//===- comgr-compiler.cpp - Comgr compiler Action internals ---------------===//
//
// Part of Comgr, under the Apache License v2.0 with LLVM Exceptions. See
// amd/comgr/LICENSE.TXT in this repository for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
///
/// \file
/// This file implements the compilation and compilation-adjacent
/// AMD_COMGR_ACTIONs. Many of these leverage Comgr's AMDGPUCompiler class.
///
//===----------------------------------------------------------------------===//

#include "comgr-compiler.h"
#include "comgr-cache.h"
#include "comgr-clang-command.h"
#include "comgr-device-libs.h"
#include "comgr-diagnostic-handler.h"
#include "comgr-env.h"
#include "comgr-libcxx-headers.h"
#include "comgr-resource-directory.h"
#include "comgr-spirv-command.h"
#include "comgr-unbundle-command.h"
#include "lld/Common/CommonLinkerContext.h"
#include "lld/Common/Driver.h"
#include "clang/CodeGen/CodeGenAction.h"
#include "clang/Driver/Compilation.h"
#include "clang/Driver/DriverDiagnostic.h"
#include "clang/Driver/Job.h"
#include "clang/Driver/OffloadBundler.h"
#include "clang/Driver/Tool.h"
#include "clang/Frontend/CompilerInstance.h"
#include "clang/Frontend/FrontendDiagnostic.h"
#include "clang/Frontend/TextDiagnosticPrinter.h"
#include "clang/FrontendTool/Utils.h"
#include "clang/Options/Options.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/Bitcode/BitcodeWriter.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/Verifier.h"
#include "llvm/IRReader/IRReader.h"
#include "llvm/Linker/Linker.h"
#include "llvm/MC/MCAsmBackend.h"
#include "llvm/MC/MCAsmInfo.h"
#include "llvm/MC/MCCodeEmitter.h"
#include "llvm/MC/MCContext.h"
#include "llvm/MC/MCInstPrinter.h"
#include "llvm/MC/MCInstrInfo.h"
#include "llvm/MC/MCObjectFileInfo.h"
#include "llvm/MC/MCObjectWriter.h"
#include "llvm/MC/MCParser/MCAsmParser.h"
#include "llvm/MC/MCParser/MCTargetAsmParser.h"
#include "llvm/MC/MCRegisterInfo.h"
#include "llvm/MC/MCStreamer.h"
#include "llvm/MC/MCSubtargetInfo.h"
#include "llvm/MC/MCTargetOptions.h"
#include "llvm/MC/TargetRegistry.h"
#include "llvm/Object/Archive.h"
#include "llvm/Support/Error.h"
#include "llvm/Support/FileSystem.h"
#include "llvm/Support/MemoryBuffer.h"
#include "llvm/Support/Path.h"
#include "llvm/Support/Process.h"
#include "llvm/Support/Signals.h"
#include "llvm/Support/VirtualFileSystem.h"
#include "llvm/Support/WithColor.h"
#include "llvm/TargetParser/Host.h"
#include "llvm/Transforms/Utils/Cloning.h"

#include "time-stat/ts-interface.h"

#ifdef COMGR_SPIRV_TRANSLATOR_AVAILABLE
#include <LLVMSPIRVLib.h>
#endif

#include <csignal>
#include <sstream>

LLD_HAS_DRIVER(elf)

using namespace llvm;
using namespace llvm::opt;
using namespace llvm::sys;
using namespace clang;
using namespace clang::driver;
using namespace clang::options;
using namespace COMGR::TimeStatistics;

namespace COMGR {

namespace {
constexpr llvm::StringLiteral LinkerJobName = "amdgpu::Linker";

/// \brief Helper class for representing a single invocation of the assembler.
struct AssemblerInvocation {
  /// @name Target Options
  /// @{

  /// The name of the target triple to assemble for.
  std::string Triple;

  /// If given, the name of the target CPU to determine which instructions
  /// are legal.
  std::string CPU;

  /// The list of target specific features to enable or disable -- this should
  /// be a list of strings starting with '+' or '-'.
  std::vector<std::string> Features;

  /// The list of symbol definitions.
  std::vector<std::string> SymbolDefs;

  /// @}
  /// @name Language Options
  /// @{

  std::vector<std::string> IncludePaths;
  unsigned NoInitialTextSection : 1;
  unsigned SaveTemporaryLabels : 1;
  unsigned GenDwarfForAssembly : 1;
  unsigned RelaxELFRelocations : 1;
  unsigned DwarfVersion;
  std::string DwarfDebugFlags;
  std::string DwarfDebugProducer;
  std::string DebugCompilationDir;
  llvm::DebugCompressionType CompressDebugSections =
      llvm::DebugCompressionType::None;
  std::string MainFileName;

  /// @}
  /// @name Frontend Options
  /// @{

  std::string InputFile;
  std::vector<std::string> LLVMArgs;
  std::string OutputPath;
  enum FileType {
    FT_Asm,  ///< Assembly (.s) output, transliterate mode.
    FT_Null, ///< No output, for timing purposes.
    FT_Obj   ///< Object file output.
  };
  FileType OutputType;
  unsigned ShowHelp : 1;
  unsigned ShowVersion : 1;

  /// @}
  /// @name Transliterate Options
  /// @{

  unsigned OutputAsmVariant;
  unsigned ShowEncoding : 1;
  unsigned ShowInst : 1;

  /// @}
  /// @name Assembler Options
  /// @{

  unsigned RelaxAll : 1;
  unsigned NoExecStack : 1;
  unsigned FatalWarnings : 1;
  unsigned IncrementalLinkerCompatible : 1;

  /// The name of the relocation model to use.
  std::string RelocationModel;

  /// @}

public:
  AssemblerInvocation() {
    Triple = "";
    NoInitialTextSection = 0;
    InputFile = "-";
    OutputPath = "-";
    OutputType = FT_Asm;
    OutputAsmVariant = 0;
    ShowInst = 0;
    ShowEncoding = 0;
    RelaxAll = 0;
    NoExecStack = 0;
    FatalWarnings = 0;
    IncrementalLinkerCompatible = 0;
    DwarfVersion = 0;
  }

  static bool createFromArgs(AssemblerInvocation &Res,
                             ArrayRef<const char *> Argv,
                             DiagnosticsEngine &Diags);
};
} // namespace

bool AssemblerInvocation::createFromArgs(AssemblerInvocation &Opts,
                                         ArrayRef<const char *> Argv,
                                         DiagnosticsEngine &Diags) {
  bool Success = true;

  // Parse the arguments.
  const OptTable &OptTbl = getDriverOptTable();

  llvm::opt::Visibility VisibilityMask(options::CC1AsOption);
  unsigned MissingArgIndex, MissingArgCount;
  InputArgList Args =
      OptTbl.ParseArgs(Argv, MissingArgIndex, MissingArgCount, VisibilityMask);

  // Check for missing argument error.
  if (MissingArgCount) {
    Diags.Report(diag::err_drv_missing_argument)
        << Args.getArgString(MissingArgIndex) << MissingArgCount;
    Success = false;
  }

  // Issue errors on unknown arguments.
  for (const Arg *A : Args.filtered(OPT_UNKNOWN)) {
    auto ArgString = A->getAsString(Args);
    std::string Nearest;
    if (OptTbl.findNearest(ArgString, Nearest, VisibilityMask) > 1) {
      Diags.Report(diag::err_drv_unknown_argument) << ArgString;
    } else {
      Diags.Report(diag::err_drv_unknown_argument_with_suggestion)
          << ArgString << Nearest;
    }
    Success = false;
  }

  // Construct the invocation.

  // Target Options
  Opts.Triple = llvm::Triple::normalize(Args.getLastArgValue(OPT_triple));
  Opts.CPU = std::string(Args.getLastArgValue(OPT_target_cpu));
  Opts.Features = Args.getAllArgValues(OPT_target_feature);

  // Use the default target triple if unspecified.
  if (Opts.Triple.empty()) {
    Opts.Triple = llvm::sys::getDefaultTargetTriple();
  }

  // Language Options
  Opts.IncludePaths = Args.getAllArgValues(OPT_I);
  Opts.NoInitialTextSection = Args.hasArg(OPT_n);
  Opts.SaveTemporaryLabels = Args.hasArg(OPT_msave_temp_labels);
  // Any DebugInfoKind implies GenDwarfForAssembly.
  Opts.GenDwarfForAssembly = Args.hasArg(OPT_debug_info_kind_EQ);

  if (const Arg *A = Args.getLastArg(OPT_compress_debug_sections,
                                     OPT_compress_debug_sections_EQ)) {
    if (A->getOption().getID() == OPT_compress_debug_sections) {
      // TODO: be more clever about the compression type auto-detection
      Opts.CompressDebugSections = llvm::DebugCompressionType::Zlib;
    } else {
      Opts.CompressDebugSections =
          llvm::StringSwitch<llvm::DebugCompressionType>(A->getValue())
              .Case("none", llvm::DebugCompressionType::None)
              .Case("zlib", llvm::DebugCompressionType::Zlib)
              .Default(llvm::DebugCompressionType::None);
    }
  }

  Opts.RelaxELFRelocations = !Args.hasArg(OPT_mrelax_relocations_no);
  Opts.DwarfVersion = getLastArgIntValue(Args, OPT_dwarf_version_EQ, 2, Diags);
  Opts.DwarfDebugFlags =
      std::string(Args.getLastArgValue(OPT_dwarf_debug_flags));
  Opts.DwarfDebugProducer =
      std::string(Args.getLastArgValue(OPT_dwarf_debug_producer));
  Opts.DebugCompilationDir =
      std::string(Args.getLastArgValue(OPT_fdebug_compilation_dir));
  Opts.MainFileName = std::string(Args.getLastArgValue(OPT_main_file_name));

  // Frontend Options
  if (Args.hasArg(OPT_INPUT)) {
    bool First = true;
    for (const Arg *A : Args.filtered(OPT_INPUT)) {
      if (First) {
        Opts.InputFile = A->getValue();
        First = false;
      } else {
        Diags.Report(diag::err_drv_unknown_argument) << A->getAsString(Args);
        Success = false;
      }
    }
  }
  Opts.LLVMArgs = Args.getAllArgValues(OPT_mllvm);
  Opts.OutputPath = std::string(Args.getLastArgValue(OPT_o));
  if (Arg *A = Args.getLastArg(OPT_filetype)) {
    StringRef Name = A->getValue();
    unsigned OutputType = StringSwitch<unsigned>(Name)
                              .Case("asm", FT_Asm)
                              .Case("null", FT_Null)
                              .Case("obj", FT_Obj)
                              .Default(~0U);
    if (OutputType == ~0U) {
      Diags.Report(diag::err_drv_invalid_value) << A->getAsString(Args) << Name;
      Success = false;
    } else {
      Opts.OutputType = FileType(OutputType);
    }
  }
  Opts.ShowHelp = Args.hasArg(OPT_help);
  Opts.ShowVersion = Args.hasArg(OPT_version);

  // Transliterate Options
  Opts.OutputAsmVariant =
      getLastArgIntValue(Args, OPT_output_asm_variant, 0, Diags);
  Opts.ShowEncoding = Args.hasArg(OPT_show_encoding);
  Opts.ShowInst = Args.hasArg(OPT_show_inst);

