RISCV.cpp

Go to the documentation of this file.

//===- RISCV.cpp ----------------------------------------------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
 
#include "ABIInfoImpl.h"
#include "TargetInfo.h"
#include "llvm/IR/IntrinsicsRISCV.h"
#include "llvm/TargetParser/RISCVTargetParser.h"
 
using namespace clang;
using namespace clang::CodeGen;
 
//===----------------------------------------------------------------------===//
// RISC-V ABI Implementation
//===----------------------------------------------------------------------===//
 
namespace {
class RISCVABIInfo : public DefaultABIInfo {
private:
  // Size of the integer ('x') registers in bits.
  unsigned XLen;
  // Size of the floating point ('f') registers in bits. Note that the target
  // ISA might have a wider FLen than the selected ABI (e.g. an RV32IF target
  // with soft float ABI has FLen==0).
  unsigned FLen;
  const int NumArgGPRs;
  const int NumArgFPRs;
  const bool EABI;
  bool detectFPCCEligibleStructHelper(QualType Ty, CharUnits CurOff,
                                      llvm::Type *&Field1Ty,
                                      CharUnits &Field1Off,
                                      llvm::Type *&Field2Ty,
                                      CharUnits &Field2Off) const;
 
  llvm::Type *detectVLSCCEligibleStruct(QualType Ty, unsigned ABIVLen) const;
 
  llvm::Type *detectHomogeneousRVVFixedLengthStruct(QualType Ty) const;
 
  unsigned getRVVFixedLengthMinElts(const VectorType *VT) const;
 
public:
  RISCVABIInfo(CodeGen::CodeGenTypes &CGT, unsigned XLen, unsigned FLen,
               bool EABI)
      : DefaultABIInfo(CGT), XLen(XLen), FLen(FLen), NumArgGPRs(EABI ? 6 : 8),
        NumArgFPRs(FLen != 0 ? 8 : 0), EABI(EABI) {}
 
  // DefaultABIInfo's classifyReturnType and classifyArgumentType are
  // non-virtual, but computeInfo is virtual, so we overload it.
  void computeInfo(CGFunctionInfo &FI) const override;
 
  ABIArgInfo classifyArgumentType(QualType Ty, bool IsFixed, int &ArgGPRsLeft,
                                  int &ArgFPRsLeft, unsigned ABIVLen) const;
  ABIArgInfo classifyReturnType(QualType RetTy, unsigned ABIVLen) const;
 
  RValue EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, QualType Ty,
                   AggValueSlot Slot) const override;
 
  ABIArgInfo extendType(QualType Ty, llvm::Type *CoerceTy = nullptr) const;
 
  bool detectFPCCEligibleStruct(QualType Ty, llvm::Type *&Field1Ty,
                                CharUnits &Field1Off, llvm::Type *&Field2Ty,
                                CharUnits &Field2Off, int &NeededArgGPRs,
                                int &NeededArgFPRs) const;
  ABIArgInfo coerceAndExpandFPCCEligibleStruct(llvm::Type *Field1Ty,
                                               CharUnits Field1Off,
                                               llvm::Type *Field2Ty,
                                               CharUnits Field2Off) const;
 
  ABIArgInfo coerceVLSVector(QualType Ty, unsigned ABIVLen = 0) const;
  // Some unsupported type e.g. bf16 without zvfbfmin or zvfbfa, should be
  // passed as same size i8 type. This function check and return the appropriate
  // fixed vector type.
  llvm::FixedVectorType *
  getVLSCCCompatibleType(llvm::FixedVectorType *FixedVecTy) const;
 
  using ABIInfo::appendAttributeMangling;
  void appendAttributeMangling(TargetClonesAttr *Attr, unsigned Index,
                               raw_ostream &Out) const override;
  void appendAttributeMangling(StringRef AttrStr,
                               raw_ostream &Out) const override;
  llvm::Value *createCoercedLoad(Address SrcAddr, const ABIArgInfo &AI,
                                 CodeGenFunction &CGF) const override;
  void createCoercedStore(llvm::Value *Val, Address DstAddr,
                          const ABIArgInfo &AI, bool DestIsVolatile,
                          CodeGenFunction &CGF) const override;
};
} // end anonymous namespace
 
void RISCVABIInfo::appendAttributeMangling(TargetClonesAttr *Attr,
                                           unsigned Index,
                                           raw_ostream &Out) const {
  appendAttributeMangling(Attr->getFeatureStr(Index), Out);
}
 
void RISCVABIInfo::appendAttributeMangling(StringRef AttrStr,
                                           raw_ostream &Out) const {
  if (AttrStr == "default") {
    Out << ".default";
    return;
  }
 
  Out << '.';
 
  SmallVector<StringRef, 8> Attrs;
  AttrStr.split(Attrs, ';');
 
  // Only consider the arch string.
  StringRef ArchStr;
  for (auto &Attr : Attrs) {
    if (Attr.starts_with("arch="))
      ArchStr = Attr;
  }
 
  // Extract features string.
  SmallVector<StringRef, 8> Features;
  ArchStr.consume_front("arch=");
  ArchStr.split(Features, ',');
 
  llvm::stable_sort(Features);
 
  for (auto Feat : Features) {
    Feat.consume_front("+");
    Out << "_" << Feat;
  }
}
 
void RISCVABIInfo::computeInfo(CGFunctionInfo &FI) const {
  unsigned ABIVLen;
  switch (FI.getExtInfo().getCC()) {
  default:
    ABIVLen = 0;
    break;
#define CC_VLS_CASE(ABI_VLEN)                                                  \
  case CallingConv::CC_RISCVVLSCall_##ABI_VLEN:                                \
    ABIVLen = ABI_VLEN;                                                        \
    break;
    CC_VLS_CASE(32)
    CC_VLS_CASE(64)
    CC_VLS_CASE(128)
    CC_VLS_CASE(256)
    CC_VLS_CASE(512)
    CC_VLS_CASE(1024)
    CC_VLS_CASE(2048)
    CC_VLS_CASE(4096)
    CC_VLS_CASE(8192)
    CC_VLS_CASE(16384)
    CC_VLS_CASE(32768)
    CC_VLS_CASE(65536)
#undef CC_VLS_CASE
  }
  QualType RetTy = FI.getReturnType();
  if (!getCXXABI().classifyReturnType(FI))
    FI.getReturnInfo() = classifyReturnType(RetTy, ABIVLen);
 
