT());
  SDValue Ops[] = {Passthru, Op, Offset, Mask, VL, PolicyOp};
  return DAG.getNode(RISCVISD::VSLIDEDOWN_VL, DL, VT, Ops);
}
 
static SDValue
getVSlideup(SelectionDAG &DAG, const RISCVSubtarget &Subtarget, const SDLoc &DL,
            EVT VT, SDValue Passthru, SDValue Op, SDValue Offset, SDValue Mask,
            SDValue VL,
            unsigned Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED) {
  if (Passthru.isUndef())
    Policy = RISCVII::TAIL_AGNOSTIC | RISCVII::MASK_AGNOSTIC;
  SDValue PolicyOp = DAG.getTargetConstant(Policy, DL, Subtarget.getXLenVT());
  SDValue Ops[] = {Passthru, Op, Offset, Mask, VL, PolicyOp};
  return DAG.getNode(RISCVISD::VSLIDEUP_VL, DL, VT, Ops);
}
 
static MVT getLMUL1VT(MVT VT) {
  assert(VT.getVectorElementType().getSizeInBits() <= 64 &&
         "Unexpected vector MVT");
  return MVT::getScalableVectorVT(
      VT.getVectorElementType(),
      RISCV::RVVBitsPerBlock / VT.getVectorElementType().getSizeInBits());
}
 
struct VIDSequence {
  int64_t StepNumerator;
  unsigned StepDenominator;
  int64_t Addend;
};
 
static std::optional<APInt> getExactInteger(const APFloat &APF,
                                            uint32_t BitWidth) {
  // We will use a SINT_TO_FP to materialize this constant so we should use a
  // signed APSInt here.
  APSInt ValInt(BitWidth, /*IsUnsigned*/ false);
  // We use an arbitrary rounding mode here. If a floating-point is an exact
  // integer (e.g., 1.0), the rounding mode does not affect the output value. If
  // the rounding mode changes the output value, then it is not an exact
  // integer.
  RoundingMode ArbitraryRM = RoundingMode::TowardZero;
  bool IsExact;
  // If it is out of signed integer range, it will return an invalid operation.
  // If it is not an exact integer, IsExact is false.
  if ((APF.convertToInteger(ValInt, ArbitraryRM, &IsExact) ==
       APFloatBase::opInvalidOp) ||
      !IsExact)
    return std::nullopt;
  return ValInt.extractBits(BitWidth, 0);
}
 
// 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.
//
// EltSizeInBits is the size of the type that the sequence will be calculated
// in, i.e. SEW for build_vectors or XLEN for address calculations.
static std::optional<VIDSequence> isSimpleVIDSequence(SDValue Op,
                                                      unsigned EltSizeInBits) {
  assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unexpected BUILD_VECTOR");
  if (!cast<BuildVectorSDNode>(Op)->isConstant())
    return std::nullopt;
  bool IsInteger = Op.getValueType().isInteger();
 
  std::optional<unsigned> SeqStepDenom;
  std::optional<APInt> SeqStepNum;
  std::optional<APInt> SeqAddend;
  std::optional<std::pair<APInt, unsigned>> PrevElt;
  assert(EltSizeInBits >= Op.getValueType().getScalarSizeInBits());
 
  // First extract the ops into a list of constant integer values. This may not
  // be possible for floats if they're not all representable as integers.
  SmallVector<std::optional<APInt>> Elts(Op.getNumOperands());
  const unsigned OpSize = Op.getScalarValueSizeInBits();
  for (auto [Idx, Elt] : enumerate(Op->op_values())) {
    if (Elt.isUndef()) {
      Elts[Idx] = std::nullopt;
      continue;
    }
    if (IsInteger) {
      Elts[Idx] = Elt->getAsAPIntVal().trunc(OpSize).zext(EltSizeInBits);
    } else {
      auto ExactInteger =
          getExactInteger(cast<ConstantFPSDNode>(Elt)->getValueAPF(), OpSize);
      if (!ExactInteger)
        return std::nullopt;
      Elts[Idx] = *ExactInteger;
    }
  }
 
  for (auto [Idx, Elt] : enumerate(Elts)) {
    // Assume undef elements match the sequence; we just have to be careful
    // when interpolating across them.
    if (!Elt)
      continue;
 
    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;
      APInt ValDiff = *Elt - PrevElt->first;
 