  // Assemble Options
  Opts.RelaxAll = Args.hasArg(OPT_mrelax_all);
  Opts.NoExecStack = Args.hasArg(OPT_mno_exec_stack);
  Opts.FatalWarnings = Args.hasArg(OPT_massembler_fatal_warnings);
  Opts.RelocationModel =
      std::string(Args.getLastArgValue(OPT_mrelocation_model, "pic"));
  Opts.IncrementalLinkerCompatible =
      Args.hasArg(OPT_mincremental_linker_compatible);
  Opts.SymbolDefs = Args.getAllArgValues(OPT_defsym);

  return Success;
}

namespace {
bool needsPreprocessing(DataObject *O) {
  if (O->DataKind != AMD_COMGR_DATA_KIND_SOURCE)
    return false;
  StringRef Ext = path::extension(O->Name);
  bool IsPreprocessedSource = Ext == ".i";
  return !IsPreprocessedSource;
}

std::unique_ptr<raw_fd_ostream> getOutputStream(AssemblerInvocation &Opts,
                                                DiagnosticsEngine &Diags,
                                                bool Binary) {
  if (Opts.OutputPath.empty()) {
    Opts.OutputPath = "-";
  }

  // Make sure that the Out file gets unlinked from the disk if we get a
  // SIGINT.
  if (Opts.OutputPath != "-") {
    sys::RemoveFileOnSignal(Opts.OutputPath);
  }

  std::error_code EC;
  auto Out = std::make_unique<raw_fd_ostream>(
      Opts.OutputPath, EC, (Binary ? sys::fs::OF_None : sys::fs::OF_Text));
  if (EC) {
    Diags.Report(diag::err_fe_unable_to_open_output)
        << Opts.OutputPath << EC.message();
    return nullptr;
  }

  return Out;
}

// clang/tools/driver/cc1as_main.cpp,  ExecuteAssemblerImpl()
bool executeAssemblerImpl(AssemblerInvocation &Opts, DiagnosticsEngine &Diags,
                          raw_ostream &LogS) {
  // Get the target specific parser.
  std::string Error;
  const Target *TheTarget = TargetRegistry::lookupTarget(
    llvm::Triple(Opts.Triple), Error);
  if (!TheTarget) {
    return Diags.Report(diag::err_target_unknown_triple) << Opts.Triple;
  }

  ErrorOr<std::unique_ptr<MemoryBuffer>> Buffer =
      MemoryBuffer::getFileOrSTDIN(Opts.InputFile);

  if (std::error_code EC = Buffer.getError()) {
    Error = EC.message();
    return Diags.Report(diag::err_fe_error_reading) << Opts.InputFile;
  }

  SourceMgr SrcMgr;
  SrcMgr.setDiagHandler(
      [](const SMDiagnostic &SMDiag, void *LogS) {
        SMDiag.print("", *(raw_ostream *)LogS, /* ShowColors */ false);
      },
      &LogS);

  // Tell SrcMgr about this buffer, which is what the parser will pick up.
  SrcMgr.AddNewSourceBuffer(std::move(*Buffer), SMLoc());

  // Record the location of the include directories so that the lexer can find
  // it later.
  SrcMgr.setIncludeDirs(Opts.IncludePaths);

  std::unique_ptr<MCRegisterInfo> MRI(TheTarget->createMCRegInfo(
      llvm::Triple(Opts.Triple)));
  assert(MRI && "Unable to create target register info!");

  llvm::MCTargetOptions MCOptions;
  MCOptions.X86RelaxRelocations = Opts.RelaxELFRelocations;
  MCOptions.CompressDebugSections = Opts.CompressDebugSections;
  std::unique_ptr<MCAsmInfo> MAI(
      TheTarget->createMCAsmInfo(*MRI, llvm::Triple(Opts.Triple), MCOptions));
  assert(MAI && "Unable to create target asm info!");

  // Ensure MCAsmInfo initialization occurs before any use, otherwise sections
  // may be created with a combination of default and explicit settings.

  bool IsBinary = Opts.OutputType == AssemblerInvocation::FT_Obj;
  std::unique_ptr<raw_fd_ostream> FDOS = getOutputStream(Opts, Diags, IsBinary);
  if (!FDOS) {
    return true;
  }

  // Build up the feature string from the target feature list.
  std::string FS;
  if (!Opts.Features.empty()) {
    FS = Opts.Features[0];
    for (unsigned I = 1, E = Opts.Features.size(); I != E; ++I) {
      FS += "," + Opts.Features[I];
    }
  }

  std::unique_ptr<MCObjectFileInfo> MOFI(new MCObjectFileInfo());
  std::unique_ptr<MCSubtargetInfo> STI(
      TheTarget->createMCSubtargetInfo(llvm::Triple(Opts.Triple), Opts.CPU, FS));

  MCContext Ctx(Triple(Opts.Triple), *MAI, *MRI, *STI, &SrcMgr);
  Ctx.setObjectFileInfo(MOFI.get());

  bool PIC = false;
  if (Opts.RelocationModel == "static") {
    PIC = false;
  } else if (Opts.RelocationModel == "pic") {
    PIC = true;
  } else {
    assert(Opts.RelocationModel == "dynamic-no-pic" && "Invalid PIC model!");
    PIC = false;
  }

  MOFI->initMCObjectFileInfo(Ctx, PIC);
  if (Opts.GenDwarfForAssembly) {
    Ctx.setGenDwarfForAssembly(true);
  }
  if (!Opts.DwarfDebugFlags.empty()) {
    Ctx.setDwarfDebugFlags(StringRef(Opts.DwarfDebugFlags));
  }
  if (!Opts.DwarfDebugProducer.empty()) {
    Ctx.setDwarfDebugProducer(StringRef(Opts.DwarfDebugProducer));
  }
  if (!Opts.DebugCompilationDir.empty()) {
    Ctx.setCompilationDir(Opts.DebugCompilationDir);
  }
  if (!Opts.MainFileName.empty()) {
    Ctx.setMainFileName(StringRef(Opts.MainFileName));
  }
  Ctx.setDwarfVersion(Opts.DwarfVersion);

  std::unique_ptr<MCStreamer> Str;
  std::unique_ptr<MCInstrInfo> MCII(TheTarget->createMCInstrInfo());

  raw_pwrite_stream *Out = FDOS.get();
  std::unique_ptr<buffer_ostream> BOS;

  // FIXME: There is a bit of code duplication with addPassesToEmitFile.
  if (Opts.OutputType == AssemblerInvocation::FT_Asm) {
    std::unique_ptr<MCInstPrinter> InstructionPrinter(
      TheTarget->createMCInstPrinter(
         llvm::Triple(Opts.Triple), Opts.OutputAsmVariant, *MAI, *MCII, *MRI));
    std::unique_ptr<MCCodeEmitter> MCE;
    std::unique_ptr<MCAsmBackend> MAB;
    if (Opts.ShowEncoding) {
      MCE.reset(TheTarget->createMCCodeEmitter(*MCII, Ctx));
      MCTargetOptions Options;
      MAB.reset(TheTarget->createMCAsmBackend(*STI, *MRI, Options));
    }
    auto FOut = std::make_unique<formatted_raw_ostream>(*Out);
    Str.reset(TheTarget->createAsmStreamer(Ctx, std::move(FOut), std::move(InstructionPrinter),
                                           std::move(MCE), std::move(MAB)));
  } else if (Opts.OutputType == AssemblerInvocation::FT_Null) {
    Str.reset(createNullStreamer(Ctx));
  } else {
    assert(Opts.OutputType == AssemblerInvocation::FT_Obj &&
           "Invalid file type!");
    if (!FDOS->supportsSeeking()) {
      BOS = std::make_unique<buffer_ostream>(*FDOS);
      Out = BOS.get();
    }

    MCCodeEmitter *CE = TheTarget->createMCCodeEmitter(*MCII, Ctx);
    MCTargetOptions Options;
    MCAsmBackend *MAB = TheTarget->createMCAsmBackend(*STI, *MRI, Options);
    Triple T(Opts.Triple);
    Str.reset(TheTarget->createMCObjectStreamer(
        T, Ctx, std::unique_ptr<MCAsmBackend>(MAB),
        MAB->createObjectWriter(*Out), std::unique_ptr<MCCodeEmitter>(CE),
        *STI));
    Str.get()->initSections(*STI);
  }

  bool Failed = false;

  std::unique_ptr<MCAsmParser> Parser(
      createMCAsmParser(SrcMgr, Ctx, *Str.get(), *MAI));

  // FIXME: init MCTargetOptions from sanitizer flags here.
  MCTargetOptions Options;
  std::unique_ptr<MCTargetAsmParser> TAP(
      TheTarget->createMCAsmParser(*STI, *Parser, *MCII));
  if (!TAP) {
    Failed = Diags.Report(diag::err_target_unknown_triple) << Opts.Triple;
  }

  // Set values for symbols, if any.
  for (auto &S : Opts.SymbolDefs) {
    auto Pair = StringRef(S).split('=');
    auto Sym = Pair.first;
    auto Val = Pair.second;
    int64_t Value;
    // We have already error checked this in the driver.
    if (!Val.getAsInteger(0, Value)) {
      Ctx.setSymbolValue(Parser->getStreamer(), Sym, Value);
    }
  }

  if (!Failed) {
    Parser->setTargetParser(*TAP.get());
    Failed = Parser->Run(Opts.NoInitialTextSection);
  }

  return Failed;
}

bool executeAssembler(AssemblerInvocation &Opts, DiagnosticsEngine &Diags,
                      raw_ostream &LogS) {
  bool Failed = executeAssemblerImpl(Opts, Diags, LogS);

  // Delete output file if there were errors.
  if (Failed && Opts.OutputPath != "-") {
    sys::fs::remove(Opts.OutputPath);
  }

  return Failed;
}

SmallString<128> getFilePath(DataObject *Object, StringRef Dir) {
  SmallString<128> Path(Dir);
  path::append(Path, Object->Name);

  // Create directories specified in the File Path so that the in-process driver
  // can successfully execute clang commands that use this file path as an
  // output argument
  if (fs::create_directories(path::parent_path(Path))) {
    return SmallString<128>();
  }

  return Path;
}

// TODO: Move inputFromFile and outputToFile within AMDGPUCompiler
//
// Currently, we only invoke these two methods in the context of AMDGPUCompiler.
// Moreover, member functions that deal with file I/O should not worry whether
// the underlying filesystem being used is virtual or real.
amd_comgr_status_t inputFromFile(DataObject *Object, StringRef Path) {
  ProfilePoint Point("FileIO");
  auto BufOrError = MemoryBuffer::getFile(Path);
  if (std::error_code EC = BufOrError.getError()) {
    return AMD_COMGR_STATUS_ERROR;
  }
  Object->setData(BufOrError.get()->getBuffer());
  return AMD_COMGR_STATUS_SUCCESS;
}

amd_comgr_status_t outputToFile(StringRef Data, StringRef Path) {
  SmallString<128> DirPath = Path;
  path::remove_filename(DirPath);
  {
    ProfilePoint Point("CreateDir");
    if (fs::create_directories(DirPath)) {
      return AMD_COMGR_STATUS_ERROR;
    }
  }
  std::error_code EC;
  ProfilePoint Point("FileIO");
  raw_fd_ostream OS(Path, EC, fs::OF_None);
  if (EC) {
    return AMD_COMGR_STATUS_ERROR;
  }
  OS << Data;
  OS.close();
  if (OS.has_error()) {
    return AMD_COMGR_STATUS_ERROR;
  }
  return AMD_COMGR_STATUS_SUCCESS;
}

amd_comgr_status_t outputToFile(DataObject *Object, StringRef Path) {
  return outputToFile(StringRef(Object->Data, Object->Size), Path);
}

void initializeCommandLineArgs(SmallVectorImpl<const char *> &Args) {
  // Workaround for flawed Driver::BuildCompilation(...) implementation,
  // which eliminates 1st argument, cause it actually awaits argv[0].
  Args.clear();
  Args.push_back("");
}