  // IsRetIndirect is true if classifyArgumentType indicated the value should
  // be passed indirect, or if the type size is a scalar greater than 2*XLen
  // and not a complex type with elements <= FLen. e.g. fp128 is passed direct
  // in LLVM IR, relying on the backend lowering code to rewrite the argument
  // list and pass indirectly on RV32.
  bool IsRetIndirect = FI.getReturnInfo().getKind() == ABIArgInfo::Indirect;
  if (!IsRetIndirect && RetTy->isScalarType() &&
      getContext().getTypeSize(RetTy) > (2 * XLen)) {
    if (RetTy->isComplexType() && FLen) {
      QualType EltTy = RetTy->castAs<ComplexType>()->getElementType();
      IsRetIndirect = getContext().getTypeSize(EltTy) > FLen;
    } else {
      // This is a normal scalar > 2*XLen, such as fp128 on RV32.
      IsRetIndirect = true;
    }
  }
 
  int ArgGPRsLeft = IsRetIndirect ? NumArgGPRs - 1 : NumArgGPRs;
  int ArgFPRsLeft = NumArgFPRs;
  int NumFixedArgs = FI.getNumRequiredArgs();
 
  int ArgNum = 0;
  for (auto &ArgInfo : FI.arguments()) {
    bool IsFixed = ArgNum < NumFixedArgs;
    ArgInfo.info = classifyArgumentType(ArgInfo.type, IsFixed, ArgGPRsLeft,
                                        ArgFPRsLeft, ABIVLen);
    ArgNum++;
  }
}
 
// Returns true if the struct is a potential candidate for the floating point
// calling convention. If this function returns true, the caller is
// responsible for checking that if there is only a single field then that
// field is a float.
bool RISCVABIInfo::detectFPCCEligibleStructHelper(QualType Ty, CharUnits CurOff,
                                                  llvm::Type *&Field1Ty,
                                                  CharUnits &Field1Off,
                                                  llvm::Type *&Field2Ty,
                                                  CharUnits &Field2Off) const {
  bool IsInt = Ty->isIntegralOrEnumerationType();
  bool IsFloat = Ty->isRealFloatingType();
 
  if (IsInt || IsFloat) {
    uint64_t Size = getContext().getTypeSize(Ty);
    if (IsInt && Size > XLen)
      return false;
    // Can't be eligible if larger than the FP registers. Handling of half
    // precision values has been specified in the ABI, so don't block those.
    if (IsFloat && Size > FLen)
      return false;
    // Can't be eligible if an integer type was already found (int+int pairs
    // are not eligible).
    if (IsInt && Field1Ty && Field1Ty->isIntegerTy())
      return false;
    if (!Field1Ty) {
      Field1Ty = CGT.ConvertType(Ty);
      Field1Off = CurOff;
      return true;
    }
    if (!Field2Ty) {
      Field2Ty = CGT.ConvertType(Ty);
      Field2Off = CurOff;
      return true;
    }
    return false;
  }
 
  if (auto CTy = Ty->getAs<ComplexType>()) {
    if (Field1Ty)
      return false;
    QualType EltTy = CTy->getElementType();
    // Only floating-point complex types (e.g. _Complex float/double) are
    // eligible to be passed in floating-point argument registers. Complex
    // integer types (a GNU extension) should be treated like a normal
    // aggregate and packed into GPRs instead.
    if (!EltTy->isRealFloatingType())
      return false;
    if (getContext().getTypeSize(EltTy) > FLen)
      return false;
    Field1Ty = CGT.ConvertType(EltTy);
    Field1Off = CurOff;
    Field2Ty = Field1Ty;
    Field2Off = Field1Off + getContext().getTypeSizeInChars(EltTy);
    return true;
  }
 
  if (const ConstantArrayType *ATy = getContext().getAsConstantArrayType(Ty)) {
    uint64_t ArraySize = ATy->getZExtSize();
    QualType EltTy = ATy->getElementType();
    // Non-zero-length arrays of empty records make the struct ineligible for
    // the FP calling convention in C++.
    if (const auto *RTy = EltTy->getAsCanonical<RecordType>()) {
      if (ArraySize != 0 && isa<CXXRecordDecl>(RTy->getDecl()) &&
          isEmptyRecord(getContext(), EltTy, true, true))
        return false;
    }
    CharUnits EltSize = getContext().getTypeSizeInChars(EltTy);
    for (uint64_t i = 0; i < ArraySize; ++i) {
      bool Ret = detectFPCCEligibleStructHelper(EltTy, CurOff, Field1Ty,
                                                Field1Off, Field2Ty, Field2Off);
      if (!Ret)
        return false;
      CurOff += EltSize;
    }
    return true;
  }
 
  if (const auto *RTy = Ty->getAsCanonical<RecordType>()) {
    // Structures with either a non-trivial destructor or a non-trivial
    // copy constructor are not eligible for the FP calling convention.
    if (getRecordArgABI(Ty, CGT.getCXXABI()))
      return false;
    if (isEmptyRecord(getContext(), Ty, true, true))
      return true;
    const RecordDecl *RD = RTy->getDecl()->getDefinitionOrSelf();
    // Unions aren't eligible unless they're empty (which is caught above).
    if (RD->isUnion())
      return false;
    const ASTRecordLayout &Layout = getContext().getASTRecordLayout(RD);
    // If this is a C++ record, check the bases first.
    if (const CXXRecordDecl *CXXRD = dyn_cast<CXXRecordDecl>(RD)) {
      for (const CXXBaseSpecifier &B : CXXRD->bases()) {
        const auto *BDecl = B.getType()->castAsCXXRecordDecl();
        CharUnits BaseOff = Layout.getBaseClassOffset(BDecl);
        bool Ret = detectFPCCEligibleStructHelper(B.getType(), CurOff + BaseOff,
                                                  Field1Ty, Field1Off, Field2Ty,
                                                  Field2Off);
        if (!Ret)
          return false;
      }
    }
    int ZeroWidthBitFieldCount = 0;
    for (const FieldDecl *FD : RD->fields()) {
      uint64_t FieldOffInBits = Layout.getFieldOffset(FD->getFieldIndex());
      QualType QTy = FD->getType();
      if (FD->isBitField()) {
        unsigned BitWidth = FD->getBitWidthValue();
        // Allow a bitfield with a type greater than XLen as long as the
        // bitwidth is XLen or less.
        if (getContext().getTypeSize(QTy) > XLen && BitWidth <= XLen)
          QTy = getContext().getIntTypeForBitwidth(XLen, false);
        // Trim type to bitwidth if possible
        else if (getContext().getTypeSize(QTy) > BitWidth) {
          bool IsSigned =
              FD->getType().getTypePtr()->hasSignedIntegerRepresentation();
          unsigned Bits = std::max(8U, (unsigned)llvm::PowerOf2Ceil(BitWidth));
          QTy = getContext().getIntTypeForBitwidth(Bits, IsSigned);
        }
        if (BitWidth == 0) {
          ZeroWidthBitFieldCount++;
          continue;
        }
      }
 
      bool Ret = detectFPCCEligibleStructHelper(
          QTy, CurOff + getContext().toCharUnitsFromBits(FieldOffInBits),
          Field1Ty, Field1Off, Field2Ty, Field2Off);
      if (!Ret)
        return false;
 
