AMDGPU.cpp

Go to the documentation of this file.

//===- AMDGPU.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 "clang/Basic/TargetOptions.h"
#include "llvm/Support/AMDGPUAddrSpace.h"
 
using namespace clang;
using namespace clang::CodeGen;
 
//===----------------------------------------------------------------------===//
// AMDGPU ABI Implementation
//===----------------------------------------------------------------------===//
 
namespace {
 
class AMDGPUABIInfo final : public DefaultABIInfo {
private:
  static const unsigned MaxNumRegsForArgsRet = 16;
 
  unsigned numRegsForType(QualType Ty) const;
 
  bool isHomogeneousAggregateBaseType(QualType Ty) const override;
  bool isHomogeneousAggregateSmallEnough(const Type *Base,
                                         uint64_t Members) const override;
 
  // Coerce HIP scalar pointer arguments from generic pointers to global ones.
  llvm::Type *coerceKernelArgumentType(llvm::Type *Ty, unsigned FromAS,
                                       unsigned ToAS) const {
    // Single value types.
    auto *PtrTy = llvm::dyn_cast<llvm::PointerType>(Ty);
    if (PtrTy && PtrTy->getAddressSpace() == FromAS)
      return llvm::PointerType::get(Ty->getContext(), ToAS);
    return Ty;
  }
 
public:
  explicit AMDGPUABIInfo(CodeGen::CodeGenTypes &CGT) :
    DefaultABIInfo(CGT) {}
 
  ABIArgInfo classifyReturnType(QualType RetTy) const;
  ABIArgInfo classifyKernelArgumentType(QualType Ty) const;
  ABIArgInfo classifyArgumentType(QualType Ty, bool Variadic,
                                  unsigned &NumRegsLeft) const;
 
  void computeInfo(CGFunctionInfo &FI) const override;
  RValue EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, QualType Ty,
                   AggValueSlot Slot) const override;
 
  llvm::FixedVectorType *
  getOptimalVectorMemoryType(llvm::FixedVectorType *T,
                             const LangOptions &Opt) const override {
    // We have legal instructions for 96-bit so 3x32 can be supported.
    // FIXME: This check should be a subtarget feature as technically SI doesn't
    // support it.
    if (T->getNumElements() == 3 && getDataLayout().getTypeSizeInBits(T) == 96)
      return T;
    return DefaultABIInfo::getOptimalVectorMemoryType(T, Opt);
  }
};
 
bool AMDGPUABIInfo::isHomogeneousAggregateBaseType(QualType Ty) const {
  return true;
}
 
bool AMDGPUABIInfo::isHomogeneousAggregateSmallEnough(
  const Type *Base, uint64_t Members) const {
  uint32_t NumRegs = (getContext().getTypeSize(Base) + 31) / 32;
 
  // Homogeneous Aggregates may occupy at most 16 registers.
  return Members * NumRegs <= MaxNumRegsForArgsRet;
}
 
/// Estimate number of registers the type will use when passed in registers.
unsigned AMDGPUABIInfo::numRegsForType(QualType Ty) const {
  unsigned NumRegs = 0;
 
  if (const VectorType *VT = Ty->getAs<VectorType>()) {
    // Compute from the number of elements. The reported size is based on the
    // in-memory size, which includes the padding 4th element for 3-vectors.
    QualType EltTy = VT->getElementType();
    unsigned EltSize = getContext().getTypeSize(EltTy);
 
    // 16-bit element vectors should be passed as packed.
    if (EltSize == 16)
      return (VT->getNumElements() + 1) / 2;
 
    unsigned EltNumRegs = (EltSize + 31) / 32;
    return EltNumRegs * VT->getNumElements();
  }
 
  if (const RecordType *RT = Ty->getAs<RecordType>()) {
    const RecordDecl *RD = RT->getDecl();
    assert(!RD->hasFlexibleArrayMember());
 
    for (const FieldDecl *Field : RD->fields()) {
      QualType FieldTy = Field->getType();
      NumRegs += numRegsForType(FieldTy);
    }
 
    return NumRegs;
  }
 
  return (getContext().getTypeSize(Ty) + 31) / 32;
}
 
void AMDGPUABIInfo::computeInfo(CGFunctionInfo &FI) const {
  llvm::CallingConv::ID CC = FI.getCallingConvention();
 
  if (!getCXXABI().classifyReturnType(FI))
    FI.getReturnInfo() = classifyReturnType(FI.getReturnType());
 
  unsigned ArgumentIndex = 0;
  const unsigned numFixedArguments = FI.getNumRequiredArgs();
 
  unsigned NumRegsLeft = MaxNumRegsForArgsRet;
  for (auto &Arg : FI.arguments()) {
    if (CC == llvm::CallingConv::AMDGPU_KERNEL) {
      Arg.info = classifyKernelArgumentType(Arg.type);
    } else {
      bool FixedArgument = ArgumentIndex++ < numFixedArguments;
      Arg.info = classifyArgumentType(Arg.type, !FixedArgument, NumRegsLeft);
    }
  }
}
 