// Parse -mllvm options
amd_comgr_status_t parseLLVMOptions(const std::vector<std::string> &Options) {
  std::vector<const char *> LLVMArgs;
  for (auto Option : Options) {
    LLVMArgs.push_back("");
    LLVMArgs.push_back(Option.c_str());
    if (!cl::ParseCommandLineOptions(LLVMArgs.size(), &LLVMArgs[0],
                                     "-mllvm options parsing")) {
      return AMD_COMGR_STATUS_ERROR_INVALID_ARGUMENT;
    }
    LLVMArgs.clear();
  }
  return AMD_COMGR_STATUS_SUCCESS;
}

amd_comgr_status_t linkWithLLD(llvm::ArrayRef<const char *> Args,
                               llvm::raw_ostream &LogS,
                               llvm::raw_ostream &LogE) {
  ArgStringList LLDArgs(llvm::iterator_range<ArrayRef<const char *>::iterator>(
      Args.begin(), Args.end()));
  LLDArgs.insert(LLDArgs.begin(), "ld.lld");
  LLDArgs.push_back("--threads=1");

  ArrayRef<const char *> ArgRefs = llvm::ArrayRef(LLDArgs);
  lld::Result LLDRet =
      lld::lldMain(ArgRefs, LogS, LogE, {{lld::Gnu, &lld::elf::link}});
  lld::CommonLinkerContext::destroy();
  if (LLDRet.retCode || !LLDRet.canRunAgain) {
    return AMD_COMGR_STATUS_ERROR;
  }
  return AMD_COMGR_STATUS_SUCCESS;
}

// Execute llvm-link in-process using llvm::Linker
// Args format: -o <output.bc> <input1.bc> <input2.bc> ...
// TODO: refactor this implementation to use a shared infra with linkBitcodeToBitcode()
amd_comgr_status_t executeLLVMLink(ArrayRef<const char *> Args,
                                   raw_ostream &LogS) {
  // Parse args: find -o <output> and collect input .bc files
  StringRef OutputPath;
  SmallVector<StringRef, 4> InputPaths;

  for (size_t I = 0; I < Args.size(); ++I) {
    StringRef Arg(Args[I]);
    if (Arg == "-o" && I + 1 < Args.size()) {
      OutputPath = Args[++I];
    } else if (Arg.ends_with(".bc")) {
      InputPaths.push_back(Arg);
    }
  }

  if (OutputPath.empty() || InputPaths.empty()) {
    LogS << "llvm-link: missing input or output files\n";
    return AMD_COMGR_STATUS_ERROR;
  }

  // Create composite module and linker
  LLVMContext Context;
  auto Composite = std::make_unique<llvm::Module>("llvm-link", Context);
  Linker L(*Composite);

  // Link each input BC
  for (StringRef InputPath : InputPaths) {
    auto BufOrErr = MemoryBuffer::getFile(InputPath);
    if (!BufOrErr) {
      LogS << "llvm-link: failed to read: " << InputPath << "\n";
      return AMD_COMGR_STATUS_ERROR;
    }
    SMDiagnostic Err;
    auto Mod = parseIR(BufOrErr->get()->getMemBufferRef(), Err, Context);
    if (!Mod) {
      Err.print("llvm-link", LogS);
      return AMD_COMGR_STATUS_ERROR;
    }
    if (L.linkInModule(std::move(Mod))) {
      LogS << "llvm-link: linking failed for: " << InputPath << "\n";
      return AMD_COMGR_STATUS_ERROR;
    }
  }

  // Write linked BC to output file
  std::error_code EC;
  raw_fd_ostream OS(OutputPath, EC);
  if (EC) {
    LogS << "llvm-link: failed to open output: " << OutputPath << "\n";
    return AMD_COMGR_STATUS_ERROR;
  }
  WriteBitcodeToFile(*Composite, OS);
  return AMD_COMGR_STATUS_SUCCESS;
}

#ifdef COMGR_SPIRV_TRANSLATOR_AVAILABLE
namespace {

// Map "SPV_…" extension name → SPIRV::ExtensionID using the x-macro shipped
// with the translator. Mirrors the table that amd-llvm-spirv builds in
// parseSPVExtOption (llvm-spirv.cpp).
const llvm::DenseMap<StringRef, SPIRV::ExtensionID> &spirvExtensionNameMap() {
  static const auto Map = []() {
    llvm::DenseMap<StringRef, SPIRV::ExtensionID> M;
#define EXT(X) M[#X] = SPIRV::ExtensionID::X;
#include "LLVMSPIRVExtensions.inc"
#undef EXT
    return M;
  }();
  return Map;
}

bool parseSpirvExtList(StringRef Spec,
                       SPIRV::TranslatorOpts::ExtensionsStatusMap &Status,
                       raw_ostream &LogS) {
  const auto &Names = spirvExtensionNameMap();
  SmallVector<StringRef, 16> Items;
  Spec.split(Items, ',', -1, false);
  for (StringRef Item : Items) {
    if (Item.empty() || (Item.front() != '+' && Item.front() != '-')) {
      LogS << "spirv-translator: invalid --spirv-ext value '" << Item
           << "' (expected +EXT or -EXT)\n";
      return false;
    }
    bool Allow = Item.front() == '+';
    StringRef Name = Item.drop_front();
    if (Name == "all") {
      for (const auto &E : Names)
        Status[E.second] = Allow;
      continue;
    }
    auto It = Names.find(Name);
    if (It == Names.end()) {
      LogS << "spirv-translator: unknown extension '" << Name
           << "' in --spirv-ext\n";
      return false;
    }
    Status[It->second] = Allow;
  }
  return true;
}

// Parse the --spirv-* subset of the amd-llvm-spirv CLI that the HIPAMD clang
// driver emits (see clang/lib/Driver/ToolChains/HIPAMD.cpp), populating Opts
// to match. Also extracts InputPath (.bc) and OutputPath (-o argument).
// Unrecognized --spirv-* flags are logged and ignored.
bool parseSPIRVTranslatorArgs(ArrayRef<const char *> Args,
                              SPIRV::TranslatorOpts &Opts, StringRef &InputPath,
                              StringRef &OutputPath, raw_ostream &LogS) {
  SPIRV::VersionNumber MaxVer = SPIRV::VersionNumber::MaximumVersion;
  SPIRV::TranslatorOpts::ExtensionsStatusMap ExtStatus;
  std::optional<SPIRV::DebugInfoEIS> DebugEIS;
  std::optional<SPIRV::TranslatorOpts::ArgList> AllowUnknownIntrinsics;
  bool PreserveAux = false;

  for (size_t I = 0, N = Args.size(); I < N; ++I) {
    if (!Args[I])
      continue;
    StringRef Arg(Args[I]);

    if (Arg == "-o") {
      if (I + 1 >= N) {
        LogS << "spirv-translator: '-o' requires an argument\n";
        return false;
      }
      OutputPath = Args[++I];
      continue;
    }
    if (Arg.ends_with(".bc")) {
      InputPath = Arg;
      continue;
    }

    // Normalize: strip leading "--" or "-" so we can compare against bare
    // names.
    StringRef Body = Arg;
    if (!Body.consume_front("--") && !Body.consume_front("-"))
      continue;

    auto EqPos = Body.find('=');
    StringRef Name = Body.substr(0, EqPos);
    bool HasValue = EqPos != StringRef::npos;
    StringRef Value = HasValue ? Body.substr(EqPos + 1) : StringRef();

    if (Name == "spirv-max-version" && HasValue) {
      using V_ = SPIRV::VersionNumber;
      auto Parsed = llvm::StringSwitch<std::optional<V_>>(Value)
                        .Case("1.0", V_::SPIRV_1_0)
                        .Case("1.1", V_::SPIRV_1_1)
                        .Case("1.2", V_::SPIRV_1_2)
                        .Case("1.3", V_::SPIRV_1_3)
                        .Case("1.4", V_::SPIRV_1_4)
                        .Case("1.5", V_::SPIRV_1_5)
                        .Case("1.6", V_::SPIRV_1_6)
                        .Default(std::nullopt);
      if (!Parsed) {
        LogS << "spirv-translator: unknown --spirv-max-version '" << Value
             << "'\n";
        return false;
      }
      MaxVer = *Parsed;
    } else if (Name == "spirv-ext" && HasValue) {
      if (!parseSpirvExtList(Value, ExtStatus, LogS))
        return false;
    } else if (Name == "spirv-debug-info-version" && HasValue) {
      using EIS_ = SPIRV::DebugInfoEIS;
      auto Parsed = llvm::StringSwitch<std::optional<EIS_>>(Value)
                        .Case("legacy", EIS_::SPIRV_Debug)
                        .Case("ocl-100", EIS_::OpenCL_DebugInfo_100)
                        .Case("nonsemantic-shader-100",
                              EIS_::NonSemantic_Shader_DebugInfo_100)
                        .Case("nonsemantic-shader-200",
                              EIS_::NonSemantic_Shader_DebugInfo_200)
                        .Default(std::nullopt);
      if (!Parsed) {
        LogS << "spirv-translator: unknown --spirv-debug-info-version '"
             << Value << "'\n";
        return false;
      }
      DebugEIS = *Parsed;
    } else if (Name == "spirv-allow-unknown-intrinsics") {
      // Bare flag → allow all unknown intrinsics. With =prefix1,prefix2 →
      // restrict to listed prefixes. Matches cl::ValueOptional semantics in
      // llvm-spirv.cpp. The translator's isUnknownIntrinsicAllowed only
      // returns true if some prefix matches; an empty prefix matches every
      // intrinsic name (StringRef::starts_with("") is true), so represent
      // "allow all" as a single empty prefix rather than an empty list.
      SPIRV::TranslatorOpts::ArgList Prefixes;
      if (HasValue) {
        SmallVector<StringRef, 4> Parts;
        Value.split(Parts, ',', -1, false);
        for (StringRef P : Parts)
          Prefixes.push_back(P);
      } else {
        Prefixes.push_back(StringRef());
      }
      AllowUnknownIntrinsics = std::move(Prefixes);
    } else if (Name == "spirv-preserve-auxdata") {
      PreserveAux = true;
    } else if (Name == "spirv-lower-const-expr") {
      // No-op: this is a global cl::opt with cl::init(true) in
      // SPIRVLowerConstExpr.cpp; the default already matches. Recognized so we
      // don't warn.
    } else if (Name.starts_with("spirv-")) {
      LogS << "spirv-translator: ignoring unrecognized flag '" << Arg << "'\n";
    }
  }

  Opts = SPIRV::TranslatorOpts(MaxVer, ExtStatus);
  if (PreserveAux)
    Opts.setPreserveAuxData(true);
  if (DebugEIS)
    Opts.setDebugInfoEIS(*DebugEIS);
  if (AllowUnknownIntrinsics)
    Opts.setSPIRVAllowUnknownIntrinsics(*AllowUnknownIntrinsics);
  return true;
}

} // namespace

// Execute amd-llvm-spirv in-process using writeSpirv.
// Args format: [options...] <input.bc> -o <output.spv>
amd_comgr_status_t executeSPIRVTranslator(ArrayRef<const char *> Args,
                                          raw_ostream &LogS) {
  StringRef InputPath, OutputPath;
  SPIRV::TranslatorOpts Opts;
  if (!parseSPIRVTranslatorArgs(Args, Opts, InputPath, OutputPath, LogS))
    return AMD_COMGR_STATUS_ERROR;

  if (InputPath.empty() || OutputPath.empty()) {
    LogS << "spirv-translator: missing input or output files\n";
    return AMD_COMGR_STATUS_ERROR;
  }

  // Read input bitcode
  auto BufOrErr = MemoryBuffer::getFile(InputPath);
  if (!BufOrErr) {
    LogS << "spirv-translator: failed to read: " << InputPath << "\n";
    return AMD_COMGR_STATUS_ERROR;
  }

  LLVMContext Context;
  SMDiagnostic Err;
  auto Mod = parseIR(BufOrErr->get()->getMemBufferRef(), Err, Context);
  if (!Mod) {
    Err.print("spirv-translator", LogS);
    return AMD_COMGR_STATUS_ERROR;
  }