      // As a quirk of the ABI, zero-width bitfields aren't ignored for fp+fp
      // or int+fp structs, but are ignored for a struct with an fp field and
      // any number of zero-width bitfields.
      if (Field2Ty && ZeroWidthBitFieldCount > 0)
        return false;
    }
    return Field1Ty != nullptr;
  }
 
  return false;
}
 
// Determine if a struct is eligible for passing according to the floating
// point calling convention (i.e., when flattened it contains a single fp
// value, fp+fp, or int+fp of appropriate size). If so, NeededArgFPRs and
// NeededArgGPRs are incremented appropriately.
bool RISCVABIInfo::detectFPCCEligibleStruct(QualType Ty, llvm::Type *&Field1Ty,
                                            CharUnits &Field1Off,
                                            llvm::Type *&Field2Ty,
                                            CharUnits &Field2Off,
                                            int &NeededArgGPRs,
                                            int &NeededArgFPRs) const {
  Field1Ty = nullptr;
  Field2Ty = nullptr;
  NeededArgGPRs = 0;
  NeededArgFPRs = 0;
  bool IsCandidate = detectFPCCEligibleStructHelper(
      Ty, CharUnits::Zero(), Field1Ty, Field1Off, Field2Ty, Field2Off);
  if (!Field1Ty)
    return false;
  // Not really a candidate if we have a single int but no float.
  if (Field1Ty && !Field2Ty && !Field1Ty->isFloatingPointTy())
    return false;
  if (!IsCandidate)
    return false;
  if (Field1Ty && Field1Ty->isFloatingPointTy())
    NeededArgFPRs++;
  else if (Field1Ty)
    NeededArgGPRs++;
  if (Field2Ty && Field2Ty->isFloatingPointTy())
    NeededArgFPRs++;
  else if (Field2Ty)
    NeededArgGPRs++;
  return true;
}
 
// Call getCoerceAndExpand for the two-element flattened struct described by
// Field1Ty, Field1Off, Field2Ty, Field2Off. This method will create an
// appropriate coerceToType and unpaddedCoerceToType.
ABIArgInfo RISCVABIInfo::coerceAndExpandFPCCEligibleStruct(
    llvm::Type *Field1Ty, CharUnits Field1Off, llvm::Type *Field2Ty,
    CharUnits Field2Off) const {
  SmallVector<llvm::Type *, 3> CoerceElts;
  SmallVector<llvm::Type *, 2> UnpaddedCoerceElts;
  if (!Field1Off.isZero())
    CoerceElts.push_back(llvm::ArrayType::get(
        llvm::Type::getInt8Ty(getVMContext()), Field1Off.getQuantity()));
 
  CoerceElts.push_back(Field1Ty);
  UnpaddedCoerceElts.push_back(Field1Ty);
 
  if (!Field2Ty) {
    return ABIArgInfo::getCoerceAndExpand(
        llvm::StructType::get(getVMContext(), CoerceElts, !Field1Off.isZero()),
        UnpaddedCoerceElts[0]);
  }
 
  CharUnits Field2Align =
      CharUnits::fromQuantity(getDataLayout().getABITypeAlign(Field2Ty));
  CharUnits Field1End = Field1Off +
      CharUnits::fromQuantity(getDataLayout().getTypeStoreSize(Field1Ty));
  CharUnits Field2OffNoPadNoPack = Field1End.alignTo(Field2Align);
 
  CharUnits Padding = CharUnits::Zero();
  if (Field2Off > Field2OffNoPadNoPack)
    Padding = Field2Off - Field2OffNoPadNoPack;
  else if (Field2Off != Field2Align && Field2Off > Field1End)
    Padding = Field2Off - Field1End;
 
  bool IsPacked = !Field2Off.isMultipleOf(Field2Align);
 
  if (!Padding.isZero())
    CoerceElts.push_back(llvm::ArrayType::get(
        llvm::Type::getInt8Ty(getVMContext()), Padding.getQuantity()));
 
  CoerceElts.push_back(Field2Ty);
  UnpaddedCoerceElts.push_back(Field2Ty);
 
  auto CoerceToType =
      llvm::StructType::get(getVMContext(), CoerceElts, IsPacked);
  auto UnpaddedCoerceToType =
      llvm::StructType::get(getVMContext(), UnpaddedCoerceElts, IsPacked);
 
  return ABIArgInfo::getCoerceAndExpand(CoerceToType, UnpaddedCoerceToType);
}
 
/// Minimum element count of the scalable vector an RVV fixed-length *data*
/// vector is coerced to, i.e. the N in <vscale x N x T>. The LMUL comes from
/// the vector kind - the type the typedef was written against - rather than
/// from VLEN and the element count.
static unsigned getRVVFixedLengthDataMinElts(VectorKind VK,
                                             unsigned EltSizeInBits) {
  // The underlying scalable type holds LMUL * RVVBitsPerBlock bits at
  // vscale=1.
  int LMulLog2 = getRVVFixedLengthDataLMulLog2(VK);
  unsigned SizeAtVScale1 = LMulLog2 >= 0
                               ? llvm::RISCV::RVVBitsPerBlock << LMulLog2
                               : llvm::RISCV::RVVBitsPerBlock >> -LMulLog2;
  assert(SizeAtVScale1 >= EltSizeInBits && "LMUL too small for element type");
  return SizeAtVScale1 / EltSizeInBits;
}
 
/// Minimum element count of the scalable vector a vbool<R>_t mask is coerced
/// to, i.e. the N in <vscale x N x i1>.
static unsigned getRVVFixedLengthMaskMinElts(VectorKind VK) {
  return llvm::RISCV::RVVBitsPerBlock >> getRVVFixedLengthMaskRatioLog2(VK);
}
 
/// Minimum element count of the scalable vector \p VT is coerced to, in units
/// of \p VT's own Clang element type. Masks are modelled as i8 vectors, so
/// their i1 lane count is rounded up to whole bytes.
unsigned RISCVABIInfo::getRVVFixedLengthMinElts(const VectorType *VT) const {
  unsigned EltSizeInBits = getContext().getTypeSize(VT->getElementType());
  if (isRVVFixedLengthDataVectorKind(VT->getVectorKind()))
    return getRVVFixedLengthDataMinElts(VT->getVectorKind(), EltSizeInBits);
  assert(isRVVFixedLengthMaskVectorKind(VT->getVectorKind()) &&
         "not an RVV fixed-length vector");
  return llvm::divideCeil(getRVVFixedLengthMaskMinElts(VT->getVectorKind()),
                          EltSizeInBits);
}
 
llvm::Type *RISCVABIInfo::detectVLSCCEligibleStruct(QualType Ty,
                                                    unsigned ABIVLen) const {
  // No riscv_vls_cc attribute.
  if (ABIVLen == 0)
    return nullptr;
 