      // 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)
        continue;
 
      int64_t Remainder = ValDiff.srem(IdxDiff);
      // Normalize the step if it's greater than 1.
      if (Remainder != ValDiff.getSExtValue()) {
        // The difference must cleanly divide the element span.
        if (Remainder != 0)
          return std::nullopt;
        ValDiff = ValDiff.sdiv(IdxDiff);
        IdxDiff = 1;
      }
 
      if (!SeqStepNum)
        SeqStepNum = ValDiff;
      else if (ValDiff != SeqStepNum)
        return std::nullopt;
 
      if (!SeqStepDenom)
        SeqStepDenom = IdxDiff;
      else if (IdxDiff != *SeqStepDenom)
        return std::nullopt;
    }
 
    // Record this non-undef element for later.
    if (!PrevElt || PrevElt->first != *Elt)
      PrevElt = std::make_pair(*Elt, Idx);
  }
 
  // We need to have logged a step for this to count as a legal index sequence.
  if (!SeqStepNum || !SeqStepDenom)
    return std::nullopt;
 
  // Loop back through the sequence and validate elements we might have skipped
  // while waiting for a valid step. While doing this, log any sequence addend.
  for (auto [Idx, Elt] : enumerate(Elts)) {
    if (!Elt)
      continue;
    APInt ExpectedVal =
        (APInt(EltSizeInBits, Idx) * *SeqStepNum).sdiv(*SeqStepDenom);
 
    APInt Addend = *Elt - ExpectedVal;
    if (!SeqAddend)
      SeqAddend = Addend;
    else if (Addend != SeqAddend)
      return std::nullopt;
  }
 
  assert(SeqAddend && "Must have an addend if we have a step");
 
  return VIDSequence{SeqStepNum->getSExtValue(), *SeqStepDenom,
                     SeqAddend->getSExtValue()};
}
 
// Match a splatted value (SPLAT_VECTOR/BUILD_VECTOR) of an EXTRACT_VECTOR_ELT
// and lower it as a VRGATHER_VX_VL from the source vector.
static SDValue matchSplatAsGather(SDValue SplatVal, MVT VT, const SDLoc &DL,
                                  SelectionDAG &DAG,
                                  const RISCVSubtarget &Subtarget) {
  if (SplatVal.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
    return SDValue();
  SDValue Vec = SplatVal.getOperand(0);
  // Only perform this optimization on vectors of the same size for simplicity.
  // Don't perform this optimization for i1 vectors.
  // FIXME: Support i1 vectors, maybe by promoting to i8?
  if (Vec.getValueType() != VT || VT.getVectorElementType() == MVT::i1)
    return SDValue();
  SDValue Idx = SplatVal.getOperand(1);
  // The index must be a legal type.
  if (Idx.getValueType() != Subtarget.getXLenVT())
    return SDValue();
 
  MVT ContainerVT = VT;
  if (VT.isFixedLengthVector()) {
    ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
    Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
  }
 
  auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
 
  SDValue Gather = DAG.getNode(RISCVISD::VRGATHER_VX_VL, DL, ContainerVT, Vec,
                               Idx, DAG.getUNDEF(ContainerVT), Mask, VL);
 
  if (!VT.isFixedLengthVector())
    return Gather;
 
  return convertFromScalableVector(VT, Gather, DAG, Subtarget);
}
 
 
/// 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.  Returns SDValue if not profitable.
/// 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.
static SDValue lowerBuildVectorViaDominantValues(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);
  auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
 
  MVT XLenVT = Subtarget.getXLenVT();
  unsigned NumElts = Op.getNumOperands();
 
  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 (0 == Count)
      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 &&
      (NumElts != 2 || ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) &&
      ((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};
 
    // We can handle an insert into the last element (of a splat) via
    // v(f)slide1down.  This is slightly better than the vslideup insert
    // lowering as it avoids the need for a vector group temporary.  It
    // is also better than using vmerge.vx as it avoids the need to
    // materialize the mask in a vector register.
    if (SDValue LastOp = Op->getOperand(Op->getNumOperands() - 1);
        !LastOp.isUndef() && ValueCounts[LastOp] == 1 &&
        LastOp != DominantValue) {
      Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
      auto OpCode =
        VT.isFloatingPoint() ? RISCVISD::VFSLIDE1DOWN_VL : RISCVISD::VSLIDE1DOWN_VL;
      if (!VT.isFloatingPoint())
        LastOp = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, LastOp);
      Vec = DAG.getNode(OpCode, DL, ContainerVT, DAG.getUNDEF(ContainerVT), Vec,
                        LastOp, Mask, VL);
      Vec = convertFromScalableVector(VT, Vec, DAG, Subtarget);
      Processed.insert(LastOp);
    }
 