  // Translate to SPIRV
  std::string ErrMsg;
  std::ostringstream OSS;
  if (!writeSpirv(Mod.get(), Opts, OSS, ErrMsg)) {
    LogS << "spirv-translator: translation failed: " << ErrMsg << "\n";
    return AMD_COMGR_STATUS_ERROR;
  }

  // Write SPIRV to output file
  std::error_code EC;
  raw_fd_ostream OS(OutputPath, EC);
  if (EC) {
    LogS << "spirv-translator: failed to open output: " << OutputPath << "\n";
    return AMD_COMGR_STATUS_ERROR;
  }
  std::string Result = OSS.str();
  OS.write(Result.data(), Result.size());
  return AMD_COMGR_STATUS_SUCCESS;
}
#endif

void logArgv(raw_ostream &OS, StringRef ProgramName,
             ArrayRef<const char *> Argv) {
  OS << "     Driver Job Args: " << ProgramName;
  for (size_t I = 0; I < Argv.size(); ++I) {
    // Skip the first argument, which we replace with ProgramName, and the last
    // argument, which is a null terminator.
    if (I && Argv[I]) {
      OS << " \"" << Argv[I] << '\"';
    }
  }
  OS << '\n';
  OS.flush();
}

amd_comgr_status_t
executeCommand(const Command &Job, raw_ostream &LogS,
               DiagnosticOptions &DiagOpts,
               IntrusiveRefCntPtr<llvm::vfs::FileSystem> FS) {
  TextDiagnosticPrinter DiagClient(LogS, DiagOpts);
  IntrusiveRefCntPtr<DiagnosticIDs> DiagID(new DiagnosticIDs);
  DiagnosticsEngine Diags(DiagID, DiagOpts, &DiagClient, false);

  auto Arguments = Job.getArguments();
  SmallVector<const char *, 128> Argv;
  initializeCommandLineArgs(Argv);
  Argv.append(Arguments.begin(), Arguments.end());
  Argv.push_back(nullptr);

  clearLLVMOptions();

  if (Argv[1] == StringRef("-cc1")) {
    if (env::shouldEmitVerboseLogs()) {
      logArgv(LogS, "clang", Argv);
    }

    std::unique_ptr<CompilerInstance> Clang(new CompilerInstance());
    Clang->setVerboseOutputStream(LogS);
    Clang->setVirtualFileSystem(FS);
    if (!Argv.back()) {
      Argv.pop_back();
    }

    if (!CompilerInvocation::CreateFromArgs(Clang->getInvocation(), Argv,
                                            Diags)) {
      return AMD_COMGR_STATUS_ERROR;
    }
    // Internally this call refers to the invocation created above, so at
    // this point the DiagnosticsEngine should accurately reflect all user
    // requested configuration from Argv.
    Clang->createDiagnostics(&DiagClient, /* ShouldOwnClient */ false);
    if (!Clang->hasDiagnostics()) {
      return AMD_COMGR_STATUS_ERROR;
    }
    if (!ExecuteCompilerInvocation(Clang.get())) {
      return AMD_COMGR_STATUS_ERROR;
    }
  } else if (Argv[1] == StringRef("-cc1as")) {
    if (env::shouldEmitVerboseLogs()) {
      logArgv(LogS, "clang", Argv);
    }
    Argv.erase(Argv.begin() + 1);
    if (!Argv.back()) {
      Argv.pop_back();
    }
    AssemblerInvocation Asm;
    if (!AssemblerInvocation::createFromArgs(Asm, Argv, Diags)) {
      return AMD_COMGR_STATUS_ERROR;
    }
    if (auto Status = parseLLVMOptions(Asm.LLVMArgs)) {
      return Status;
    }
    if (executeAssembler(Asm, Diags, LogS)) {
      return AMD_COMGR_STATUS_ERROR;
    }
  } else if (Job.getCreator().getName() == LinkerJobName) {
    if (env::shouldEmitVerboseLogs()) {
      logArgv(LogS, "lld", Argv);
    }
    if (auto Status = linkWithLLD(Arguments, LogS, LogS)) {
      return Status;
    }
  } else {
    // Check executable name for additional tools (e.g., from AMDGCN::Linker)
    StringRef Executable = Job.getExecutable();
    StringRef ExeName = sys::path::filename(Executable);

    if (ExeName.contains("llvm-link")) {
      if (env::shouldEmitVerboseLogs()) {
        logArgv(LogS, "llvm-link", Argv);
      }
      return executeLLVMLink(Arguments, LogS);
    }
#ifdef COMGR_SPIRV_TRANSLATOR_AVAILABLE
    if (ExeName.contains("llvm-spirv")) {
      if (env::shouldEmitVerboseLogs()) {
        logArgv(LogS, "amd-llvm-spirv", Argv);
      }
      return executeSPIRVTranslator(Arguments, LogS);
    }
#endif

    LogS << "     Unhandled Job: " << Job.getCreator().getName()
         << " (executable: " << ExeName << ")\n";
    return AMD_COMGR_STATUS_ERROR;
  }
  return AMD_COMGR_STATUS_SUCCESS;
}

std::string getStableCUID(const DataSet *InSet) {
  using Hash = CachedCommandAdaptor::HashAlgorithm;
  Hash H;
  for (const DataObject *Input : InSet->DataObjects) {
    CachedCommandAdaptor::addFileContents(H,
                                          StringRef{Input->Data, Input->Size});
  }
  return toHex(H.final());
}
} // namespace

amd_comgr_status_t
AMDGPUCompiler::executeInProcessDriver(ArrayRef<const char *> Args) {
  // A DiagnosticsEngine is required at several points:
  //  * By the Driver in order to diagnose option parsing.
  //  * By the CompilerInvocation in order to diagnose option parsing.
  //  * By the CompilerInstance in order to diagnose everything else.
  // It is a chicken-and-egg problem in that you need some form of diagnostics
  // in order to diagnose options which further influence diagnostics. The code
  // here is mostly copy-and-pasted from driver.cpp/cc1_main.cpp/various Clang
  // tests to try to approximate the same behavior as running the `clang`
  // executable.
  std::unique_ptr<DiagnosticOptions> DiagOpts(new DiagnosticOptions);
  unsigned MissingArgIndex, MissingArgCount;
  InputArgList ArgList = getDriverOptTable().ParseArgs(
      Args.slice(1), MissingArgIndex, MissingArgCount);
  // We ignore MissingArgCount and the return value of ParseDiagnosticArgs. Any
  // errors that would be diagnosed here will also be diagnosed later, when the
  // DiagnosticsEngine actually exists.
  (void)ParseDiagnosticArgs(*DiagOpts, ArgList);
  TextDiagnosticPrinter *DiagClient =
      new TextDiagnosticPrinter(LogS, *DiagOpts);
  IntrusiveRefCntPtr<DiagnosticIDs> DiagID(new DiagnosticIDs);
  DiagnosticsEngine Diags(DiagID, *DiagOpts, DiagClient);

  ProcessWarningOptions(Diags, *DiagOpts, *OverlayFS, /*ReportDiags=*/false);

  Driver TheDriver((Twine(env::getLLVMPath()) + "/bin/clang").str(),
                   llvm::sys::getDefaultTargetTriple(), Diags,
                   "AMDGPU Code Object Manager", OverlayFS);
  TheDriver.setCheckInputsExist(false);

  // We do not want the driver to promote -include into -include-pch.
  // Otherwise, the driver may pick PCH in the wrong format, without permissions,
  // in the process's CWD.
  TheDriver.setProbePrecompiled(false);

  // Log arguments used to build compilation
  if (env::shouldEmitVerboseLogs()) {
    LogS << "    Compilation Args: ";
    for (size_t I = 1; I < Args.size(); ++I) {
      if (Args[I]) {
        LogS << " \"" << Args[I] << '\"';
      }
    }
    LogS << '\n';
    LogS.flush();
  }

  std::unique_ptr<Compilation> C(TheDriver.BuildCompilation(Args));
  if (!C || C->containsError()) {
    return AMD_COMGR_STATUS_ERROR;
  }

  auto Cache = CommandCache::get(LogS);
  for (auto &Job : C->getJobs()) {
    ClangCommand C(Job, *DiagOpts, OverlayFS, executeCommand);
    if (Cache) {
      if (auto Status = Cache->execute(C, LogS)) {
        return Status;
      }
    } else {
      if (auto Status = C.execute(LogS)) {
        return Status;
      }
    }
  }
  return AMD_COMGR_STATUS_SUCCESS;
}

amd_comgr_status_t AMDGPUCompiler::createTmpDirs() {
  static std::atomic<unsigned> Id = 0;
  static Process::Pid Pid = Process::getProcessId();

  std::string TmpDirPrefix("comgr-" + std::to_string(Pid) + "-" +
                           std::to_string(Id++));

  ProfilePoint Point("CreateDir");
  if (std::error_code EC = fs::createUniqueDirectory(TmpDirPrefix, TmpDir)) {
    if (env::shouldEmitVerboseLogs()) {
      LogS << "comgr-compiler: failed to create temporary directory '"
           << TmpDirPrefix << "': " << EC.message() << "\n";
      const char *TmpDirEnv = std::getenv("TMPDIR");
      if (TmpDirEnv)
        LogS << "comgr-compiler: TMPDIR='" << TmpDirEnv
             << "' may not exist or be writable\n";
    }
    return AMD_COMGR_STATUS_ERROR;
  }

  InputDir = TmpDir;
  path::append(InputDir, "input");
  if (fs::create_directory(InputDir)) {
    return AMD_COMGR_STATUS_ERROR;
  }

  OutputDir = TmpDir;
  path::append(OutputDir, "output");
  if (fs::create_directory(OutputDir)) {
    return AMD_COMGR_STATUS_ERROR;
  }

  IncludeDir = TmpDir;
  path::append(IncludeDir, "include");
  if (fs::create_directory(IncludeDir)) {
    return AMD_COMGR_STATUS_ERROR;
  }

  return AMD_COMGR_STATUS_SUCCESS;
}

// On windows fs::remove_directories takes huge time so use fs::remove.
#ifdef _WIN32
amd_comgr_status_t removeDirectory(const StringRef DirName) {
  std::error_code EC;
  for (fs::directory_iterator Dir(DirName, EC), DirEnd; Dir != DirEnd && !EC;
       Dir.increment(EC)) {
    const StringRef Path = Dir->path();

    fs::file_status Status;
    EC = fs::status(Path, Status);
    if (EC) {
      return AMD_COMGR_STATUS_ERROR;
    }

    switch (Status.type()) {
    case fs::file_type::regular_file:
      if (fs::remove(Path)) {
        return AMD_COMGR_STATUS_ERROR;
      }
      break;
    case fs::file_type::directory_file:
      if (removeDirectory(Path)) {
        return AMD_COMGR_STATUS_ERROR;
      }

      if (fs::remove(Path)) {
        return AMD_COMGR_STATUS_ERROR;
      }
      break;
    default:
      return AMD_COMGR_STATUS_ERROR_INVALID_ARGUMENT;
    }
  }

  if (fs::remove(DirName)) {
    return AMD_COMGR_STATUS_ERROR;
  }

  return AMD_COMGR_STATUS_SUCCESS;
}
#endif

amd_comgr_status_t AMDGPUCompiler::removeTmpDirs() {
  if (TmpDir.empty()) {
    return AMD_COMGR_STATUS_SUCCESS;
  }
  ProfilePoint Point("RemoveDir");
#ifndef _WIN32
  if (fs::remove_directories(TmpDir)) {
    return AMD_COMGR_STATUS_ERROR;
  }
  return AMD_COMGR_STATUS_SUCCESS;
#else
  return removeDirectory(TmpDir);
#endif
}

amd_comgr_status_t AMDGPUCompiler::processFile(DataObject *Input,
                                               const char *InputFilePath,
                                               const char *OutputFilePath) {
  SmallVector<const char *, 128> Argv = Args;

  for (auto &Option : ActionInfo->getOptions()) {
    Argv.push_back(Option.c_str());
    if (Option.rfind("--rocm-path", 0) == 0) {
      NoGpuLib = false;
    }
  }