  // Legal struct for VLS calling convention should fulfill following rules:
  // 1. Struct element should be either "homogeneous fixed-length vectors" or "a
  //    fixed-length vector array".
  // 2. Number of struct elements or array elements should be greater or equal
  //    to 1 and less or equal to 8
  // 3. Total number of vector registers needed should not exceed 8.
  //
  // Examples: Assume ABI_VLEN = 128.
  // These are legal structs:
  //   a. Structs with 1~8 "same" fixed-length vectors, e.g.
  //   struct {
  //     __attribute__((vector_size(16))) int a;
  //     __attribute__((vector_size(16))) int b;
  //   }
  //
  //   b. Structs with "single" fixed-length vector array with lengh 1~8, e.g.
  //   struct {
  //     __attribute__((vector_size(16))) int a[3];
  //   }
  // These are illegal structs:
  //   a. Structs with 9 fixed-length vectors, e.g.
  //   struct {
  //     __attribute__((vector_size(16))) int a;
  //     __attribute__((vector_size(16))) int b;
  //     __attribute__((vector_size(16))) int c;
  //     __attribute__((vector_size(16))) int d;
  //     __attribute__((vector_size(16))) int e;
  //     __attribute__((vector_size(16))) int f;
  //     __attribute__((vector_size(16))) int g;
  //     __attribute__((vector_size(16))) int h;
  //     __attribute__((vector_size(16))) int i;
  //   }
  //
  //   b. Structs with "multiple" fixed-length vector array, e.g.
  //   struct {
  //     __attribute__((vector_size(16))) int a[2];
  //     __attribute__((vector_size(16))) int b[2];
  //   }
  //
  //   c. Vector registers needed exceeds 8, e.g.
  //   struct {
  //     // Registers needed for single fixed-length element:
  //     // 64 * 8 / ABI_VLEN = 4
  //     __attribute__((vector_size(64))) int a;
  //     __attribute__((vector_size(64))) int b;
  //     __attribute__((vector_size(64))) int c;
  //     __attribute__((vector_size(64))) int d;
  //   }
  //
  // 1. Struct of 1 fixed-length vector is passed as a scalable vector.
  // 2. Struct of >1 fixed-length vectors are passed as vector tuple.
  // 3. Struct of an array with 1 element of fixed-length vectors is passed as a
  //    scalable vector.
  // 4. Struct of an array with >1 elements of fixed-length vectors is passed as
  //    vector tuple.
  // 5. Otherwise, pass the struct indirectly.
 
  llvm::StructType *STy = dyn_cast<llvm::StructType>(CGT.ConvertType(Ty));
  if (!STy)
    return nullptr;
 
  unsigned NumElts = STy->getStructNumElements();
  if (NumElts > 8)
    return nullptr;
 
  auto *FirstEltTy = STy->getElementType(0);
  if (!STy->containsHomogeneousTypes())
    return nullptr;
 
  if (auto *ArrayTy = dyn_cast<llvm::ArrayType>(FirstEltTy)) {
    // Only struct of single array is accepted
    if (NumElts != 1)
      return nullptr;
    FirstEltTy = ArrayTy->getArrayElementType();
    NumElts = ArrayTy->getNumElements();
  }
 
  auto *FixedVecTy = dyn_cast<llvm::FixedVectorType>(FirstEltTy);
  if (!FixedVecTy)
    return nullptr;
 
  // Check registers needed <= 8.
  if (NumElts * llvm::divideCeil(
                    FixedVecTy->getNumElements() *
                        FixedVecTy->getElementType()->getScalarSizeInBits(),
                    ABIVLen) >
      8)
    return nullptr;
 
  // Turn them into scalable vector type or vector tuple type if legal.
  if (NumElts == 1) {
    // Handle single fixed-length vector.
    llvm::FixedVectorType *VLSTy = getVLSCCCompatibleType(FixedVecTy);
    return llvm::ScalableVectorType::get(
        VLSTy->getElementType(),
        llvm::divideCeil(VLSTy->getNumElements() * llvm::RISCV::RVVBitsPerBlock,
                         ABIVLen));
  }
 
  // LMUL
  // = fixed-length vector size / ABIVLen
  // = 8 * I8EltCount / RVVBitsPerBlock
  // =>
  // I8EltCount
  // = (fixed-length vector size * RVVBitsPerBlock) / (ABIVLen * 8)
  unsigned I8EltCount =
      llvm::divideCeil(FixedVecTy->getNumElements() *
                           FixedVecTy->getElementType()->getScalarSizeInBits() *
                           llvm::RISCV::RVVBitsPerBlock,
                       ABIVLen * 8);
  return llvm::TargetExtType::get(
      getVMContext(), "riscv.vector.tuple",
      llvm::ScalableVectorType::get(llvm::Type::getInt8Ty(getVMContext()),
                                    I8EltCount),
      NumElts);
}
 
llvm::Type *
RISCVABIInfo::detectHomogeneousRVVFixedLengthStruct(QualType Ty) const {
  const auto *RT = Ty->getAsCanonical<RecordType>();
  if (!RT)
    return nullptr;
 
  const RecordDecl *RD = RT->getDecl()->getDefinitionOrSelf();
  if (RD->isUnion())
    return nullptr;
  if (getRecordArgABI(Ty, getCXXABI()))
    return nullptr;
 
  // Reject C++ types with base classes.
  if (const auto *CXXRD = dyn_cast<CXXRecordDecl>(RD))
    if (CXXRD->getNumBases() != 0)
      return nullptr;
 
  SmallVector<const FieldDecl *, 8> Fields(RD->fields());
 
  if (Fields.empty())
    return nullptr;
 
  auto IsFixedLengthRVVVector = [](const VectorType *VT) {
    return isRVVFixedLengthVectorKind(VT->getVectorKind());
  };
 
  QualType CommonTy;
  unsigned Count = 0;
 
  // Single array field: struct { fixed-length RVV T a[N]; }
  if (Fields.size() == 1) {
    QualType FieldTy = Fields[0]->getType().getCanonicalType();
    if (const ConstantArrayType *AT =
            getContext().getAsConstantArrayType(FieldTy)) {
      QualType EltTy = AT->getElementType().getCanonicalType();
      if (const auto *VT = EltTy->getAs<VectorType>();
          VT && IsFixedLengthRVVVector(VT)) {
        CommonTy = EltTy;
        Count = AT->getZExtSize();
      }
    }
  }
 
  // All fields are the same fixed-length RVV vector type (data or mask).
  if (CommonTy.isNull()) {
    if (Fields.size() > 8)
      return nullptr;
    for (const FieldDecl *FD : Fields) {
      QualType FieldTy = FD->getType().getCanonicalType();
      const auto *VT = FieldTy->getAs<VectorType>();
      if (!VT || !IsFixedLengthRVVVector(VT))
        return nullptr;
      if (CommonTy.isNull())
        CommonTy = FieldTy;
      else if (!getContext().hasSameType(CommonTy, FieldTy))
        return nullptr;
    }
    Count = Fields.size();
  }
 
  if (Count == 0 || Count > 8)
    return nullptr;
 
  const auto *VT = CommonTy->castAs<VectorType>();
  llvm::Type *EltType = CGT.ConvertType(VT->getElementType());
  unsigned MinElts = getRVVFixedLengthMinElts(VT);
 