    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.getVectorIdxConstant(OpIdx.index(), DL));
      } 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 lowerBuildVectorOfConstants(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);
  auto [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::clamp(NumElts, 8u, Subtarget.getXLen());
    NumViaIntegerBits = std::min(NumViaIntegerBits, Subtarget.getELen());
    // 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.
    unsigned IntegerViaVecElts = divideCeil(NumElts, NumViaIntegerBits);
    MVT IntegerViaVecVT =
      MVT::getVectorVT(MVT::getIntegerVT(NumViaIntegerBits),
                       IntegerViaVecElts);
 
    uint64_t Bits = 0;
    unsigned BitPos = 0, IntegerEltIdx = 0;
    SmallVector<SDValue, 8> Elts(IntegerViaVecElts);
 
    for (unsigned I = 0; I < NumElts;) {
      SDValue V = Op.getOperand(I);
      bool BitValue = !V.isUndef() && V->getAsZExtVal();
      Bits |= ((uint64_t)BitValue << BitPos);
      ++BitPos;
      ++I;
 
      // Once we accumulate enough bits to fill our scalar type or process the
      // last element, insert into our vector and clear our accumulated data.
      if (I % NumViaIntegerBits == 0 || I == NumElts) {
        if (NumViaIntegerBits <= 32)
          Bits = SignExtend64<32>(Bits);
        SDValue Elt = DAG.getConstant(Bits, DL, XLenVT);
        Elts[IntegerEltIdx] = Elt;
        Bits = 0;
        BitPos = 0;
        IntegerEltIdx++;
      }
    }
 
    SDValue Vec = DAG.getBuildVector(IntegerViaVecVT, DL, Elts);
 
    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;
  }
 
  if (SDValue Splat = cast<BuildVectorSDNode>(Op)->getSplatValue()) {
    unsigned Opc = VT.isFloatingPoint() ? RISCVISD::VFMV_V_F_VL
                                        : RISCVISD::VMV_V_X_VL;
    if (!VT.isFloatingPoint())
      Splat = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, Splat);
    Splat =
        DAG.getNode(Opc, DL, ContainerVT, DAG.getUNDEF(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, Op.getScalarValueSizeInBits())) {
    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;
    // Exclude INT64_MIN to avoid passing it to std::abs. We won't optimize it
    // anyway as the shift of 63 won't fit in uimm5.
    if (StepNumerator != 1 && StepNumerator != INT64_MIN &&
        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) &&
        (SplatStepVal >= 0 || StepDenominator == 1) && isInt<5>(Addend)) {
      MVT VIDVT =
          VT.isFloatingPoint() ? VT.changeVectorElementTypeToInteger() : VT;
      MVT VIDContainerVT =
          getContainerForFixedLengthVector(DAG, VIDVT, Subtarget);
      SDValue VID = DAG.getNode(RISCVISD::VID_VL, DL, VIDContainerVT, 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(VIDVT, VID, DAG, Subtarget);
      if ((StepOpcode == ISD::MUL && SplatStepVal != 1) ||
          (StepOpcode == ISD::SHL && SplatStepVal != 0)) {
        SDValue SplatStep = DAG.getConstant(SplatStepVal, DL, VIDVT);
        VID = DAG.getNode(StepOpcode, DL, VIDVT, VID, SplatStep);
      }
      if (StepDenominator != 1) {
        SDValue SplatStep =
            DAG.getConstant(Log2_64(StepDenominator), DL, VIDVT);
        VID = DAG.getNode(ISD::SRL, DL, VIDVT, VID, SplatStep);
      }
      if (Addend != 0 || Negate) {
        SDValue SplatAddend = DAG.getConstant(Addend, DL, VIDVT);
        VID = DAG.getNode(Negate ? ISD::SUB : ISD::ADD, DL, VIDVT, SplatAddend,
                          VID);
      }
      if (VT.isFloatingPoint()) {
        // TODO: Use vfwcvt to reduce register pressure.
        VID = DAG.getNode(ISD::SINT_TO_FP, DL, VT, VID);
      }
      return VID;
    }
  }
 