  // The ROCm device library should be provided via --rocm-path. Otherwise
  // we can pass -nogpulib to build without the ROCm device library
  if (NoGpuLib) {
    Argv.push_back("-nogpulib");
  }

  // Auto-inject embedded libc++ headers as a fallback include path.
  // Using -idirafter places them AFTER all other include paths, so:
  //   - System libstdc++ or libc++ headers take priority when available
  //   - User-provided -I paths take priority
  //   - Embedded headers only kick in when no other C++ headers are found
  // This ensures backward compatibility while providing headers on systems
  // without C++ development headers (e.g., driver-only installs).
  if (HasEmbeddedHeaders && getLanguage() == AMD_COMGR_LANGUAGE_HIP) {
    SmallString<256> LibcxxPath(env::getLLVMPath());
    sys::path::append(LibcxxPath, "include", "c++", "v1");
    Argv.push_back("-idirafter");
    Argv.push_back(Saver.save(StringRef(LibcxxPath)).data());
  }

  // TODO: Enable this for OpenCL as well (SWDEV-377546)
  if (getLanguage() == AMD_COMGR_LANGUAGE_HIP && env::shouldSaveLLVMTemps()) {
    Argv.push_back("-save-temps=obj");
  }

  // Add SPIR-V flags
  for (auto Flag : Input->SpirvFlags) {
    Argv.push_back("-Xclang");
    Argv.push_back(Flag);
  }

  // By default clang driver will ask CC1 to leak memory.
  Argv.push_back("-Xclang");
  Argv.push_back("-no-disable-free");

  // Append options from AMD_COMGR_DRIVER_OPTIONS_APPEND environment variable.
  // Options are space-separated and appended after all other options.
  StringRef EnvOptions = env::getDriverOptionsAppend();
  if (!EnvOptions.empty()) {
    SmallVector<StringRef, 8> Options;
    EnvOptions.split(Options, ' ', /*MaxSplit=*/-1, /*KeepEmpty=*/false);
    for (StringRef Opt : Options)
      Argv.push_back(Saver.save(Opt).data());
  }

  Argv.push_back(InputFilePath);

  Argv.push_back("-o");
  Argv.push_back(OutputFilePath);

  return executeInProcessDriver(Argv);
}

amd_comgr_status_t
AMDGPUCompiler::processFiles(amd_comgr_data_kind_t OutputKind,
                             const char *OutputSuffix) {
  return processFiles(OutputKind, OutputSuffix, InSet);
}

amd_comgr_status_t
AMDGPUCompiler::processFiles(amd_comgr_data_kind_t OutputKind,
                             const char *OutputSuffix, DataSet *InSet) {
  for (auto *Input : InSet->DataObjects) {
    if (Input->DataKind != AMD_COMGR_DATA_KIND_INCLUDE) {
      continue;
    }
    auto IncludeFilePath = getFilePath(Input, IncludeDir);
    if (auto Status = outputToFile(Input, IncludeFilePath)) {
      return Status;
    }
  }

  for (auto *Input : InSet->DataObjects) {
    if (Input->DataKind != AMD_COMGR_DATA_KIND_SOURCE &&
        Input->DataKind != AMD_COMGR_DATA_KIND_BC &&
        Input->DataKind != AMD_COMGR_DATA_KIND_RELOCATABLE &&
        Input->DataKind != AMD_COMGR_DATA_KIND_EXECUTABLE) {
      continue;
    }

    auto InputFilePath = getFilePath(Input, InputDir);
    if (auto Status = outputToFile(Input, InputFilePath)) {
      return Status;
    }

    amd_comgr_data_t OutputT;
    if (auto Status = amd_comgr_create_data(OutputKind, &OutputT)) {
      return Status;
    }

    // OutputT can be released after addition to the data_set
    ScopedDataObjectReleaser SDOR(OutputT);

    DataObject *Output = DataObject::convert(OutputT);

    SmallString<128> OutputName(Input->Name);
    sys::path::replace_extension(OutputName, OutputSuffix);
    Output->setName(OutputName);

    auto OutputFilePath = getFilePath(Output, OutputDir);

    if (auto Status =
            processFile(Input, InputFilePath.c_str(), OutputFilePath.c_str())) {
      return Status;
    }

    if (auto Status = inputFromFile(Output, OutputFilePath)) {
      return Status;
    }

    if (auto Status = amd_comgr_data_set_add(OutSetT, OutputT)) {
      return Status;
    }
  }

  return AMD_COMGR_STATUS_SUCCESS;
}

amd_comgr_status_t AMDGPUCompiler::addIncludeFlags() {
  if (none_of(InSet->DataObjects, needsPreprocessing))
    return AMD_COMGR_STATUS_SUCCESS;

  amd_comgr_language_t Language = ActionInfo->Language;
  switch (Language) {
  case AMD_COMGR_LANGUAGE_OPENCL_1_2:
  case AMD_COMGR_LANGUAGE_OPENCL_2_0: {
    SmallString<128> OpenCLCBasePath = IncludeDir;
    sys::path::append(OpenCLCBasePath, "opencl-c-base.h");
    if (auto Status =
            outputToFile(getOpenCLCBaseHeaderContents(), OpenCLCBasePath)) {
      return Status;
    }
    Args.push_back("-include");
    Args.push_back(Saver.save(OpenCLCBasePath.c_str()).data());
    Args.push_back("-Xclang");
    Args.push_back("-fdeclare-opencl-builtins");
    break;
  }
  default:
    break;
  }

  if (ActionInfo->Path) {
    Args.push_back("-I");
    Args.push_back(ActionInfo->Path);
  }

  Args.push_back("-I");
  Args.push_back(IncludeDir.c_str());

  for (auto *Input : InSet->DataObjects) {
    if (Input->DataKind != AMD_COMGR_DATA_KIND_PRECOMPILED_HEADER) {
      continue;
    }
    PrecompiledHeaders.push_back(getFilePath(Input, IncludeDir));
    auto &PrecompiledHeaderPath = PrecompiledHeaders.back();
    if (auto Status = outputToFile(Input, PrecompiledHeaderPath)) {
      return Status;
    }
    Args.push_back("-include-pch");
    Args.push_back(PrecompiledHeaderPath.c_str());
    Args.push_back("-Xclang");
    Args.push_back("-fno-validate-pch");
  }

  bool CacheEnabled = CommandCache::get(LogS) != nullptr;
  if (PrecompiledHeaders.empty() && CacheEnabled) {
    // The -no-integrated-cpp is used to split the preprocessing stage from the
    // rest of the compilation jobs. The cache doesn't handle source-code input,
    // but can handle preprocessed input (to avoid dealing with includes).
    Args.push_back("-no-integrated-cpp");
    // The -dD option is used to keep the #define directives in the preprocessed
    // output. When -fdeclare-opencl-builtins is used, the opencl builtin
    // semantic analysis queries the preprocessor for macro definitions that
    // signal that an OpenCL feature is enabled. After preprocessing these
    // #define are gone, so the semantic analysis during the compilation stage
    // fails. This flag is used to keep them such that they are present during
    // the compilation stage.
    // Additionally, we need to keep the definitions for #pragma directives.
    // The preprocessor doesn't expand macro identifiers in #pragmas, and if we
    // do not pass -dD the definitions would be missing when clang parses the
    // code
    Args.push_back("-dD");
  }

  return AMD_COMGR_STATUS_SUCCESS;
}

amd_comgr_status_t
AMDGPUCompiler::addTargetIdentifierFlags(llvm::StringRef IdentStr,
                                         bool CompilingSrc = false) {
  TargetIdentifier Ident;
  if (auto Status = parseTargetIdentifier(IdentStr, Ident)) {
    return Status;
  }

  std::string GPUArch = Twine(Ident.Processor).str();
  if (!Ident.Features.empty()) {
    GPUArch += ":" + join(Ident.Features, ":");
  }

  if (CompilingSrc && getLanguage() == AMD_COMGR_LANGUAGE_HIP) {
    // OffloadArch
    Args.push_back(Saver.save(Twine("--offload-arch=") + GPUArch).data());
  } else {
    // Triple and CPU
    Args.push_back("-target");
    Args.push_back(
        Saver.save(Twine(Ident.Arch) + "-" + Ident.Vendor + "-" + Ident.OS)
            .data());
    Args.push_back(Saver.save(Twine("-mcpu=") + GPUArch).data());
  }

  return AMD_COMGR_STATUS_SUCCESS;
}

amd_comgr_status_t AMDGPUCompiler::addCompilationFlags() {
  // Default to O3 for all contexts
  Args.push_back("-O3");

  Args.push_back("-x");

  bool NeedsPreprocessing = any_of(InSet->DataObjects, needsPreprocessing);

  switch (ActionInfo->Language) {
  case AMD_COMGR_LANGUAGE_LLVM_IR:
    Args.push_back("ir");
    break;
  case AMD_COMGR_LANGUAGE_OPENCL_1_2:
    Args.push_back(NeedsPreprocessing ? "cl" : "cl-cpp-output");
    Args.push_back("-std=cl1.2");
    Args.push_back("-cl-no-stdinc");
    break;
  case AMD_COMGR_LANGUAGE_OPENCL_2_0:
    Args.push_back(NeedsPreprocessing ? "cl" : "cl-cpp-output");
    Args.push_back("-std=cl2.0");
    Args.push_back("-cl-no-stdinc");
    break;
  case AMD_COMGR_LANGUAGE_HIP:
    Args.push_back(NeedsPreprocessing ? "hip" : "hip-cpp-output");
    Args.push_back("--offload-device-only");
    // Pass a cuid that depends on the input files
    // Otherwise, a random (which depends on the /tmp/comgr-xxxxx path) cuid is
    // generated which causes a cache miss on every run.
    Args.push_back(Saver.save("-cuid=" + getStableCUID(InSet)).data());
    break;
  default:
    return AMD_COMGR_STATUS_ERROR_INVALID_ARGUMENT;
  }
  return AMD_COMGR_STATUS_SUCCESS;
}

amd_comgr_status_t AMDGPUCompiler::outputResource(llvm::StringRef Path,
                                                  llvm::StringRef FileContent) {
  // TODO: We should abstract the logic of deciding whether to use the VFS
  // or the real file system within inputFromFile and outputToFile.
  if (UseVFS) {
    if (!InMemoryFS->addFile(Path, /* ModificationTime */ 0,
                             llvm::MemoryBuffer::getMemBuffer(FileContent))) {
      return AMD_COMGR_STATUS_ERROR;
    }
  } else {
    if (auto Status = outputToFile(FileContent, Path)) {
      return Status;
    }
  }

  return AMD_COMGR_STATUS_SUCCESS;
}

amd_comgr_status_t AMDGPUCompiler::addDeviceLibraries() {
  SmallString<256> ClangBinaryPath(env::getLLVMPath());
  sys::path::append(ClangBinaryPath, "bin", "clang");

  std::string ClangResourceDir = GetResourcesPath(ClangBinaryPath);

  NoGpuLib = false;

  for (ResourceDirResource ResourceDirEntry : getResourceDirectoryFiles()) {
    llvm::SmallString<128> ResourcePath(ClangResourceDir);
    path::append(ResourcePath, ResourceDirEntry.RelativePath);

    amd_comgr_status_t Status =
        outputResource(ResourcePath, ResourceDirEntry.FileContent);
    if (Status != AMD_COMGR_STATUS_SUCCESS) {
      return Status;
    }
  }

  // TODO: This manual handling of device libs is redundant. Remove it in the
  // future when device-libs is converted to using the runtimes build to the
  // resource directory.

