NVPTX.h

Go to the documentation of this file.

//===--- NVPTX.h - Declare NVPTX target feature support ---------*- C++ -*-===//
//
// 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 declares NVPTX TargetInfo objects.
//
//===----------------------------------------------------------------------===//
 
#ifndef LLVM_CLANG_LIB_BASIC_TARGETS_NVPTX_H
#define LLVM_CLANG_LIB_BASIC_TARGETS_NVPTX_H
 
#include "clang/Basic/Cuda.h"
#include "clang/Basic/TargetInfo.h"
#include "clang/Basic/TargetOptions.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/NVPTXAddrSpace.h"
#include "llvm/TargetParser/Triple.h"
#include <optional>
 
namespace clang {
namespace targets {
 
static constexpr LangASMap NVPTXAddrSpaceMap = {
    {LangAS::opencl_global, 1},
    {LangAS::opencl_local, 3},
    {LangAS::opencl_constant, 4},
    // FIXME: generic has to be added to the target
    {LangAS::opencl_generic, 0},
    {LangAS::opencl_global_device, 1},
    {LangAS::opencl_global_host, 1},
    {LangAS::cuda_device, 1},
    {LangAS::cuda_constant, 4},
    {LangAS::cuda_shared, 3},
    {LangAS::sycl_global, 1},
    {LangAS::sycl_global_device, 1},
    {LangAS::sycl_global_host, 1},
    {LangAS::sycl_local, 3},
};
 
/// The DWARF address class. Taken from
/// https://docs.nvidia.com/cuda/archive/10.0/ptx-writers-guide-to-interoperability/index.html#cuda-specific-dwarf
static const int NVPTXDWARFAddrSpaceMap[] = {
    -1, // Default, opencl_private or opencl_generic - not defined
    5,  // opencl_global
    -1,
    8,  // opencl_local or cuda_shared
    4,  // opencl_constant or cuda_constant
};
 
class LLVM_LIBRARY_VISIBILITY NVPTXTargetInfo : public TargetInfo {
  static const char *const GCCRegNames[];
  OffloadArch GPU;
  uint32_t PTXVersion;
  std::unique_ptr<TargetInfo> HostTarget;
 
public:
  NVPTXTargetInfo(const llvm::Triple &Triple, const TargetOptions &Opts,
                  unsigned TargetPointerWidth);
 
  void getTargetDefines(const LangOptions &Opts,
                        MacroBuilder &Builder) const override;
 
  llvm::SmallVector<Builtin::InfosShard> getTargetBuiltins() const override;
 
  bool isCLZForZeroUndef() const override { return false; }
 
  bool
  initFeatureMap(llvm::StringMap<bool> &Features, DiagnosticsEngine &Diags,
                 StringRef CPU,
                 const std::vector<std::string> &FeaturesVec) const override {
    if (GPU != OffloadArch::Unused)
      Features[OffloadArchToString(GPU)] = true;
    // Only add PTX feature if explicitly requested. Otherwise, let the backend
    // use the minimum required PTX version for the target SM.
    if (PTXVersion != 0)
      Features["ptx" + std::to_string(PTXVersion)] = true;
    return TargetInfo::initFeatureMap(Features, Diags, CPU, FeaturesVec);
  }
 
  bool hasFeature(StringRef Feature) const override;
 
  virtual bool isAddressSpaceSupersetOf(LangAS A, LangAS B) const override {
    // The generic address space AS(0) is a superset of all the other address
    // spaces used by the backend target.
    return A == B ||
           ((A == LangAS::Default ||
             (isTargetAddressSpace(A) &&
              toTargetAddressSpace(A) ==
                  llvm::NVPTXAS::ADDRESS_SPACE_GENERIC)) &&
            isTargetAddressSpace(B) &&
            toTargetAddressSpace(B) >= llvm::NVPTXAS::ADDRESS_SPACE_GENERIC &&
            toTargetAddressSpace(B) <= llvm::NVPTXAS::ADDRESS_SPACE_LOCAL &&
            toTargetAddressSpace(B) != 2);
  }
 