  // For very small build_vectors, use a single scalar insert of a constant.
  // TODO: Base this on constant rematerialization cost, not size.
  const unsigned EltBitSize = VT.getScalarSizeInBits();
  if (VT.getSizeInBits() <= 32 &&
      ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
    MVT ViaIntVT = MVT::getIntegerVT(VT.getSizeInBits());
    assert((ViaIntVT == MVT::i16 || ViaIntVT == MVT::i32) &&
           "Unexpected sequence type");
    // If we can use the original VL with the modified element type, this
    // means we only have a VTYPE toggle, not a VL toggle.  TODO: Should this
    // be moved into InsertVSETVLI?
    unsigned ViaVecLen =
      (Subtarget.getRealMinVLen() >= VT.getSizeInBits() * NumElts) ? NumElts : 1;
    MVT ViaVecVT = MVT::getVectorVT(ViaIntVT, ViaVecLen);
 
    uint64_t EltMask = maskTrailingOnes<uint64_t>(EltBitSize);
    uint64_t SplatValue = 0;
    // Construct the amalgamated value at this larger vector type.
    for (const auto &OpIdx : enumerate(Op->op_values())) {
      const auto &SeqV = OpIdx.value();
      if (!SeqV.isUndef())
        SplatValue |=
            ((SeqV->getAsZExtVal() & EltMask) << (OpIdx.index() * EltBitSize));
    }
 
    // 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<32>(SplatValue);
 
    SDValue Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ViaVecVT,
                              DAG.getUNDEF(ViaVecVT),
                              DAG.getConstant(SplatValue, DL, XLenVT),
                              DAG.getVectorIdxConstant(0, DL));
    if (ViaVecLen != 1)
      Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
                        MVT::getVectorVT(ViaIntVT, 1), Vec,
                        DAG.getConstant(0, DL, XLenVT));
    return DAG.getBitcast(VT, Vec);
  }
 
 
  // 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;
  const auto *BV = cast<BuildVectorSDNode>(Op);
  if (VT.isInteger() && EltBitSize < Subtarget.getELen() &&
      ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
      BV->getRepeatedSequence(Sequence) &&
      (Sequence.size() * EltBitSize) <= Subtarget.getELen()) {
    unsigned SeqLen = Sequence.size();
    MVT ViaIntVT = MVT::getIntegerVT(EltBitSize * SeqLen);
    assert((ViaIntVT == MVT::i16 || ViaIntVT == MVT::i32 ||
            ViaIntVT == MVT::i64) &&
           "Unexpected sequence type");
 
    // If we can use the original VL with the modified element type, this
    // means we only have a VTYPE toggle, not a VL toggle.  TODO: Should this
    // be moved into InsertVSETVLI?
    const unsigned RequiredVL = NumElts / SeqLen;
    const unsigned ViaVecLen =
      (Subtarget.getRealMinVLen() >= ViaIntVT.getSizeInBits() * NumElts) ?
      NumElts : RequiredVL;
    MVT ViaVecVT = MVT::getVectorVT(ViaIntVT, ViaVecLen);
 
    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 |=
            ((SeqV->getAsZExtVal() & 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<32>(SplatValue);
 
    // 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.getUNDEF(ViaContainerVT),
                      DAG.getConstant(SplatValue, DL, XLenVT), ViaVL);
      Splat = convertFromScalableVector(ViaVecVT, Splat, DAG, Subtarget);
      if (ViaVecLen != RequiredVL)
        Splat = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
                            MVT::getVectorVT(ViaIntVT, RequiredVL), Splat,
                            DAG.getConstant(0, DL, XLenVT));
      return DAG.getBitcast(VT, Splat);
    }
  }
 
  // If the number of signbits allows, see if we can lower as a <N x i8>.
  // Our main goal here is to reduce LMUL (and thus work) required to
  // build the constant, but we will also narrow if the resulting
  // narrow vector is known to materialize cheaply.
  // TODO: We really should be costing the smaller vector.  There are
  // profitable cases this misses.
  if (EltBitSize > 8 && VT.isInteger() &&
      (NumElts <= 4 || VT.getSizeInBits() > Subtarget.getRealMinVLen())) {
    unsigned SignBits = DAG.ComputeNumSignBits(Op);
    if (EltBitSize - SignBits < 8) {
      SDValue Source = DAG.getBuildVector(VT.changeVectorElementType(MVT::i8),
                                          DL, Op->ops());
      Source = convertToScalableVector(ContainerVT.changeVectorElementType(MVT::i8),
                                       Source, DAG, Subtarget);
      SDValue Res = DAG.getNode(RISCVISD::VSEXT_VL, DL, ContainerVT, Source, Mask, VL);
      return convertFromScalableVector(VT, Res, DAG, Subtarget);
    }
  }
 
  if (SDValue Res = lowerBuildVectorViaDominantValues(Op, DAG, Subtarget))
    return Res;
 