  SmallString<256> DeviceLibPath(ClangResourceDir);
  sys::path::append(DeviceLibPath, "lib");

  SmallString<256> DeviceCodeDir(DeviceLibPath);
  sys::path::append(DeviceCodeDir, "amdgcn", "bitcode");

  if (llvm::sys::fs::exists(DeviceCodeDir)) {
    Args.push_back(Saver.save(Twine("--rocm-path=") + DeviceLibPath).data());
  } else {
    llvm::SmallString<128> FakeRocmDir = TmpDir;
    path::append(FakeRocmDir, "rocm");
    llvm::SmallString<128> DeviceLibsDir = FakeRocmDir;
    path::append(DeviceLibsDir, "amdgcn", "bitcode");
    if (fs::create_directory(InputDir)) {
      return AMD_COMGR_STATUS_ERROR;
    }
    Args.push_back(Saver.save(Twine("--rocm-path=") + FakeRocmDir).data());

    for (auto DeviceLib : getDeviceLibraries()) {
      llvm::SmallString<128> DeviceLibPath = DeviceLibsDir;
      path::append(DeviceLibPath, std::get<0>(DeviceLib));

      amd_comgr_status_t Status =
          outputResource(DeviceLibPath, std::get<1>(DeviceLib));
      if (Status != AMD_COMGR_STATUS_SUCCESS)
        return Status;
    }
  }

  return AMD_COMGR_STATUS_SUCCESS;
}

amd_comgr_status_t AMDGPUCompiler::preprocessToSource() {
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  if (ActionInfo->IsaName) {
    if (auto Status = addTargetIdentifierFlags(ActionInfo->IsaName, true)) {
      return Status;
    }
  }

  if (auto Status = addIncludeFlags()) {
    return Status;
  }

  if (auto Status = addCompilationFlags()) {
    return Status;
  }

  Args.push_back("-E");

  return processFiles(AMD_COMGR_DATA_KIND_SOURCE, ".i");
}

amd_comgr_status_t AMDGPUCompiler::compileToBitcode(bool WithDeviceLibs) {
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  if (ActionInfo->IsaName) {
    if (auto Status = addTargetIdentifierFlags(ActionInfo->IsaName, true)) {
      return Status;
    }
  }

  if (auto Status = addIncludeFlags()) {
    return Status;
  }

  if (auto Status = addCompilationFlags()) {
    return Status;
  }

  Args.push_back("-c");
  Args.push_back("-emit-llvm");

#if _WIN32
  Args.push_back("-fshort-wchar");
#endif

  // TODO: Deprecate WithDeviceLibs in favor of ActionInfo->ShouldLinkDeviceLibs
  if (WithDeviceLibs || ActionInfo->ShouldLinkDeviceLibs) {
    if (auto Status = addDeviceLibraries()) {
      return Status;
    }

    // Currently linking postopt is only needed for OpenCL. If this becomes
    // necessary for HIP (for example if HIP adopts the same AMDGPUSimplifyLibs
    // strategy that potentially introduces undefined device-library symbols),
    // we will need also apply this option in compileToRelocatable().
    Args.push_back("-Xclang");
    Args.push_back("-mlink-builtin-bitcode-postopt");
  }

  return processFiles(AMD_COMGR_DATA_KIND_BC, ".bc");
}

amd_comgr_status_t AMDGPUCompiler::compileToExecutable() {
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  if (ActionInfo->IsaName) {
    if (auto Status = addTargetIdentifierFlags(ActionInfo->IsaName, true)) {
      return Status;
    }
  }

  if (auto Status = addIncludeFlags()) {
    return Status;
  }

  if (auto Status = addCompilationFlags()) {
    return Status;
  }

#if _WIN32
  Args.push_back("-fshort-wchar");
#endif

  // TODO: Remove "true" conditional once dependent APIs have included new
  // new *_set_device_lib_linking API
  if (ActionInfo->ShouldLinkDeviceLibs || true) {
    if (auto Status = addDeviceLibraries()) {
      return Status;
    }
  }

  return processFiles(AMD_COMGR_DATA_KIND_EXECUTABLE, ".so");
}

amd_comgr_status_t AMDGPUCompiler::compileToRelocatable() {
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  if (ActionInfo->Language != AMD_COMGR_LANGUAGE_HIP) {
    return AMD_COMGR_STATUS_ERROR_INVALID_ARGUMENT;
  }

  if (ActionInfo->IsaName) {
    if (auto Status = addTargetIdentifierFlags(ActionInfo->IsaName, true)) {
      return Status;
    }
  }

  Args.push_back("-c");
  Args.push_back("-fhip-emit-relocatable");
  Args.push_back("-mllvm");
  Args.push_back("-amdgpu-internalize-symbols");

  if (auto Status = addIncludeFlags()) {
    return Status;
  }

  if (auto Status = addCompilationFlags()) {
    return Status;
  }

#if _WIN32
  Args.push_back("-fshort-wchar");
#endif

  // TODO: Remove "true" conditional once dependent APIs have included new
  // new *_set_device_lib_linking API
  if (ActionInfo->ShouldLinkDeviceLibs || true) {
    if (auto Status = addDeviceLibraries()) {
      return Status;
    }
  }

  return processFiles(AMD_COMGR_DATA_KIND_RELOCATABLE, ".o");
}

amd_comgr_status_t AMDGPUCompiler::unbundle() {
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  // Collect bitcode memory buffers from bitcodes, bundles, and archives
  auto Cache = CommandCache::get(LogS);
  for (auto *Input : InSet->DataObjects) {

    const char *FileExtension;
    amd_comgr_data_kind_t UnbundledDataKind;
    switch (Input->DataKind) {
    case AMD_COMGR_DATA_KIND_BC_BUNDLE:
      FileExtension = "bc";
      UnbundledDataKind = AMD_COMGR_DATA_KIND_BC;
      break;
    case AMD_COMGR_DATA_KIND_AR_BUNDLE:
      FileExtension = "a";
      UnbundledDataKind = AMD_COMGR_DATA_KIND_AR;
      break;
    case AMD_COMGR_DATA_KIND_OBJ_BUNDLE:
      FileExtension = "o";
      UnbundledDataKind = AMD_COMGR_DATA_KIND_EXECUTABLE;
      break;
    default:
      return AMD_COMGR_STATUS_ERROR_INVALID_ARGUMENT;
    }

    // Configure Offload Bundler
    OffloadBundlerConfig BundlerConfig;
    BundlerConfig.AllowMissingBundles = true;
    BundlerConfig.FilesType = FileExtension;
    BundlerConfig.HipOpenmpCompatible = 1;
    BundlerConfig.AllowNoHost = 1;

    // Generate random name if none provided
    if (!strcmp(Input->Name, "")) {
      const size_t BufSize = sizeof(char) * 30;
      char *Buf = (char *)malloc(BufSize);
      snprintf(Buf, BufSize, "comgr-bundle-%d.%s", std::rand() % 10000,
               FileExtension);
      Input->Name = Buf;
    }

    // Write input file system so that OffloadBundler API can process
    // TODO: Switch write to VFS
    SmallString<128> InputFilePath = getFilePath(Input, InputDir);
    if (auto Status = outputToFile(Input, InputFilePath)) {
      return Status;
    }

    // Bundler input name
    BundlerConfig.InputFileNames.emplace_back(InputFilePath);

    // Generate prefix for output files
    StringRef OutputPrefix = Input->Name;
    size_t Index = OutputPrefix.find_last_of(".");
    OutputPrefix = OutputPrefix.substr(0, Index);

    // TODO: Log Command (see linkBitcodeToBitcode() unbundling)
    if (env::shouldEmitVerboseLogs()) {
      LogS << "   Extracting Bundle:\n"
           << "   Input Filename: " << BundlerConfig.InputFileNames[0] << "\n"
           << "   Unbundled Files Extension: ." << FileExtension << "\n";
    }

    for (StringRef Entry : ActionInfo->BundleEntryIDs) {
      // Add an output file for each target
      SmallString<128> OutputFilePath = OutputDir;
      sys::path::append(OutputFilePath,
                        OutputPrefix + "-" + Entry + "." + FileExtension);

      BundlerConfig.TargetNames.emplace_back(Entry);
      BundlerConfig.OutputFileNames.emplace_back(OutputFilePath);

      if (env::shouldEmitVerboseLogs()) {
        LogS << "\tBundle Entry ID: " << Entry << "\n"
             << "\tOutput Filename: " << OutputFilePath << "\n";
        LogS.flush();
      }
    }

    UnbundleCommand Unbundle(Input->DataKind, BundlerConfig);
    if (Cache) {
      if (auto Status = Cache->execute(Unbundle, LogS)) {
        return Status;
      }
    } else {
      if (auto Status = Unbundle.execute(LogS)) {
        return Status;
      }
    }

    // Add new bitcodes to OutSetT
    for (StringRef OutputFilePath : BundlerConfig.OutputFileNames) {

      amd_comgr_data_t ResultT;

      if (auto Status = amd_comgr_create_data(UnbundledDataKind, &ResultT))
        return Status;

      // ResultT can be released after addition to the data_set
      ScopedDataObjectReleaser SDOR(ResultT);

      DataObject *Result = DataObject::convert(ResultT);
      if (auto Status = inputFromFile(Result, OutputFilePath))
        return Status;

      StringRef OutputFileName = sys::path::filename(OutputFilePath);
      Result->setName(OutputFileName);

      if (auto Status = amd_comgr_data_set_add(OutSetT, ResultT)) {
        return Status;
      }
    }
  }

  return AMD_COMGR_STATUS_SUCCESS;
}

amd_comgr_status_t AMDGPUCompiler::linkBitcodeToBitcode() {
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  SMDiagnostic SMDiag;
  LLVMContext Context;
  Context.setDiagnosticHandler(
      std::make_unique<AMDGPUCompilerDiagnosticHandler>(this->LogS), true);

  auto Composite = std::make_unique<llvm::Module>("llvm-link", Context);
  Linker L(*Composite);
  unsigned ApplicableFlags = Linker::Flags::None;

  // Collect bitcode memory buffers from bitcodes, bundles, and archives
  for (auto *Input : InSet->DataObjects) {

    if (!strcmp(Input->Name, "")) {
      // If the calling API doesn't provide a DataObject name, generate a random
      // string to assign. This string is used when the DataObject is written
      // to the file system via SAVE_TEMPS, or if the object is a bundle which
      // also needs a file system write for unpacking
      std::string Name =
          "comgr-anon-bitcode-" + std::to_string(std::rand() % 10000) + ".bc";
      Input->setName(Name);
    }

    if (env::shouldSaveTemps()) {
      if (auto Status = outputToFile(Input, getFilePath(Input, InputDir))) {
        return Status;
      }
    }

    if (Input->DataKind == AMD_COMGR_DATA_KIND_BC) {
      if (env::shouldEmitVerboseLogs()) {
        LogS << "\t     Linking Bitcode: " << InputDir << path::get_separator() << Input->Name
             << "\n";
      }

      // The data in Input outlives Mod, and the linker destructs Mod after
      // linking it into composite (i.e. ownership is not transferred to the
      // composite) so MemoryBuffer::getMemBuffer is sufficient.
      auto Mod =
          getLazyIRModule(MemoryBuffer::getMemBuffer(
                              StringRef(Input->Data, Input->Size), "", false),
                          SMDiag, Context, true);

      if (!Mod) {
        SMDiag.print(Input->Name, LogS, /* ShowColors */ false);
        return AMD_COMGR_STATUS_ERROR;
      }
      if (verifyModule(*Mod, &LogS))
        return AMD_COMGR_STATUS_ERROR;
      if (L.linkInModule(std::move(Mod), ApplicableFlags))
        return AMD_COMGR_STATUS_ERROR;
    } else if (Input->DataKind == AMD_COMGR_DATA_KIND_BC_BUNDLE) {
      if (env::shouldEmitVerboseLogs()) {
        LogS << "      Linking Bundle: " << InputDir << path::get_separator() << Input->Name
             << "\n";
      }