  // Ensure total register usage does not exceed 8. Each vector occupies
  // max(1, LMUL) registers.
  uint64_t RegsPerVec =
      llvm::divideCeil((uint64_t)MinElts * EltType->getScalarSizeInBits(),
                       llvm::RISCV::RVVBitsPerBlock);
  if (Count > 1 && Count * RegsPerVec > 8)
    return nullptr;
 
  if (Count == 1)
    return llvm::ScalableVectorType::get(EltType, MinElts);
 
  unsigned I8EltCount =
      llvm::divideCeil((uint64_t)MinElts * EltType->getScalarSizeInBits(), 8);
  auto *I8Vec = llvm::ScalableVectorType::get(
      llvm::Type::getInt8Ty(getVMContext()), I8EltCount);
  return llvm::TargetExtType::get(getVMContext(), "riscv.vector.tuple", I8Vec,
                                  Count);
}
 
llvm::FixedVectorType *
RISCVABIInfo::getVLSCCCompatibleType(llvm::FixedVectorType *FixedVecTy) const {
  llvm::Type *EltType = FixedVecTy->getElementType();
  const TargetInfo &TI = getContext().getTargetInfo();
  if ((EltType->isHalfTy() && !TI.hasFeature("zvfhmin")) ||
      (EltType->isBFloatTy() &&
       !(TI.hasFeature("zvfbfmin") || TI.hasFeature("experimental-zvfbfa"))) ||
      (EltType->isFloatTy() && !TI.hasFeature("zve32f")) ||
      (EltType->isDoubleTy() && !TI.hasFeature("zve64d")) ||
      (EltType->isIntegerTy(64) && !TI.hasFeature("zve64x")) ||
      EltType->isIntegerTy(128))
    return llvm::FixedVectorType::get(llvm::Type::getInt8Ty(getVMContext()),
                                      FixedVecTy->getNumElements() *
                                          EltType->getScalarSizeInBits() / 8);
  return FixedVecTy;
}
 
// Fixed-length RVV vectors are represented as scalable vectors in function
// args/return and must be coerced from fixed vectors.
ABIArgInfo RISCVABIInfo::coerceVLSVector(QualType Ty, unsigned ABIVLen) const {
  assert(Ty->isVectorType() && "expected vector type!");
 
  const auto *VT = Ty->castAs<VectorType>();
  assert(VT->getElementType()->isBuiltinType() && "expected builtin type!");
 
  auto VScale = getContext().getTargetInfo().getVScaleRange(
      getContext().getLangOpts(), TargetInfo::ArmStreamingKind::NotStreaming);
 
  unsigned NumElts;
  llvm::Type *EltType;
  if (isRVVFixedLengthMaskVectorKind(VT->getVectorKind())) {
    NumElts = VT->getRVVFixedLengthMaskLanes();
    EltType = llvm::Type::getInt1Ty(getVMContext());
  } else {
    assert((VT->getVectorKind() == VectorKind::Generic ||
            isRVVFixedLengthDataVectorKind(VT->getVectorKind())) &&
           "Unexpected vector kind");
    NumElts = VT->getNumElements();
    EltType = CGT.ConvertType(VT->getElementType());
  }
 
  llvm::ScalableVectorType *ResType;
 
  if (ABIVLen == 0) {
    unsigned MinScalableElts;
    if (isRVVFixedLengthDataVectorKind(VT->getVectorKind())) {
      // Use the LMUL of the type the typedef was written against.
      MinScalableElts = getRVVFixedLengthDataMinElts(
          VT->getVectorKind(), EltType->getScalarSizeInBits());
    } else if (isRVVFixedLengthMaskVectorKind(VT->getVectorKind())) {
      // Use the ratio of the vbool<R>_t type the typedef was written against.
      MinScalableElts = getRVVFixedLengthMaskMinElts(VT->getVectorKind());
    } else {
      // The MinNumElts is simplified from equation:
      // NumElts / VScale =
      //  (EltSize * NumElts / (VScale * RVVBitsPerBlock))
      //    * (RVVBitsPerBlock / EltSize)
      MinScalableElts = NumElts / VScale->first;
    }
    ResType = llvm::ScalableVectorType::get(EltType, MinScalableElts);
  } else {
    // Check registers needed <= 8.
    if ((EltType->getScalarSizeInBits() * NumElts / ABIVLen) > 8)
      return getNaturalAlignIndirect(
          Ty, /*AddrSpace=*/getDataLayout().getAllocaAddrSpace(),
          /*ByVal=*/false);
 
    // Generic vector
    // The number of elements needs to be at least 1.
    llvm::FixedVectorType *VLSTy =
        getVLSCCCompatibleType(llvm::FixedVectorType::get(EltType, NumElts));
    ResType = llvm::ScalableVectorType::get(
        VLSTy->getElementType(),
        llvm::divideCeil(VLSTy->getNumElements() * llvm::RISCV::RVVBitsPerBlock,
                         ABIVLen));
  }
 
  return ABIArgInfo::getDirect(ResType);
}
 
ABIArgInfo RISCVABIInfo::classifyArgumentType(QualType Ty, bool IsFixed,
                                              int &ArgGPRsLeft,
                                              int &ArgFPRsLeft,
                                              unsigned ABIVLen) const {
  assert(ArgGPRsLeft <= NumArgGPRs && "Arg GPR tracking underflow");
  Ty = useFirstFieldIfTransparentUnion(Ty);
 
  // Structures with either a non-trivial destructor or a non-trivial
  // copy constructor are always passed indirectly.
  if (CGCXXABI::RecordArgABI RAA = getRecordArgABI(Ty, getCXXABI())) {
    if (ArgGPRsLeft)
      ArgGPRsLeft -= 1;
    return getNaturalAlignIndirect(
        Ty, /*AddrSpace=*/getDataLayout().getAllocaAddrSpace(),
        /*ByVal=*/RAA == CGCXXABI::RAA_DirectInMemory);
  }
 
  uint64_t Size = getContext().getTypeSize(Ty);
 
  // Ignore empty structs/unions whose size is zero. According to the calling
  // convention empty structs/unions are required to be sized types in C++.
  if (isEmptyRecord(getContext(), Ty, true) && Size == 0)
    return ABIArgInfo::getIgnore();
 
  // Pass floating point values via FPRs if possible.
  if (IsFixed && Ty->isFloatingType() && !Ty->isComplexType() &&
      FLen >= Size && ArgFPRsLeft) {
    ArgFPRsLeft--;
    return ABIArgInfo::getDirect();
  }
 