  // For constant vectors, use generic constant pool lowering.  Otherwise,
  // we'd have to materialize constants in GPRs just to move them into the
  // vector.
  return SDValue();
}
 
static unsigned getPACKOpcode(unsigned DestBW,
                              const RISCVSubtarget &Subtarget) {
  switch (DestBW) {
  default:
    llvm_unreachable("Unsupported pack size");
  case 16:
    return RISCV::PACKH;
  case 32:
    return Subtarget.is64Bit() ? RISCV::PACKW : RISCV::PACK;
  case 64:
    assert(Subtarget.is64Bit());
    return RISCV::PACK;
  }
}
 
/// Double the element size of the build vector to reduce the number
/// of vslide1down in the build vector chain.  In the worst case, this
/// trades three scalar operations for 1 vector operation.  Scalar
/// operations are generally lower latency, and for out-of-order cores
/// we also benefit from additional parallelism.
static SDValue lowerBuildVectorViaPacking(SDValue Op, SelectionDAG &DAG,
                                          const RISCVSubtarget &Subtarget) {
  SDLoc DL(Op);
  MVT VT = Op.getSimpleValueType();
  assert(VT.isFixedLengthVector() && "Unexpected vector!");
  MVT ElemVT = VT.getVectorElementType();
  if (!ElemVT.isInteger())
    return SDValue();
 
  // TODO: Relax these architectural restrictions, possibly with costing
  // of the actual instructions required.
  if (!Subtarget.hasStdExtZbb() || !Subtarget.hasStdExtZba())
    return SDValue();
 
  unsigned NumElts = VT.getVectorNumElements();
  unsigned ElemSizeInBits = ElemVT.getSizeInBits();
  if (ElemSizeInBits >= std::min(Subtarget.getELen(), Subtarget.getXLen()) ||
      NumElts % 2 != 0)
    return SDValue();
 
  // Produce [B,A] packed into a type twice as wide.  Note that all
  // scalars are XLenVT, possibly masked (see below).
  MVT XLenVT = Subtarget.getXLenVT();
  SDValue Mask = DAG.getConstant(
      APInt::getLowBitsSet(XLenVT.getSizeInBits(), ElemSizeInBits), DL, XLenVT);
  auto pack = [&](SDValue A, SDValue B) {
    // Bias the scheduling of the inserted operations to near the
    // definition of the element - this tends to reduce register
    // pressure overall.
    SDLoc ElemDL(B);
    if (Subtarget.hasStdExtZbkb())
      // Note that we're relying on the high bits of the result being
      // don't care.  For PACKW, the result is *sign* extended.
      return SDValue(
          DAG.getMachineNode(getPACKOpcode(ElemSizeInBits * 2, Subtarget),
                             ElemDL, XLenVT, A, B),
          0);
 
    A = DAG.getNode(ISD::AND, SDLoc(A), XLenVT, A, Mask);
    B = DAG.getNode(ISD::AND, SDLoc(B), XLenVT, B, Mask);
    SDValue ShtAmt = DAG.getConstant(ElemSizeInBits, ElemDL, XLenVT);
    SDNodeFlags Flags;
    Flags.setDisjoint(true);
    return DAG.getNode(ISD::OR, ElemDL, XLenVT, A,
                       DAG.getNode(ISD::SHL, ElemDL, XLenVT, B, ShtAmt), Flags);
  };
 
  SmallVector<SDValue> NewOperands;
  NewOperands.reserve(NumElts / 2);
  for (unsigned i = 0; i < VT.getVectorNumElements(); i += 2)
    NewOperands.push_back(pack(Op.getOperand(i), Op.getOperand(i + 1)));
  assert(NumElts == NewOperands.size() * 2);
  MVT WideVT = MVT::getIntegerVT(ElemSizeInBits * 2);
  MVT WideVecVT = MVT::getVectorVT(WideVT, NumElts / 2);
  return DAG.getNode(ISD::BITCAST, DL, VT,
                     DAG.getBuildVector(WideVecVT, DL, NewOperands));
}
 