      // Determine desired bundle entry ID
      // TODO: Move away from using ActionInfo->IsaName
      //   Use ActionInfo->BundleEntryIDs instead
      if (!ActionInfo->IsaName)
        return AMD_COMGR_STATUS_ERROR_INVALID_ARGUMENT;

      std::string IsaName = ActionInfo->IsaName;
      size_t Index = IsaName.find("gfx");
      std::string BundleEntryId =
          "hip-amdgcn-amd-amdhsa--gfx" + IsaName.substr(Index + 3);

      // Write data to file system so that Offload Bundler can process, assuming
      // we didn't already write due to shouldSaveTemps() conditional above
      // TODO: Switch write to VFS
      if (!env::shouldSaveTemps()) {
        if (auto Status = outputToFile(Input, getFilePath(Input, InputDir))) {
          return Status;
        }
      }

      // Configure Offload Bundler
      OffloadBundlerConfig BundlerConfig;
      BundlerConfig.AllowMissingBundles = true;
      BundlerConfig.FilesType = "bc";

      BundlerConfig.TargetNames.push_back(BundleEntryId);
      std::string InputFilePath = getFilePath(Input, InputDir).str().str();
      BundlerConfig.InputFileNames.push_back(InputFilePath);

      // Generate prefix for output files
      std::string OutputPrefix = std::string(Input->Name);
      Index = OutputPrefix.find_last_of(".");
      OutputPrefix = OutputPrefix.substr(0, Index);
      std::string OutputFileName = OutputPrefix + '-' + BundleEntryId + ".bc";

      // ISA name may contain ':', which is an invalid character in file names
      // on Windows. Replace with '_'
      std::replace(OutputFileName.begin(), OutputFileName.end(), ':', '_');

      std::string OutputFilePath = OutputDir.str().str() + path::get_separator().str() + OutputFileName;
      BundlerConfig.OutputFileNames.push_back(OutputFilePath);

      OffloadBundler Bundler(BundlerConfig);

      // Execute unbundling
      if (env::shouldEmitVerboseLogs()) {
        LogS << "Extracting Bitcode Bundle:\n"
             << "\t  Bundle Entry ID: " << BundlerConfig.TargetNames[0] << "\n"
             << "\t   Input Filename: " << BundlerConfig.InputFileNames[0]
             << "\n"
             << "\t  Output Filename: " << BundlerConfig.OutputFileNames[0]
             << "\n";
        LogS << "\t          Command: clang-offload-bundler -unbundle -type=bc"
                " -targets="
             << BundleEntryId << " -input=" << InputFilePath
             << " -output=" << OutputFilePath << "\n";
        LogS.flush();
      }

      llvm::Error Err = Bundler.UnbundleFiles();
      llvm::logAllUnhandledErrors(std::move(Err), llvm::errs(),
                                  "UnbundleFiles error: ");

      // Read unbundled bitcode from file system in order to pass to linker
      amd_comgr_data_t ResultT;
      if (auto Status = amd_comgr_create_data(AMD_COMGR_DATA_KIND_BC, &ResultT))
        return Status;

      // ResultT can be released after addition to the data_set
      ScopedDataObjectReleaser SDOR(ResultT);

      DataObject *Result = DataObject::convert(ResultT);
      if (auto Status = inputFromFile(Result, StringRef(OutputFilePath)))
        return Status;

      Result->Name = strdup(OutputFileName.c_str());

      auto Mod =
          getLazyIRModule(MemoryBuffer::getMemBuffer(
                              StringRef(Result->Data, Result->Size), "", false),
                          SMDiag, Context, true);

      if (!Mod) {
        SMDiag.print(Result->Name, LogS, /* ShowColors */ false);
        return AMD_COMGR_STATUS_ERROR;
      }
      if (verifyModule(*Mod, &LogS))
        return AMD_COMGR_STATUS_ERROR;
      if (L.linkInModule(std::move(Mod), ApplicableFlags))
        return AMD_COMGR_STATUS_ERROR;
    }
    // Unbundle bitcode archive
    else if (Input->DataKind == AMD_COMGR_DATA_KIND_AR_BUNDLE) {
      if (env::shouldEmitVerboseLogs()) {
        LogS << "\t     Linking Archive: " << InputDir << path::get_separator() << Input->Name
             << "\n";
      }

      // Determine desired bundle entry ID
      // TODO: Move away from using ActionInfo->IsaName
      //   Use ActionInfo->BundleEntryIDs instead
      if (!ActionInfo->IsaName)
        return AMD_COMGR_STATUS_ERROR_INVALID_ARGUMENT;

      std::string IsaName = ActionInfo->IsaName;
      size_t Index = IsaName.find("gfx");
      std::string BundleEntryId =
          "hip-amdgcn-amd-amdhsa--gfx" + IsaName.substr(Index + 3);

      // Write data to file system so that Offload Bundler can process, assuming
      // we didn't already write due to shouldSaveTemps() conditional above
      // TODO: Switch write to VFS
      if (!env::shouldSaveTemps()) {
        if (auto Status = outputToFile(Input, getFilePath(Input, InputDir))) {
          return Status;
        }
      }

      // Configure Offload Bundler
      OffloadBundlerConfig BundlerConfig;
      BundlerConfig.AllowMissingBundles = true;
      BundlerConfig.FilesType = "a";
      BundlerConfig.HipOpenmpCompatible = 1;
      BundlerConfig.AllowNoHost = 1;

      BundlerConfig.TargetNames.push_back(BundleEntryId);
      std::string InputFilePath = getFilePath(Input, InputDir).str().str();
      BundlerConfig.InputFileNames.push_back(InputFilePath);

      // Generate prefix for output files
      std::string OutputPrefix = std::string(Input->Name);
      Index = OutputPrefix.find_last_of(".");
      OutputPrefix = OutputPrefix.substr(0, Index);

      std::string OutputFileName = OutputPrefix + '-' + BundleEntryId + ".a";

      // ISA name may contain ':', which is an invalid character in file names
      // on Windows. Replace with '_'
      std::replace(OutputFileName.begin(), OutputFileName.end(), ':', '_');

      std::string OutputFilePath = OutputDir.str().str() + path::get_separator().str() + OutputFileName;
      BundlerConfig.OutputFileNames.push_back(OutputFilePath);

      OffloadBundler Bundler(BundlerConfig);

      // Execute unbundling
      if (env::shouldEmitVerboseLogs()) {
        LogS << "    Extracting Bitcode Archive:\n"
             << "\t  Bundle Entry ID: " << BundlerConfig.TargetNames[0] << "\n"
             << "\t   Input Filename: " << BundlerConfig.InputFileNames[0]
             << "\n"
             << "\t  Output Filename: " << BundlerConfig.OutputFileNames[0]
             << "\n";
        LogS << "\t          Command: clang-offload-bundler -unbundle -type=a "
                " -targets="
             << BundleEntryId << " -input=" << InputFilePath
             << " -output=" << OutputFilePath << "\n";
        LogS.flush();
      }
      llvm::Error Err = Bundler.UnbundleArchive();
      llvm::logAllUnhandledErrors(std::move(Err), llvm::errs(),
                                  "UnbundleArchive error: ");

      // Read archive back into Comgr
      amd_comgr_data_t ResultT;
      if (auto Status = amd_comgr_create_data(AMD_COMGR_DATA_KIND_AR, &ResultT))
        return Status;

      // ResultT can be released after addition to the data_set
      ScopedDataObjectReleaser SDOR(ResultT);

      DataObject *Result = DataObject::convert(ResultT);
      if (auto Status = inputFromFile(Result, StringRef(OutputFilePath)))
        return Status;

      // Get memory buffer for each bitcode in archive file
      //   Modeled after static loadArFile in llvm-link.cpp
      std::string ArchiveName = "comgr.ar";
      llvm::StringRef ArchiveBuf = StringRef(Result->Data, Result->Size);
      auto ArchiveOrError =
          object::Archive::create(MemoryBufferRef(ArchiveBuf, ArchiveName));

      if (!ArchiveOrError) {
        llvm::logAllUnhandledErrors(ArchiveOrError.takeError(), llvm::errs(),
                                    "Unpack Archives error: ");
        return AMD_COMGR_STATUS_ERROR;
      }

      auto Archive = std::move(ArchiveOrError.get());

      Err = Error::success();
      for (const object::Archive::Child &C : Archive->children(Err)) {

        // Get child name
        Expected<StringRef> Ename = C.getName();
        if (Error E = Ename.takeError()) {
          errs() << ": ";
          WithColor::error() << " failed to read name of archive member"
                             << ArchiveName << "'\n";
          return AMD_COMGR_STATUS_ERROR;
        }
        std::string ChildName = Ename.get().str();

        // Get memory buffer
        SMDiagnostic ParseErr;
        Expected<MemoryBufferRef> MemBuf = C.getMemoryBufferRef();
        if (Error E = MemBuf.takeError()) {
          errs() << ": ";
          WithColor::error()
              << " loading memory for member '"
              << "' of archive library failed'" << ArchiveName << "'\n";
          return AMD_COMGR_STATUS_ERROR;
        };

        // Link memory buffer into composite
        auto Mod = getLazyIRModule(MemoryBuffer::getMemBuffer(MemBuf.get()),
                                   SMDiag, Context, true);

        if (!Mod) {
          SMDiag.print(ChildName.c_str(), LogS, /* ShowColors */ false);
          return AMD_COMGR_STATUS_ERROR;
        }
        if (verifyModule(*Mod, &LogS))
          return AMD_COMGR_STATUS_ERROR;
        if (L.linkInModule(std::move(Mod), ApplicableFlags))
          return AMD_COMGR_STATUS_ERROR;
      }

      llvm::logAllUnhandledErrors(std::move(Err), llvm::errs(),
                                  "Unpack Archives error: ");
    } else
      continue;
  }

  if (verifyModule(*Composite, &LogS)) {
    return AMD_COMGR_STATUS_ERROR;
  }

  SmallString<0> OutBuf;
  BitcodeWriter Writer(OutBuf);
  Writer.writeModule(*Composite, false, nullptr, false, nullptr);
  Writer.writeSymtab();
  Writer.writeStrtab();

  amd_comgr_data_t OutputT;
  if (auto Status = amd_comgr_create_data(AMD_COMGR_DATA_KIND_BC, &OutputT)) {
    return Status;
  }

  // OutputT can be released after addition to the data_set
  ScopedDataObjectReleaser SDOR(OutputT);

  DataObject *Output = DataObject::convert(OutputT);
  Output->setName("linked.bc");
  Output->setData(OutBuf);

  return amd_comgr_data_set_add(OutSetT, OutputT);
}

amd_comgr_status_t AMDGPUCompiler::codeGenBitcodeToRelocatable() {
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  if (ActionInfo->IsaName) {
    if (auto Status = addTargetIdentifierFlags(ActionInfo->IsaName)) {
      return Status;
    }
  }

  if (ActionInfo->ShouldLinkDeviceLibs) {
    if (auto Status = addDeviceLibraries()) {
      return Status;
    }
  }

  Args.push_back("-c");

  Args.push_back("-mllvm");
  Args.push_back("-amdgpu-internalize-symbols");

  return processFiles(AMD_COMGR_DATA_KIND_RELOCATABLE, ".o");
}

amd_comgr_status_t AMDGPUCompiler::codeGenBitcodeToAssembly() {
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  if (ActionInfo->IsaName) {
    if (auto Status = addTargetIdentifierFlags(ActionInfo->IsaName)) {
      return Status;
    }
  }

  if (ActionInfo->ShouldLinkDeviceLibs) {
    if (auto Status = addDeviceLibraries()) {
      return Status;
    }
  }