  // Complex types for the hard float ABI must be passed direct rather than
  // using CoerceAndExpand.
  if (IsFixed && Ty->isComplexType() && FLen && ArgFPRsLeft >= 2) {
    QualType EltTy = Ty->castAs<ComplexType>()->getElementType();
    if (getContext().getTypeSize(EltTy) <= FLen) {
      ArgFPRsLeft -= 2;
      return ABIArgInfo::getDirect();
    }
  }
 
  if (IsFixed && FLen && Ty->isStructureOrClassType()) {
    llvm::Type *Field1Ty = nullptr;
    llvm::Type *Field2Ty = nullptr;
    CharUnits Field1Off = CharUnits::Zero();
    CharUnits Field2Off = CharUnits::Zero();
    int NeededArgGPRs = 0;
    int NeededArgFPRs = 0;
    bool IsCandidate =
        detectFPCCEligibleStruct(Ty, Field1Ty, Field1Off, Field2Ty, Field2Off,
                                 NeededArgGPRs, NeededArgFPRs);
    if (IsCandidate && NeededArgGPRs <= ArgGPRsLeft &&
        NeededArgFPRs <= ArgFPRsLeft) {
      ArgGPRsLeft -= NeededArgGPRs;
      ArgFPRsLeft -= NeededArgFPRs;
      return coerceAndExpandFPCCEligibleStruct(Field1Ty, Field1Off, Field2Ty,
                                               Field2Off);
    }
  }
 
  if (IsFixed && Ty->isStructureOrClassType()) {
    if (llvm::Type *CoerceTy = detectHomogeneousRVVFixedLengthStruct(Ty))
      return ABIArgInfo::getTargetSpecific(CoerceTy);
 
    if (llvm::Type *VLSType = detectVLSCCEligibleStruct(Ty, ABIVLen))
      return ABIArgInfo::getTargetSpecific(VLSType);
  }
 
  uint64_t NeededAlign = getContext().getTypeAlign(Ty);
  // Determine the number of GPRs needed to pass the current argument
  // according to the ABI. 2*XLen-aligned varargs are passed in "aligned"
  // register pairs, so may consume 3 registers.
  // TODO: To be compatible with GCC's behaviors, we don't align registers
  // currently if we are using ILP32E calling convention. This behavior may be
  // changed when RV32E/ILP32E is ratified.
  int NeededArgGPRs = 1;
  if (!IsFixed && NeededAlign == 2 * XLen)
    NeededArgGPRs = 2 + (EABI ? 0 : (ArgGPRsLeft % 2));
  else if (Size > XLen && Size <= 2 * XLen)
    NeededArgGPRs = 2;
 
  if (NeededArgGPRs > ArgGPRsLeft) {
    NeededArgGPRs = ArgGPRsLeft;
  }
 
  ArgGPRsLeft -= NeededArgGPRs;
 
  if (!isAggregateTypeForABI(Ty) && !Ty->isVectorType()) {
    // Treat an enum type as its underlying type.
    if (const auto *ED = Ty->getAsEnumDecl())
      Ty = ED->getIntegerType();
 
    if (const auto *EIT = Ty->getAs<BitIntType>()) {
 
      if (XLen == 64 && EIT->getNumBits() == 32)
        return extendType(Ty, CGT.ConvertType(Ty));
 
      if (EIT->getNumBits() <= 2 * XLen)
        return ABIArgInfo::getExtend(Ty, CGT.ConvertType(Ty));
      return getNaturalAlignIndirect(
          Ty, /*AddrSpace=*/getDataLayout().getAllocaAddrSpace(),
          /*ByVal=*/false);
    }
 
    // All integral types are promoted to XLen width
    if (Size < XLen && Ty->isIntegralOrEnumerationType())
      return extendType(Ty, CGT.ConvertType(Ty));
 
    return ABIArgInfo::getDirect();
  }
 
  // TODO: _BitInt is not handled yet in VLS calling convention since _BitInt
  // ABI is also not merged yet in RISC-V:
  // https://github.com/riscv-non-isa/riscv-elf-psabi-doc/pull/419
  if (const VectorType *VT = Ty->getAs<VectorType>();
      VT && !VT->getElementType()->isBitIntType()) {
    if (isRVVFixedLengthVectorKind(VT->getVectorKind()))
      return coerceVLSVector(Ty);
    if (VT->getVectorKind() == VectorKind::Generic && ABIVLen != 0)
      // Generic vector without riscv_vls_cc should fall through and pass by
      // reference.
      return coerceVLSVector(Ty, ABIVLen);
  }
 
  // Aggregates which are <= 2*XLen will be passed in registers if possible,
  // so coerce to integers.
  if (Size <= 2 * XLen) {
    unsigned Alignment = getContext().getTypeAlign(Ty);
 
    if (Size <= XLen) {
      // Use the smallest integer type we can.
      return ABIArgInfo::getDirect(
          llvm::IntegerType::get(getVMContext(), Size));
    }
    // Use 2*XLen if 2*XLen alignment is required.
    if (Alignment == 2 * XLen)
      return ABIArgInfo::getDirect(
          llvm::IntegerType::get(getVMContext(), 2 * XLen));
    // Use 2-element XLen array if only XLen alignment is required.
    return ABIArgInfo::getDirect(
        llvm::ArrayType::get(llvm::IntegerType::get(getVMContext(), XLen), 2));
  }
  return getNaturalAlignIndirect(
      Ty, /*AddrSpace=*/getDataLayout().getAllocaAddrSpace(),
      /*ByVal=*/false);
}
 
ABIArgInfo RISCVABIInfo::classifyReturnType(QualType RetTy,
                                            unsigned ABIVLen) const {
  if (RetTy->isVoidType())
    return ABIArgInfo::getIgnore();
 
  int ArgGPRsLeft = 2;
  int ArgFPRsLeft = FLen ? 2 : 0;
 
  // The rules for return and argument types are the same, so defer to
  // classifyArgumentType.
  return classifyArgumentType(RetTy, /*IsFixed=*/true, ArgGPRsLeft, ArgFPRsLeft,
                              ABIVLen);
}
 
RValue RISCVABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
                               QualType Ty, AggValueSlot Slot) const {
  CharUnits SlotSize = CharUnits::fromQuantity(XLen / 8);
 
  // Empty records are ignored for parameter passing purposes.
  if (isEmptyRecord(getContext(), Ty, true))
    return Slot.asRValue();
 
  auto TInfo = getContext().getTypeInfoInChars(Ty);
 
  // TODO: To be compatible with GCC's behaviors, we force arguments with
  // 2×XLEN-bit alignment and size at most 2×XLEN bits like `long long`,
  // `unsigned long long`, `double` and `__int128` to have XLEN-byte alignment
  // for ILP32E and LP64E. This behavior may be changed when RV32E/ILP32E is
  // ratified.
  if (EABI)
    TInfo.Align = std::min(TInfo.Align, SlotSize);
 