RValue AMDGPUABIInfo::EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
                                QualType Ty, AggValueSlot Slot) const {
  const bool IsIndirect = false;
  const bool AllowHigherAlign = false;
  return emitVoidPtrVAArg(CGF, VAListAddr, Ty, IsIndirect,
                          getContext().getTypeInfoInChars(Ty),
                          CharUnits::fromQuantity(4), AllowHigherAlign, Slot);
}
 
ABIArgInfo AMDGPUABIInfo::classifyReturnType(QualType RetTy) const {
  if (isAggregateTypeForABI(RetTy)) {
    // Records with non-trivial destructors/copy-constructors should not be
    // returned by value.
    if (!getRecordArgABI(RetTy, getCXXABI())) {
      // Ignore empty structs/unions.
      if (isEmptyRecord(getContext(), RetTy, true))
        return ABIArgInfo::getIgnore();
 
      // Lower single-element structs to just return a regular value.
      if (const Type *SeltTy = isSingleElementStruct(RetTy, getContext()))
        return ABIArgInfo::getDirect(CGT.ConvertType(QualType(SeltTy, 0)));
 
      if (const RecordType *RT = RetTy->getAs<RecordType>()) {
        const RecordDecl *RD = RT->getDecl();
        if (RD->hasFlexibleArrayMember())
          return DefaultABIInfo::classifyReturnType(RetTy);
      }
 
      // Pack aggregates <= 4 bytes into single VGPR or pair.
      uint64_t Size = getContext().getTypeSize(RetTy);
      if (Size <= 16)
        return ABIArgInfo::getDirect(llvm::Type::getInt16Ty(getVMContext()));
 
      if (Size <= 32)
        return ABIArgInfo::getDirect(llvm::Type::getInt32Ty(getVMContext()));
 
      if (Size <= 64) {
        llvm::Type *I32Ty = llvm::Type::getInt32Ty(getVMContext());
        return ABIArgInfo::getDirect(llvm::ArrayType::get(I32Ty, 2));
      }
 
      if (numRegsForType(RetTy) <= MaxNumRegsForArgsRet)
        return ABIArgInfo::getDirect();
    }
  }
 
  // Otherwise just do the default thing.
  return DefaultABIInfo::classifyReturnType(RetTy);
}
 
/// For kernels all parameters are really passed in a special buffer. It doesn't
/// make sense to pass anything byval, so everything must be direct.
ABIArgInfo AMDGPUABIInfo::classifyKernelArgumentType(QualType Ty) const {
  Ty = useFirstFieldIfTransparentUnion(Ty);
 
  // TODO: Can we omit empty structs?
 
  if (const Type *SeltTy = isSingleElementStruct(Ty, getContext()))
    Ty = QualType(SeltTy, 0);
 
  llvm::Type *OrigLTy = CGT.ConvertType(Ty);
  llvm::Type *LTy = OrigLTy;
  if (getContext().getLangOpts().HIP) {
    LTy = coerceKernelArgumentType(
        OrigLTy, /*FromAS=*/getContext().getTargetAddressSpace(LangAS::Default),
        /*ToAS=*/getContext().getTargetAddressSpace(LangAS::cuda_device));
  }
 
  // FIXME: Should also use this for OpenCL, but it requires addressing the
  // problem of kernels being called.
  //
  // FIXME: This doesn't apply the optimization of coercing pointers in structs
  // to global address space when using byref. This would require implementing a
  // new kind of coercion of the in-memory type when for indirect arguments.
  if (!getContext().getLangOpts().OpenCL && LTy == OrigLTy &&
      isAggregateTypeForABI(Ty)) {
    return ABIArgInfo::getIndirectAliased(
        getContext().getTypeAlignInChars(Ty),
        getContext().getTargetAddressSpace(LangAS::opencl_constant),
        false /*Realign*/, nullptr /*Padding*/);
  }
 
  // If we set CanBeFlattened to true, CodeGen will expand the struct to its
  // individual elements, which confuses the Clover OpenCL backend; therefore we
  // have to set it to false here. Other args of getDirect() are just defaults.
  return ABIArgInfo::getDirect(LTy, 0, nullptr, false);
}
 
ABIArgInfo AMDGPUABIInfo::classifyArgumentType(QualType Ty, bool Variadic,
                                               unsigned &NumRegsLeft) const {
  assert(NumRegsLeft <= MaxNumRegsForArgsRet && "register estimate underflow");
 