  ArrayRef<const char *> getGCCRegNames() const override;
 
  ArrayRef<TargetInfo::GCCRegAlias> getGCCRegAliases() const override {
    // No aliases.
    return {};
  }
 
  bool validateAsmConstraint(const char *&Name,
                             TargetInfo::ConstraintInfo &Info) const override {
    switch (*Name) {
    default:
      return false;
    case 'c':
    case 'h':
    case 'r':
    case 'l':
    case 'f':
    case 'd':
    case 'q':
      Info.setAllowsRegister();
      return true;
    }
  }
 
  std::string_view getClobbers() const override {
    // FIXME: Is this really right?
    return "";
  }
 
  BuiltinVaListKind getBuiltinVaListKind() const override {
    return TargetInfo::CharPtrBuiltinVaList;
  }
 
  bool isValidCPUName(StringRef Name) const override {
    return StringToOffloadArch(Name) != OffloadArch::Unknown;
  }
 
  void fillValidCPUList(SmallVectorImpl<StringRef> &Values) const override {
    for (int i = static_cast<int>(OffloadArch::SM_20);
         i < static_cast<int>(OffloadArch::Generic); ++i)
      Values.emplace_back(OffloadArchToString(static_cast<OffloadArch>(i)));
  }
 
  bool setCPU(StringRef Name) override {
    GPU = StringToOffloadArch(Name);
    return GPU != OffloadArch::Unknown;
  }
 
  void setSupportedOpenCLOpts() override {
    auto &Opts = getSupportedOpenCLOpts();
    Opts["cl_clang_storage_class_specifiers"] = true;
    Opts["__cl_clang_function_pointers"] = true;
    Opts["__cl_clang_variadic_functions"] = true;
    Opts["__cl_clang_function_scope_local_variables"] = true;
    Opts["__cl_clang_non_portable_kernel_param_types"] = true;
    Opts["__cl_clang_bitfields"] = true;
 
    Opts["cl_khr_fp64"] = true;
    Opts["__opencl_c_fp64"] = true;
    Opts["cl_khr_byte_addressable_store"] = true;
    Opts["cl_khr_global_int32_base_atomics"] = true;
    Opts["cl_khr_global_int32_extended_atomics"] = true;
    Opts["cl_khr_local_int32_base_atomics"] = true;
    Opts["cl_khr_local_int32_extended_atomics"] = true;
 
    Opts["__opencl_c_images"] = true;
    Opts["__opencl_c_3d_image_writes"] = true;
    Opts["cl_khr_3d_image_writes"] = true;
 
    Opts["__opencl_c_generic_address_space"] = true;
  }
 
  const llvm::omp::GV &getGridValue() const override {
    return llvm::omp::NVPTXGridValues;
  }
 
  /// \returns If a target requires an address within a target specific address
  /// space \p AddressSpace to be converted in order to be used, then return the
  /// corresponding target specific DWARF address space.
  ///
  /// \returns Otherwise return std::nullopt and no conversion will be emitted
  /// in the DWARF.
  std::optional<unsigned>
  getDWARFAddressSpace(unsigned AddressSpace) const override {
    if (AddressSpace >= std::size(NVPTXDWARFAddrSpaceMap) ||
        NVPTXDWARFAddrSpaceMap[AddressSpace] < 0)
      return std::nullopt;
    return NVPTXDWARFAddrSpaceMap[AddressSpace];
  }
 
  CallingConvCheckResult checkCallingConvention(CallingConv CC) const override {
    // CUDA compilations support all of the host's calling conventions.
    //
    // TODO: We should warn if you apply a non-default CC to anything other than
    // a host function.
    if (HostTarget)
      return HostTarget->checkCallingConvention(CC);
    return CC == CC_DeviceKernel ? CCCR_OK : CCCR_Warning;
  }
 
  bool hasBitIntType() const override { return true; }
  bool hasBFloat16Type() const override { return true; }
 
  OffloadArch getGPU() const { return GPU; }
};
} // namespace targets
} // namespace clang
#endif // LLVM_CLANG_LIB_BASIC_TARGETS_NVPTX_H