  // Arguments bigger than 2*Xlen bytes are passed indirectly.
  bool IsIndirect = TInfo.Width > 2 * SlotSize;
 
  return emitVoidPtrVAArg(CGF, VAListAddr, Ty, IsIndirect, TInfo, SlotSize,
                          /*AllowHigherAlign=*/true, Slot);
}
 
ABIArgInfo RISCVABIInfo::extendType(QualType Ty, llvm::Type *CoerceTy) const {
  int TySize = getContext().getTypeSize(Ty);
  // RV64 ABI requires unsigned 32 bit integers to be sign extended.
  if (XLen == 64 && Ty->isUnsignedIntegerOrEnumerationType() && TySize == 32)
    return ABIArgInfo::getSignExtend(Ty, CoerceTy);
  return ABIArgInfo::getExtend(Ty, CoerceTy);
}
 
llvm::Value *RISCVABIInfo::createCoercedLoad(Address Src, const ABIArgInfo &AI,
                                             CodeGenFunction &CGF) const {
  llvm::Type *Ty = AI.getCoerceToType();
  llvm::Type *SrcTy = Src.getElementType();
  llvm::StructType *SrcSTy = cast<llvm::StructType>(SrcTy);
  assert((Ty->isScalableTy() || Ty->isTargetExtTy()) &&
         "Only scalable vector type and vector tuple type are allowed for load "
         "type.");
  if (llvm::TargetExtType *TupTy = dyn_cast<llvm::TargetExtType>(Ty)) {
    // In RISC-V VLS calling convention, struct of fixed vectors or struct of
    // array of fixed vector of length >1 might be lowered using vector tuple
    // type, we consider it as a valid load, e.g.
    // struct i32x4x2 {
    //     __attribute__((vector_size(16))) int i;
    //     __attribute__((vector_size(16))) int i;
    // };
    // or
    // struct i32x4 {
    //     __attribute__((vector_size(16))) int i[2];
    // };
    // is lowered to target("riscv.vector.tuple", <vscale x 8 x i8>, 2)
    // when ABI_VLEN = 128 bits, please checkout
    // clang/test/CodeGen/RISCV/riscv-vector-callingconv-llvm-ir.c
    // for more information.
    assert(TupTy->getName() == "riscv.vector.tuple");
    llvm::Type *EltTy = TupTy->getTypeParameter(0);
    unsigned NumElts = TupTy->getIntParameter(0);
 
    if (auto *ArrayTy = dyn_cast<llvm::ArrayType>(SrcSTy->getElementType(0)))
      Src = Src.withElementType(ArrayTy);
 
    // Perform extract element and load
    llvm::Value *TupleVal = llvm::PoisonValue::get(Ty);
    auto *Load = CGF.Builder.CreateLoad(Src);
    for (unsigned i = 0; i < NumElts; ++i) {
      // Extract from struct
      llvm::Value *ExtractFromLoad = CGF.Builder.CreateExtractValue(Load, i);
      auto *FixedVecTy =
          cast<llvm::FixedVectorType>(ExtractFromLoad->getType());
      llvm::FixedVectorType *VLSTy = getVLSCCCompatibleType(FixedVecTy);
      if (VLSTy != FixedVecTy)
        ExtractFromLoad = CGF.Builder.CreateBitCast(ExtractFromLoad, VLSTy);
      // Element in vector tuple type is always i8, so we need to cast back to
      // it's original element type.
      EltTy =
          cast<llvm::ScalableVectorType>(llvm::VectorType::getWithSizeAndScalar(
              cast<llvm::VectorType>(EltTy), VLSTy));
      llvm::Value *VectorVal = llvm::PoisonValue::get(EltTy);
      // Insert to scalable vector
      VectorVal = CGF.Builder.CreateInsertVector(
          EltTy, VectorVal, ExtractFromLoad, uint64_t(0), "cast.scalable");
      // Insert scalable vector to vector tuple
      llvm::Value *Idx = CGF.Builder.getInt32(i);
      TupleVal =
          CGF.Builder.CreateIntrinsic(llvm::Intrinsic::riscv_tuple_insert,
                                      {Ty, EltTy}, {TupleVal, VectorVal, Idx});
    }
    return TupleVal;
  }
 
  // In RISC-V VLS calling convention, struct of fixed vector or struct of
  // fixed vector array of length 1 might be lowered using scalable vector,
  // we consider it as a valid load, e.g.
  // struct i32x4 {
  //     __attribute__((vector_size(16))) int i;
  // };
  // or
  // struct i32x4 {
  //     __attribute__((vector_size(16))) int i[1];
  // };
  // is lowered to <vscale x 2 x i32>
  // when ABI_VLEN = 128 bits, please checkout
  // clang/test/CodeGen/RISCV/riscv-vector-callingconv-llvm-ir.c
  // for more information.
  auto *ScalableDstTy = cast<llvm::ScalableVectorType>(Ty);
  SrcTy = SrcSTy->getElementType(0);
  if (auto *ArrayTy = dyn_cast<llvm::ArrayType>(SrcTy))
    SrcTy = ArrayTy->getElementType();
  Src = Src.withElementType(SrcTy);
  auto *FixedSrcTy = cast<llvm::FixedVectorType>(SrcTy);
  llvm::Value *Load = CGF.Builder.CreateLoad(Src);
  llvm::FixedVectorType *VLSTy = getVLSCCCompatibleType(FixedSrcTy);
  if (VLSTy != FixedSrcTy)
    Load = CGF.Builder.CreateBitCast(Load, VLSTy);
  auto *VectorVal = llvm::PoisonValue::get(ScalableDstTy);
  llvm::Value *Result = CGF.Builder.CreateInsertVector(
      ScalableDstTy, VectorVal, Load, uint64_t(0), "cast.scalable");
  return Result;
}
 
void RISCVABIInfo::createCoercedStore(llvm::Value *Val, Address Dst,
                                      const ABIArgInfo &AI, bool DestIsVolatile,
                                      CodeGenFunction &CGF) const {
  llvm::Type *SrcTy = Val->getType();
  llvm::StructType *DstSTy = cast<llvm::StructType>(Dst.getElementType());
  assert((SrcTy->isScalableTy() || SrcTy->isTargetExtTy()) &&
         "Only scalable vector type and vector tuple type are allowed for "
         "store value.");
  if (llvm::TargetExtType *TupTy = dyn_cast<llvm::TargetExtType>(SrcTy)) {
    // In RISC-V VLS calling convention, struct of fixed vectors or struct
    // of array of fixed vector of length >1 might be lowered using vector
    // tuple type, we consider it as a valid load, e.g.
    // struct i32x4x2 {
    //     __attribute__((vector_size(16))) int i;
    //     __attribute__((vector_size(16))) int i;
    // };
    // or
    // struct i32x4 {
    //     __attribute__((vector_size(16))) int i[2];
    // };
    // is lowered to target("riscv.vector.tuple", <vscale x 8 x i8>, 2)
    // when ABI_VLEN = 128 bits, please checkout
    // clang/test/CodeGen/RISCV/riscv-vector-callingconv-llvm-ir.c
    // for more information.
    assert(TupTy->getName() == "riscv.vector.tuple");
    llvm::Type *EltTy = TupTy->getTypeParameter(0);
    unsigned NumElts = TupTy->getIntParameter(0);
 
    llvm::Type *FixedVecTy = DstSTy->getElementType(0);
    if (auto *ArrayTy = dyn_cast<llvm::ArrayType>(DstSTy->getElementType(0))) {
      Dst = Dst.withElementType(ArrayTy);
      FixedVecTy = ArrayTy->getArrayElementType();
    }
 
    llvm::FixedVectorType *VLSTy =
        getVLSCCCompatibleType(cast<llvm::FixedVectorType>(FixedVecTy));
 
