Helpers.h

Go to the documentation of this file.

#ifndef PROTEUS_PASS_HELPERS_H
#define PROTEUS_PASS_HELPERS_H
 
#include "proteus/Error.h"
#include "proteus/impl/CompilerInterfaceRuntimeConstantInfo.h"
#include "proteus/impl/Logger.h"
 
#include <llvm/ADT/SetVector.h>
#include <llvm/Demangle/Demangle.h>
#include <llvm/IR/Module.h>
#include <llvm/TargetParser/Triple.h>
 
#define DEBUG_TYPE "proteus-pass"
#define DEBUG(x)                                                               \
  do                                                                           \
    if (isDebugOutputEnabled()) {                                              \
      x;                                                                       \
    }                                                                          \
  while (0);
 
#if PROTEUS_ENABLE_HIP
constexpr char const *RegisterFunctionName = "__hipRegisterFunction";
constexpr char const *LaunchFunctionName = "hipLaunchKernel";
constexpr char const *RegisterVarName = "__hipRegisterVar";
constexpr char const *RegisterFatBinaryName = "__hipRegisterFatBinary";
#elif PROTEUS_ENABLE_CUDA
constexpr char const *RegisterFunctionName = "__cudaRegisterFunction";
constexpr char const *LaunchFunctionName = "cudaLaunchKernel";
constexpr char const *RegisterVarName = "__cudaRegisterVar";
constexpr char const *RegisterFatBinaryName = "__cudaRegisterFatBinary";
#else
constexpr char const *RegisterFunctionName = nullptr;
constexpr char const *LaunchFunctionName = nullptr;
constexpr char const *RegisterVarName = nullptr;
constexpr char const *RegisterFatBinaryName = nullptr;
#endif
 
namespace proteus {
 
using namespace llvm;
 
struct JitFunctionInfo {
  SmallSetVector<RuntimeConstantInfo, 16> ConstantArgs;
  std::string ModuleIR;
};
 
struct ModuleInfo {
  const Module &M;
  ModuleInfo(const Module &M) : M(M) {}
};
 
inline bool isDebugOutputEnabled() {
  auto GetEnvVar = []() {
    const char *EnvValue = std::getenv("PROTEUS_DEBUG_OUTPUT");
    return EnvValue ? static_cast<bool>(std::stoi(EnvValue)) : false;
  };
 
  static bool IsEnabled = GetEnvVar();
  return IsEnabled;
}
 
bool inline isDeviceCompilation(Module &M) {
  Triple TargetTriple(M.getTargetTriple());
  DEBUG(Logger::logs("proteus-pass")
        << "TargetTriple " << TargetTriple.str() << "\n");
  if (TargetTriple.isNVPTX() || TargetTriple.isAMDGCN())
    return true;
 
  return false;
}
 
inline std::string getUniqueFileID(Module &M) {
  llvm::sys::fs::UniqueID ID;
  if (auto EC = llvm::sys::fs::getUniqueID(M.getSourceFileName(), ID))
    reportFatalError("Could not get unique id for source file " + EC.message());
 
  SmallString<64> Out;
  llvm::raw_svector_ostream OutStr(Out);
  OutStr << llvm::format("%x_%x", ID.getDevice(), ID.getFile());
 
  return std::string(Out);
}
 
// Resolve the module source path so consumers of the embedded bitcode can
// locate the translation unit independently of the compile working directory.
inline std::string getCanonicalSourceFileName(Module &M) {
  SmallString<256> RealPath;
  if (llvm::sys::fs::real_path(M.getSourceFileName(), RealPath))
    return M.getSourceFileName();
 
  return std::string(RealPath);
}
 
inline bool isDeviceKernel(const Function *F) {
  if (!F)
    reportFatalError("Expected non-null function");
 
#if PROTEUS_ENABLE_CUDA
#if LLVM_VERSION_MAJOR >= 20
  return (F->getCallingConv() == CallingConv::PTX_Kernel);
#else
  const Module &M = *F->getParent();
  auto GetDeviceKernels = [&M]() {
    SmallPtrSet<Function *, 16> Kernels;
    NamedMDNode *MD = M.getNamedMetadata("nvvm.annotations");
 
    if (!MD)
      return Kernels;
 
    for (auto *Op : MD->operands()) {
      if (Op->getNumOperands() < 2)
        continue;
      MDString *KindID = dyn_cast<MDString>(Op->getOperand(1));
      if (!KindID || KindID->getString() != "kernel")
        continue;
 
      Function *KernelFn =
          mdconst::dyn_extract_or_null<Function>(Op->getOperand(0));
      if (!KernelFn)
        continue;
 
      Kernels.insert(KernelFn);
    }
 
    return Kernels;
  };
 
  // Create a kernel cache per module, assumes we don't insert/remove kernels
  // after parsing nvvm.annotations.
  static DenseMap<const Module *, SmallPtrSet<Function *, 16>> KernelCache;
  auto It = KernelCache.find(&M);
  if (It == KernelCache.end())
    It = KernelCache.insert({&M, GetDeviceKernels()}).first;
  const auto &KernelSet = It->second;
  if (KernelSet.contains(F))
    return true;
 
  return false;
#endif
#endif
 
#if PROTEUS_ENABLE_HIP
  return (F->getCallingConv() == CallingConv::AMDGPU_KERNEL);
#endif
 
  return false;
}
 
} // namespace proteus
 
namespace llvm {
 
using namespace proteus;
 
template <> struct DenseMapInfo<RuntimeConstantInfo> {
  static inline RuntimeConstantInfo getEmptyKey() {
    RuntimeConstantInfo K{BEGIN, -1};
    return K;
  }
 
  static inline RuntimeConstantInfo getTombstoneKey() {
    RuntimeConstantInfo K{END, -1};
    return K;
  }
 
  static unsigned getHashValue(const RuntimeConstantInfo &Val) {
    return hash_combine(Val.ArgInfo.Type, Val.ArgInfo.Pos);
  }
 
  static bool isEqual(const RuntimeConstantInfo &LHS,
                      const RuntimeConstantInfo &RHS) {
    return ((LHS.ArgInfo.Type == RHS.ArgInfo.Type) &&
            (LHS.ArgInfo.Pos == RHS.ArgInfo.Pos));
  }
};
 
inline bool isDeviceKernelHostStub(
    const DenseMap<Value *, GlobalVariable *> &StubToKernelMap, Function &Fn) {
  if (StubToKernelMap.contains(&Fn))
    return true;
 
  return false;
}
 
} // namespace llvm
 
#endif