// Convert to an vXf16 build_vector to vXi16 with bitcasts.
static SDValue lowerBUILD_VECTORvXf16(SDValue Op, SelectionDAG &DAG) {
  MVT VT = Op.getSimpleValueType();
  MVT IVT = VT.changeVectorElementType(MVT::i16);
  SmallVector<SDValue, 16> NewOps(Op.getNumOperands());
  for (unsigned I = 0, E = Op.getNumOperands(); I != E; ++I)
    NewOps[I] = DAG.getBitcast(MVT::i16, Op.getOperand(I));
  SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), IVT, NewOps);
  return DAG.getBitcast(VT, Res);
}
 
static SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
                                 const RISCVSubtarget &Subtarget) {
  MVT VT = Op.getSimpleValueType();
  assert(VT.isFixedLengthVector() && "Unexpected vector!");
 
  // If we don't have scalar f16, we need to bitcast to an i16 vector.
  if (VT.getVectorElementType() == MVT::f16 &&
      !Subtarget.hasStdExtZfhmin())
    return lowerBUILD_VECTORvXf16(Op, DAG);
 
  if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) ||
      ISD::isBuildVectorOfConstantFPSDNodes(Op.getNode()))
    return lowerBuildVectorOfConstants(Op, DAG, Subtarget);
 
  MVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
 
  SDLoc DL(Op);
  auto [Mask, VL] = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget);
 
  MVT XLenVT = Subtarget.getXLenVT();
 
  if (VT.getVectorElementType() == MVT::i1) {
    // 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.
      Splat = DAG.getNode(ISD::AND, DL, Splat.getValueType(), Splat,
                          DAG.getConstant(1, DL, Splat.getValueType()));
      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()) {
    if (auto Gather = matchSplatAsGather(Splat, VT, DL, DAG, Subtarget))
      return Gather;
    unsigned Opc = VT.isFloatingPoint() ? RISCVISD::VFMV_V_F_VL
                                        : RISCVISD::VMV_V_X_VL;
    if (!VT.isFloatingPoint())
      Splat = DAG.getNode(ISD::ANY_EXTEND, DL, XLenVT, Splat);
    Splat =
        DAG.getNode(Opc, DL, ContainerVT, DAG.getUNDEF(ContainerVT), Splat, VL);
    return convertFromScalableVector(VT, Splat, DAG, Subtarget);
  }
 
  if (SDValue Res = lowerBuildVectorViaDominantValues(Op, DAG, Subtarget))
    return Res;
 
  // If we're compiling for an exact VLEN value, we can split our work per
  // register in the register group.
  if (const auto VLen = Subtarget.getRealVLen();
      VLen && VT.getSizeInBits().getKnownMinValue() > *VLen) {
    MVT ElemVT = VT.getVectorElementType();
    unsigned ElemsPerVReg = *VLen / ElemVT.getFixedSizeInBits();
    EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT, Subtarget);
    MVT OneRegVT = MVT::getVectorVT(ElemVT, ElemsPerVReg);
    MVT M1VT = getContainerForFixedLengthVector(DAG, OneRegVT, Subtarget);
    assert(M1VT == getLMUL1VT(M1VT));
 
    // The following semantically builds up a fixed length concat_vector
    // of the component build_vectors.  We eagerly lower to scalable and
    // insert_subvector here to avoid DAG combining it back to a large
    // build_vector.
    SmallVector<SDValue> BuildVectorOps(Op->ops());
    unsigned NumOpElts = M1VT.getVectorMinNumElements();
    SDValue Vec = DAG.getUNDEF(ContainerVT);
    for (unsigned i = 0; i < VT.getVectorNumElements(); i += ElemsPerVReg) {
      auto OneVRegOfOps = ArrayRef(BuildVectorOps).slice(i, ElemsPerVReg);
      SDValue SubBV =
          DAG.getNode(ISD::BUILD_VECTOR, DL, OneRegVT, OneVRegOfOps);
      SubBV = convertToScalableVector(M1VT, SubBV, DAG, Subtarget);
      unsigned InsertIdx = (i / ElemsPerVReg) * NumOpElts;
      Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ContainerVT, Vec, SubBV,
                        DAG.getVectorIdxConstant(InsertIdx, DL));
    }
    return convertFromScalableVector(VT, Vec, DAG, Subtarget);
  }
 