  Ty = useFirstFieldIfTransparentUnion(Ty);
 
  if (Variadic) {
    return ABIArgInfo::getDirect(/*T=*/nullptr,
                                 /*Offset=*/0,
                                 /*Padding=*/nullptr,
                                 /*CanBeFlattened=*/false,
                                 /*Align=*/0);
  }
 
  if (isAggregateTypeForABI(Ty)) {
    // Records with non-trivial destructors/copy-constructors should not be
    // passed by value.
    if (auto RAA = getRecordArgABI(Ty, getCXXABI()))
      return getNaturalAlignIndirect(Ty, RAA == CGCXXABI::RAA_DirectInMemory);
 
    // Ignore empty structs/unions.
    if (isEmptyRecord(getContext(), Ty, true))
      return ABIArgInfo::getIgnore();
 
    // Lower single-element structs to just pass a regular value. TODO: We
    // could do reasonable-size multiple-element structs too, using getExpand(),
    // though watch out for things like bitfields.
    if (const Type *SeltTy = isSingleElementStruct(Ty, getContext()))
      return ABIArgInfo::getDirect(CGT.ConvertType(QualType(SeltTy, 0)));
 
    if (const RecordType *RT = Ty->getAs<RecordType>()) {
      const RecordDecl *RD = RT->getDecl();
      if (RD->hasFlexibleArrayMember())
        return DefaultABIInfo::classifyArgumentType(Ty);
    }
 
    // Pack aggregates <= 8 bytes into single VGPR or pair.
    uint64_t Size = getContext().getTypeSize(Ty);
    if (Size <= 64) {
      unsigned NumRegs = (Size + 31) / 32;
      NumRegsLeft -= std::min(NumRegsLeft, NumRegs);
 
      if (Size <= 16)
        return ABIArgInfo::getDirect(llvm::Type::getInt16Ty(getVMContext()));
 
      if (Size <= 32)
        return ABIArgInfo::getDirect(llvm::Type::getInt32Ty(getVMContext()));
 
      // XXX: Should this be i64 instead, and should the limit increase?
      llvm::Type *I32Ty = llvm::Type::getInt32Ty(getVMContext());
      return ABIArgInfo::getDirect(llvm::ArrayType::get(I32Ty, 2));
    }
 
    if (NumRegsLeft > 0) {
      unsigned NumRegs = numRegsForType(Ty);
      if (NumRegsLeft >= NumRegs) {
        NumRegsLeft -= NumRegs;
        return ABIArgInfo::getDirect();
      }
    }
 
    // Use pass-by-reference in stead of pass-by-value for struct arguments in
    // function ABI.
    return ABIArgInfo::getIndirectAliased(
        getContext().getTypeAlignInChars(Ty),
        getContext().getTargetAddressSpace(LangAS::opencl_private));
  }
 
  // Otherwise just do the default thing.
  ABIArgInfo ArgInfo = DefaultABIInfo::classifyArgumentType(Ty);
  if (!ArgInfo.isIndirect()) {
    unsigned NumRegs = numRegsForType(Ty);
    NumRegsLeft -= std::min(NumRegs, NumRegsLeft);
  }
 
  return ArgInfo;
}
 
class AMDGPUTargetCodeGenInfo : public TargetCodeGenInfo {
public:
  AMDGPUTargetCodeGenInfo(CodeGenTypes &CGT)
      : TargetCodeGenInfo(std::make_unique<AMDGPUABIInfo>(CGT)) {}
 
  void setFunctionDeclAttributes(const FunctionDecl *FD, llvm::Function *F,
                                 CodeGenModule &CGM) const;
 
  void emitTargetGlobals(CodeGen::CodeGenModule &CGM) const override;
 
  void setTargetAttributes(const Decl *D, llvm::GlobalValue *GV,
                           CodeGen::CodeGenModule &M) const override;
  unsigned getOpenCLKernelCallingConv() const override;
 
  llvm::Constant *getNullPointer(const CodeGen::CodeGenModule &CGM,
      llvm::PointerType *T, QualType QT) const override;
 