    // Perform extract element and store
    for (unsigned i = 0; i < NumElts; ++i) {
      // Element in vector tuple type is always i8, so we need to cast back
      // to it's original element type.
      EltTy =
          cast<llvm::ScalableVectorType>(llvm::VectorType::getWithSizeAndScalar(
              cast<llvm::VectorType>(EltTy), VLSTy));
      // Extract scalable vector from tuple
      llvm::Value *Idx = CGF.Builder.getInt32(i);
      auto *TupleElement = CGF.Builder.CreateIntrinsic(
          llvm::Intrinsic::riscv_tuple_extract, {EltTy, TupTy}, {Val, Idx});
 
      // Extract fixed vector from scalable vector
      llvm::Value *ExtractVec =
          CGF.Builder.CreateExtractVector(VLSTy, TupleElement, uint64_t(0));
      if (VLSTy != FixedVecTy)
        ExtractVec = CGF.Builder.CreateBitCast(ExtractVec, FixedVecTy);
      // Store fixed vector to corresponding address
      Address EltPtr = Address::invalid();
      if (Dst.getElementType()->isStructTy())
        EltPtr = CGF.Builder.CreateStructGEP(Dst, i);
      else
        EltPtr = CGF.Builder.CreateConstArrayGEP(Dst, i);
      auto *I = CGF.Builder.CreateStore(ExtractVec, EltPtr, DestIsVolatile);
      CGF.addInstToCurrentSourceAtom(I, ExtractVec);
    }
    return;
  }
 
  // In RISC-V VLS calling convention, struct of fixed vector or struct of
  // fixed vector array of length 1 might be lowered using scalable
  // vector, we consider it as a valid load, e.g.
  // struct i32x4 {
  //     __attribute__((vector_size(16))) int i;
  // };
  // or
  // struct i32x4 {
  //     __attribute__((vector_size(16))) int i[1];
  // };
  // is lowered to <vscale x 2 x i32>
  // when ABI_VLEN = 128 bits, please checkout
  // clang/test/CodeGen/RISCV/riscv-vector-callingconv-llvm-ir.c
  // for more information.
  llvm::Type *EltTy = DstSTy->getElementType(0);
  if (auto *ArrayTy = dyn_cast<llvm::ArrayType>(EltTy)) {
    assert(ArrayTy->getNumElements() == 1);
    EltTy = ArrayTy->getElementType();
  }
  auto *FixedVecTy = cast<llvm::FixedVectorType>(EltTy);
  llvm::FixedVectorType *VLSTy = getVLSCCCompatibleType(FixedVecTy);
  llvm::Value *Coerced =
      CGF.Builder.CreateExtractVector(VLSTy, Val, uint64_t(0));
  if (VLSTy != FixedVecTy)
    Coerced = CGF.Builder.CreateBitCast(Coerced, FixedVecTy);
  auto *I = CGF.Builder.CreateStore(Coerced, Dst, DestIsVolatile);
  CGF.addInstToCurrentSourceAtom(I, Val);
}
 
namespace {
class RISCVTargetCodeGenInfo : public TargetCodeGenInfo {
public:
  RISCVTargetCodeGenInfo(CodeGen::CodeGenTypes &CGT, unsigned XLen,
                         unsigned FLen, bool EABI)
      : TargetCodeGenInfo(
            std::make_unique<RISCVABIInfo>(CGT, XLen, FLen, EABI)) {
    SwiftInfo =
        std::make_unique<SwiftABIInfo>(CGT, /*SwiftErrorInRegister=*/false);
  }
 
  void setTargetAttributes(const Decl *D, llvm::GlobalValue *GV,
                           CodeGen::CodeGenModule &CGM) const override {
    const auto *FD = dyn_cast_or_null<FunctionDecl>(D);
    if (!FD) return;
 
    auto *Fn = cast<llvm::Function>(GV);
 
    if (CGM.getCodeGenOpts().CFProtectionReturn)
      Fn->addFnAttr("hw-shadow-stack");
 
    const auto *Attr = FD->getAttr<RISCVInterruptAttr>();
    if (!Attr)
      return;
 
    StringRef Kind = "machine";
    bool HasSiFiveCLICPreemptible = false;
    bool HasSiFiveCLICStackSwap = false;
    for (RISCVInterruptAttr::InterruptType type : Attr->interrupt()) {
      switch (type) {
      case RISCVInterruptAttr::machine:
        // Do not update `Kind` because `Kind` is already "machine", or the
        // kinds also contains SiFive types which need to be applied.
        break;
      case RISCVInterruptAttr::supervisor:
        Kind = "supervisor";
        break;
      case RISCVInterruptAttr::rnmi:
        Kind = "rnmi";
        break;
      case RISCVInterruptAttr::qcinest:
        Kind = "qci-nest";
        break;
      case RISCVInterruptAttr::qcinonest:
        Kind = "qci-nonest";
        break;
      // There are three different LLVM IR attribute values for SiFive CLIC
      // interrupt kinds, one for each kind and one extra for their combination.
      case RISCVInterruptAttr::SiFiveCLICPreemptible: {
        HasSiFiveCLICPreemptible = true;
        Kind = HasSiFiveCLICStackSwap ? "SiFive-CLIC-preemptible-stack-swap"
                                      : "SiFive-CLIC-preemptible";
        break;
      }
      case RISCVInterruptAttr::SiFiveCLICStackSwap: {
        HasSiFiveCLICStackSwap = true;
        Kind = HasSiFiveCLICPreemptible ? "SiFive-CLIC-preemptible-stack-swap"
                                        : "SiFive-CLIC-stack-swap";
        break;
      }
      }
    }
 
    Fn->addFnAttr("interrupt", Kind);
  }
};
} // namespace
 
std::unique_ptr<TargetCodeGenInfo>
CodeGen::createRISCVTargetCodeGenInfo(CodeGenModule &CGM, unsigned XLen,
                                      unsigned FLen, bool EABI) {
  return std::make_unique<RISCVTargetCodeGenInfo>(CGM.getTypes(), XLen, FLen,
                                                  EABI);
}