  // If we're about to resort to vslide1down (or stack usage), pack our
  // elements into the widest scalar type we can.  This will force a VL/VTYPE
  // toggle, but reduces the critical path, the number of vslide1down ops
  // required, and possibly enables scalar folds of the values.
  if (SDValue Res = lowerBuildVectorViaPacking(Op, DAG, Subtarget))
    return Res;
 
  // For m1 vectors, if we have non-undef values in both halves of our vector,
  // split the vector into low and high halves, build them separately, then
  // use a vselect to combine them.  For long vectors, this cuts the critical
  // path of the vslide1down sequence in half, and gives us an opportunity
  // to special case each half independently.  Note that we don't change the
  // length of the sub-vectors here, so if both fallback to the generic
  // vslide1down path, we should be able to fold the vselect into the final
  // vslidedown (for the undef tail) for the first half w/ masking.
  unsigned NumElts = VT.getVectorNumElements();
  unsigned NumUndefElts =
      count_if(Op->op_values(), [](const SDValue &V) { return V.isUndef(); });
  unsigned NumDefElts = NumElts - NumUndefElts;
  if (NumDefElts >= 8 && NumDefElts > NumElts / 2 &&
      ContainerVT.bitsLE(getLMUL1VT(ContainerVT))) {
    SmallVector<SDValue> SubVecAOps, SubVecBOps;
    SmallVector<SDValue> MaskVals;
    SDValue UndefElem = DAG.getUNDEF(Op->getOperand(0)->getValueType(0));
    SubVecAOps.reserve(NumElts);
    SubVecBOps.reserve(NumElts);
    for (unsigned i = 0; i < NumElts; i++) {
      SDValue Elem = Op->getOperand(i);
      if (i < NumElts / 2) {
        SubVecAOps.push_back(Elem);
        SubVecBOps.push_back(UndefElem);
      } else {
        SubVecAOps.push_back(UndefElem);
        SubVecBOps.push_back(Elem);
      }
      bool SelectMaskVal = (i < NumElts / 2);
      MaskVals.push_back(DAG.getConstant(SelectMaskVal, DL, XLenVT));
    }
    assert(SubVecAOps.size() == NumElts && SubVecBOps.size() == NumElts &&
           MaskVals.size() == NumElts);
 
    SDValue SubVecA = DAG.getBuildVector(VT, DL, SubVecAOps);
    SDValue SubVecB = DAG.getBuildVector(VT, DL, SubVecBOps);
    MVT MaskVT = MVT::getVectorVT(MVT::i1, NumElts);
    SDValue SelectMask = DAG.getBuildVector(MaskVT, DL, MaskVals);
    return DAG.getNode(ISD::VSELECT, DL, VT, SelectMask, SubVecA, SubVecB);
  }
 
  // Cap the cost at a value linear to the number of elements in the vector.
  // The default lowering is to use the stack.  The vector store + scalar loads
  // is linear in VL.  However, at high lmuls vslide1down and vslidedown end up
  // being (at least) linear in LMUL.  As a result, using the vslidedown
  // lowering for every element ends up being VL*LMUL..
  // TODO: Should we be directly costing the stack alternative?  Doing so might
  // give us a more accurate upper bound.
  InstructionCost LinearBudget = VT.getVectorNumElements() * 2;
 
  // TODO: unify with TTI getSlideCost.
  InstructionCost PerSlideCost = 1;
  switch (RISCVTargetLowering::getLMUL(ContainerVT)) {
  default: break;
  case RISCVII::VLMUL::LMUL_2:
    PerSlideCost = 2;
    break;
  case RISCVII::VLMUL::LMUL_4:
    PerSlideCost = 4;
    break;
  case RISCVII::VLMUL::LMUL_8:
    PerSlideCost = 8;
    break;
  }
 
  // TODO: Should we be using the build instseq then cost + evaluate scheme
  // we use for integer constants here?
  unsigned UndefCount = 0;
  for (const SDValue &V : Op->ops()) {
    if (V.isUndef()) {
      UndefCount++;
      continue;
    }
    if (UndefCount) {
      LinearBudget -= PerSlideCost;
      UndefCount = 0;
    }
    LinearBudget -= PerSlideCost;
  }
  if (UndefCount) {
    LinearBudget -= PerSlideCost;
  }
 
  if (LinearBudget < 0)
    return SDValue();
 
  assert((!VT.isFloatingPoint() ||
          VT.getVectorElementType().getSizeInBits() <= Subtarget.getFLen()) &&
         "Illegal type which will result in reserved encoding");
 
  const unsigned Policy = RISCVII::TAIL_AGNOSTIC | RISCVII::MASK_AGNOSTIC;
 