  LangAS getASTAllocaAddressSpace() const override {
    return getLangASFromTargetAS(
        getABIInfo().getDataLayout().getAllocaAddrSpace());
  }
  LangAS getGlobalVarAddressSpace(CodeGenModule &CGM,
                                  const VarDecl *D) const override;
  llvm::SyncScope::ID getLLVMSyncScopeID(const LangOptions &LangOpts,
                                         SyncScope Scope,
                                         llvm::AtomicOrdering Ordering,
                                         llvm::LLVMContext &Ctx) const override;
  void setTargetAtomicMetadata(CodeGenFunction &CGF,
                               llvm::Instruction &AtomicInst,
                               const AtomicExpr *Expr = nullptr) const override;
  llvm::Value *createEnqueuedBlockKernel(CodeGenFunction &CGF,
                                         llvm::Function *BlockInvokeFunc,
                                         llvm::Type *BlockTy) const override;
  bool shouldEmitStaticExternCAliases() const override;
  bool shouldEmitDWARFBitFieldSeparators() const override;
  void setCUDAKernelCallingConvention(const FunctionType *&FT) const override;
};
}
 
static bool requiresAMDGPUProtectedVisibility(const Decl *D,
                                              llvm::GlobalValue *GV) {
  if (GV->getVisibility() != llvm::GlobalValue::HiddenVisibility)
    return false;
 
  return !D->hasAttr<OMPDeclareTargetDeclAttr>() &&
         (D->hasAttr<OpenCLKernelAttr>() ||
          (isa<FunctionDecl>(D) && D->hasAttr<CUDAGlobalAttr>()) ||
          (isa<VarDecl>(D) &&
           (D->hasAttr<CUDADeviceAttr>() || D->hasAttr<CUDAConstantAttr>() ||
            cast<VarDecl>(D)->getType()->isCUDADeviceBuiltinSurfaceType() ||
            cast<VarDecl>(D)->getType()->isCUDADeviceBuiltinTextureType())));
}
 
void AMDGPUTargetCodeGenInfo::setFunctionDeclAttributes(
    const FunctionDecl *FD, llvm::Function *F, CodeGenModule &M) const {
  const auto *ReqdWGS =
      M.getLangOpts().OpenCL ? FD->getAttr<ReqdWorkGroupSizeAttr>() : nullptr;
  const bool IsOpenCLKernel =
      M.getLangOpts().OpenCL && FD->hasAttr<OpenCLKernelAttr>();
  const bool IsHIPKernel = M.getLangOpts().HIP && FD->hasAttr<CUDAGlobalAttr>();
 
  const auto *FlatWGS = FD->getAttr<AMDGPUFlatWorkGroupSizeAttr>();
  if (ReqdWGS || FlatWGS) {
    M.handleAMDGPUFlatWorkGroupSizeAttr(F, FlatWGS, ReqdWGS);
  } else if (IsOpenCLKernel || IsHIPKernel) {
    // By default, restrict the maximum size to a value specified by
    // --gpu-max-threads-per-block=n or its default value for HIP.
    const unsigned OpenCLDefaultMaxWorkGroupSize = 256;
    const unsigned DefaultMaxWorkGroupSize =
        IsOpenCLKernel ? OpenCLDefaultMaxWorkGroupSize
                       : M.getLangOpts().GPUMaxThreadsPerBlock;
    std::string AttrVal =
        std::string("1,") + llvm::utostr(DefaultMaxWorkGroupSize);
    F->addFnAttr("amdgpu-flat-work-group-size", AttrVal);
  }
 
  if (const auto *Attr = FD->getAttr<AMDGPUWavesPerEUAttr>())
    M.handleAMDGPUWavesPerEUAttr(F, Attr);
 
  if (const auto *Attr = FD->getAttr<AMDGPUNumSGPRAttr>()) {
    unsigned NumSGPR = Attr->getNumSGPR();
 
    if (NumSGPR != 0)
      F->addFnAttr("amdgpu-num-sgpr", llvm::utostr(NumSGPR));
  }
 
  if (const auto *Attr = FD->getAttr<AMDGPUNumVGPRAttr>()) {
    uint32_t NumVGPR = Attr->getNumVGPR();
 
    if (NumVGPR != 0)
      F->addFnAttr("amdgpu-num-vgpr", llvm::utostr(NumVGPR));
  }
 
  if (const auto *Attr = FD->getAttr<AMDGPUMaxNumWorkGroupsAttr>()) {
    uint32_t X = Attr->getMaxNumWorkGroupsX()
                     ->EvaluateKnownConstInt(M.getContext())
                     .getExtValue();
    // Y and Z dimensions default to 1 if not specified
    uint32_t Y = Attr->getMaxNumWorkGroupsY()
                     ? Attr->getMaxNumWorkGroupsY()
                           ->EvaluateKnownConstInt(M.getContext())
                           .getExtValue()
                     : 1;
    uint32_t Z = Attr->getMaxNumWorkGroupsZ()
                     ? Attr->getMaxNumWorkGroupsZ()
                           ->EvaluateKnownConstInt(M.getContext())
                           .getExtValue()
                     : 1;
 
    llvm::SmallString<32> AttrVal;