  SDValue Vec;
  UndefCount = 0;
  for (SDValue V : Op->ops()) {
    if (V.isUndef()) {
      UndefCount++;
      continue;
    }
 
    // Start our sequence with a TA splat in the hopes that hardware is able to
    // recognize there's no dependency on the prior value of our temporary
    // register.
    if (!Vec) {
      Vec = DAG.getSplatVector(VT, DL, V);
      Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget);
      UndefCount = 0;
      continue;
    }
 
    if (UndefCount) {
      const SDValue Offset = DAG.getConstant(UndefCount, DL, Subtarget.getXLenVT());
      Vec = getVSlidedown(DAG, Subtarget, DL, ContainerVT, DAG.getUNDEF(ContainerVT),
                          Vec, Offset, Mask, VL, Policy);
      UndefCount = 0;
    }
    auto OpCode =
      VT.isFloatingPoint() ? RISCVISD::VFSLIDE1DOWN_VL : RISCVISD::VSLIDE1DOWN_VL;
    if (!VT.isFloatingPoint())
      V = DAG.getNode(ISD::ANY_EXTEND, DL, Subtarget.getXLenVT(), V);
    Vec = DAG.getNode(OpCode, DL, ContainerVT, DAG.getUNDEF(ContainerVT), Vec,
                      V, Mask, VL);
  }
  if (UndefCount) {
    const SDValue Offset = DAG.getConstant(UndefCount, DL, Subtarget.getXLenVT());
    Vec = getVSlidedown(DAG, Subtarget, DL, ContainerVT, DAG.getUNDEF(ContainerVT),
                        Vec, Offset, Mask, VL, Policy);
  }
  return convertFromScalableVector(VT, Vec, DAG, Subtarget);
}
 
static SDValue splatPartsI64WithVL(const SDLoc &DL, MVT VT, SDValue Passthru,
                                   SDValue Lo, SDValue Hi, SDValue VL,
                                   SelectionDAG &DAG) {
  if (!Passthru)
    Passthru = DAG.getUNDEF(VT);
  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, Passthru, Lo, VL);
 
    // If vl is equal to VLMAX or fits in 4 bits and Hi constant is equal to Lo,
    // we could use vmv.v.x whose EEW = 32 to lower it. This allows us to use
    // vlmax vsetvli or vsetivli to change the VL.
    // FIXME: Support larger constants?
    // FIXME: Support non-constant VLs by saturating?
    if (LoC == HiC) {
      SDValue NewVL;
      if (isAllOnesConstant(VL) ||
          (isa<RegisterSDNode>(VL) &&
           cast<RegisterSDNode>(VL)->getReg() == RISCV::X0))
        NewVL = DAG.getRegister(RISCV::X0, MVT::i32);
      else if (isa<ConstantSDNode>(VL) && isUInt<4>(VL->getAsZExtVal()))
        NewVL = DAG.getNode(ISD::ADD, DL, VL.getValueType(), VL, VL);
 
      if (NewVL) {
        MVT InterVT =
            MVT::getVectorVT(MVT::i32, VT.getVectorElementCount() * 2);
        auto InterVec = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, InterVT,
                                    DAG.getUNDEF(InterVT), Lo, NewVL);
        return DAG.getNode(ISD::BITCAST, DL, VT, InterVec);
      }
    }
  }
 
  // Detect cases where Hi is (SRA Lo, 31) which means Hi is Lo sign extended.
  if (Hi.getOpcode() == ISD::SRA && Hi.getOperand(0) == Lo &&
      isa<ConstantSDNode>(Hi.getOperand(1)) &&
      Hi.getConstantOperandVal(1) == 31)
    return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, Passthru, Lo, VL);
 
  // If the hi bits of the splat are undefined, then it's fine to just splat Lo
  // even if it might be sign extended.
  if (Hi.isUndef())
    return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, VT, Passthru, Lo, VL);
 
  // Fall back to a stack store and stride x0 vector load.
  return DAG.getNode(RISCVISD::SPLAT_VECTOR_SPLIT_I64_VL, DL, VT, Passthru, 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 Passthru,
</