  Args.push_back("-S");

  return processFiles(AMD_COMGR_DATA_KIND_SOURCE, ".s");
}

amd_comgr_status_t AMDGPUCompiler::assembleToRelocatable() {
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  if (ActionInfo->IsaName) {
    if (auto Status = addTargetIdentifierFlags(ActionInfo->IsaName)) {
      return Status;
    }
  }

  if (auto Status = addIncludeFlags()) {
    return Status;
  }

  if (ActionInfo->ShouldLinkDeviceLibs) {
    if (auto Status = addDeviceLibraries()) {
      return Status;
    }
  }

  Args.push_back("-c");
  Args.push_back("-x");
  Args.push_back("assembler");

  // -nogpulib option not needed for assembling to relocatable
  NoGpuLib = false;

  return processFiles(AMD_COMGR_DATA_KIND_RELOCATABLE, ".o");
}

amd_comgr_status_t AMDGPUCompiler::linkToRelocatable() {
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  for (auto &Option : ActionInfo->getOptions()) {
    Args.push_back(Option.c_str());
  }

  SmallVector<SmallString<128>, 128> Inputs;
  for (auto *Input : InSet->DataObjects) {
    if (Input->DataKind != AMD_COMGR_DATA_KIND_RELOCATABLE) {
      continue;
    }

    Inputs.push_back(getFilePath(Input, InputDir));
    if (auto Status = outputToFile(Input, Inputs.back())) {
      return Status;
    }
    Args.push_back(Inputs.back().c_str());
  }

  amd_comgr_data_t OutputT;
  if (auto Status =
          amd_comgr_create_data(AMD_COMGR_DATA_KIND_RELOCATABLE, &OutputT)) {
    return Status;
  }

  // OutputT can be released after addition to the data_set
  ScopedDataObjectReleaser SDOR(OutputT);

  DataObject *Output = DataObject::convert(OutputT);
  Output->setName("a.o");
  auto OutputFilePath = getFilePath(Output, OutputDir);
  Args.push_back("-o");
  Args.push_back(OutputFilePath.c_str());

  Args.push_back("-r");

  if (auto Status = linkWithLLD(Args, LogS, LogS)) {
    return Status;
  }

  if (auto Status = inputFromFile(Output, OutputFilePath)) {
    return Status;
  }

  return amd_comgr_data_set_add(OutSetT, OutputT);
}

amd_comgr_status_t AMDGPUCompiler::linkToExecutable() {
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  if (ActionInfo->IsaName) {
    if (auto Status = addTargetIdentifierFlags(ActionInfo->IsaName)) {
      return Status;
    }
  }

  for (auto &Option : ActionInfo->getOptions()) {
    Args.push_back(Option.c_str());
  }

  SmallVector<SmallString<128>, 128> Inputs;
  for (auto *Input : InSet->DataObjects) {
    if (Input->DataKind != AMD_COMGR_DATA_KIND_RELOCATABLE) {
      continue;
    }

    Inputs.push_back(getFilePath(Input, InputDir));
    if (auto Status = outputToFile(Input, Inputs.back())) {
      return Status;
    }
    Args.push_back(Inputs.back().c_str());
  }

  if (ActionInfo->ShouldLinkDeviceLibs) {
    if (auto Status = addDeviceLibraries()) {
      return Status;
    }
  }

  amd_comgr_data_t OutputT;
  if (auto Status =
          amd_comgr_create_data(AMD_COMGR_DATA_KIND_EXECUTABLE, &OutputT)) {
    return Status;
  }
  // OutputT can be released after addition to the data_set
  ScopedDataObjectReleaser SDOR(OutputT);

  DataObject *Output = DataObject::convert(OutputT);
  Output->setName("a.so");
  auto OutputFilePath = getFilePath(Output, OutputDir);
  Args.push_back("-o");
  Args.push_back(OutputFilePath.c_str());

  if (auto Status = executeInProcessDriver(Args)) {
    return Status;
  }

  if (auto Status = inputFromFile(Output, OutputFilePath)) {
    return Status;
  }

  return amd_comgr_data_set_add(OutSetT, OutputT);
}

// TODO: Generalize this list to include all -cc1 flags and arguments that are
//   still valid in a bitcode compilation context
static inline const std::unordered_set<std::string_view> ValidSpirvFlags{
    "-fapprox-func",
    "-fcolor-diagnostics",
    "-fconvergent-functions",
    "-fexceptions",
    "-ffast-math",
    "-ffinite-math-only",
    "-ffp-contract=fast",
    "-ffp-contract=fast-honor-pragmas",
    "-ffp-contract=on",
    "-fgpu-rdc",
    "-finline-functions",
    "-fno-autolink",
    "-fno-experimental-relative-c++-abi-vtables",
    "-fno-rounding-math",
    "-fno-signed-zeros",
    "-fno-threadsafe-statics",
    "-freciprocal-math",
    "-funsafe-math-optimizations",
    "-fvisibility=hidden",
    "-O0",
    "-O1",
    "-O2",
    "-O3",
    "--save-temps"};

amd_comgr_status_t AMDGPUCompiler::extractSpirvFlags(DataSet *BcSet) {

  for (auto *Bc : BcSet->DataObjects) {
    // Create SPIR-V IR Module from Bitcode Buffer
    SMDiagnostic SMDiag;
    LLVMContext Context;
    Context.setDiagnosticHandler(
        std::make_unique<AMDGPUCompilerDiagnosticHandler>(this->LogS), true);

    auto Mod = getLazyIRModule(
        MemoryBuffer::getMemBuffer(StringRef(Bc->Data, Bc->Size), "", false),
        SMDiag, Context, true);

    if (!Mod) {
      SMDiag.print("SPIR-V Bitcode", LogS, /* ShowColors */ false);
      return AMD_COMGR_STATUS_ERROR;
    }

    if (verifyModule(*Mod, &LogS))
      return AMD_COMGR_STATUS_ERROR;

    // Fetch @llvm.cmdline
    GlobalVariable *CmdLine = Mod->getNamedGlobal("llvm.cmdline");

    // Return if no @llvm.cmdline
    if (!CmdLine)
      return AMD_COMGR_STATUS_SUCCESS;

    if (ConstantDataSequential *CDS =
            dyn_cast<ConstantDataSequential>(CmdLine->getInitializer())) {

      // Add each valid null-terminated '\0' string to Flags
      std::string Tmp;
      StringRef CmdLineRaw = CDS->getRawDataValues();
      std::stringstream ss(CmdLineRaw.str());
      while (getline(ss, Tmp, '\0')) {
        if (Tmp == "--hipstdpar" || Tmp == "-amdgpu-enable-hipstdpar") {
          Bc->SpirvFlags.push_back("-mllvm");
          Bc->SpirvFlags.push_back("-amdgpu-enable-hipstdpar");
        } else if (Tmp == "-amdgpu-spill-cfi-saved-regs") {
          Bc->SpirvFlags.push_back("-mllvm");
          Bc->SpirvFlags.push_back("-amdgpu-spill-cfi-saved-regs");
        } else if (ValidSpirvFlags.count(Tmp)) {
          Bc->SpirvFlags.push_back(Saver.save(Tmp.c_str()).data());
        }
      }
    }

    if (env::shouldEmitVerboseLogs()) {
      LogS << "        SPIR-V Flags: " << Bc->Name << "\n";
      for (auto Flag : Bc->SpirvFlags)
        LogS << "          " << Flag << "\n";
    }
  }

  return AMD_COMGR_STATUS_SUCCESS;
}

amd_comgr_status_t AMDGPUCompiler::translateSpirvToBitcode() {
  return translateSpirvToBitcodeImpl(InSet, DataSet::convert(OutSetT));
}

amd_comgr_status_t
AMDGPUCompiler::translateSpirvToBitcodeImpl(DataSet *SpirvInSet,
                                            DataSet *BcOutSet) {
#ifndef COMGR_SPIRV_TRANSLATOR_AVAILABLE
#ifdef COMGR_DISABLE_SPIRV
  LogS << "Calling AMDGPUCompiler::translateSpirvToBitcodeImpl() not "
       << "supported. Comgr was built with -DCOMGR_DISABLE_SPIRV=ON.\n";
#else
  LogS << "Calling AMDGPUCompiler::translateSpirvToBitcodeImpl() not "
       << "supported. The LLVM-SPIRV-Translator was not found when Comgr "
       << "was configured.\n";
#endif
  return AMD_COMGR_STATUS_ERROR;
#else
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  // Extract GPU processor from ISA name if set, for SPIR-V feature predicate
  // resolution. TODO: Make ISA name required for this action once users have
  // migrated.
  StringRef OffloadArch;
  TargetIdentifier Ident;
  if (ActionInfo->IsaName) {
    if (auto Status = parseTargetIdentifier(ActionInfo->IsaName, Ident))
      return Status;
    OffloadArch = Ident.Processor;
  }

  auto Cache = CommandCache::get(LogS);

  for (auto *Input : SpirvInSet->DataObjects) {

    if (env::shouldSaveTemps()) {
      if (auto Status = outputToFile(Input, getFilePath(Input, InputDir))) {
        return Status;
      }
    }

    if (Input->DataKind != AMD_COMGR_DATA_KIND_SPIRV) {
      return AMD_COMGR_STATUS_ERROR_INVALID_ARGUMENT;
    }

    SmallString<0> OutBuf;
    SPIRVCommand SPIRV(Input, OutBuf, OffloadArch);

    amd_comgr_status_t Status;
    if (!Cache) {
      Status = SPIRV.execute(LogS);
    } else {
      Status = Cache->execute(SPIRV, LogS);
    }

    if (Status) {
      return Status;
    }

    amd_comgr_data_t OutputT;
    if (auto Status = amd_comgr_create_data(AMD_COMGR_DATA_KIND_BC, &OutputT)) {
      return Status;
    }

    // OutputT can be released after addition to the data_set
    ScopedDataObjectReleaser SDOR(OutputT);

    DataObject *Output = DataObject::convert(OutputT);
    Output->setName(std::string(Input->Name) + std::string(".bc"));
    Output->setData(OutBuf);

    if (auto Status =
            amd_comgr_data_set_add(DataSet::convert(BcOutSet), OutputT)) {
      return Status;
    }

    if (env::shouldEmitVerboseLogs()) {
      LogS << "SPIR-V Translation: amd-llvm-spirv -r --spirv-target-env=CL2.0 "
              "--spirv-preserve-auxdata "
           << getFilePath(Input, InputDir) << " -o "
           << getFilePath(Output, OutputDir) << " (command line equivalent)\n";
    }

    if (env::shouldSaveTemps()) {
      if (auto Status = outputToFile(Output, getFilePath(Output, OutputDir))) {
        return Status;
      }
    }
  }

  // If block sizes are specified, clone kernels for each block size
  if (!ActionInfo->BlockSizes.empty()) {
    if (auto Status = cloneKernelsInBitcode(BcOutSet)) {
      return Status;
    }
  }

  return AMD_COMGR_STATUS_SUCCESS;
#endif
}

amd_comgr_status_t AMDGPUCompiler::cloneKernelsInBitcode(DataSet *BcSet) {
  if (ActionInfo->BlockSizes.empty()) {
    // Nothing to do
    return AMD_COMGR_STATUS_SUCCESS;
  }

  if (env::shouldEmitVerboseLogs()) {
    LogS << "Cloning kernels for block sizes:";
    for (size_t BlockSize : ActionInfo->BlockSizes) {
      LogS << " " << BlockSize;
    }
    LogS << "\n";
  }

  // For each bitcode module, clone kernels for each block size
  LLVMContext Context;
  Context.setDiagnosticHandler(
      std::make_unique<AMDGPUCompilerDiagnosticHandler>(LogS), true);

  // We need to clone all bitcode modules and replace them in the set
  SmallVector<DataObject *, 8> OriginalBitcodes =
      BcSet->DataObjects.takeVector();

  for (auto *Input : OriginalBitcodes) {
    if (Input->DataKind != AMD_COMGR_DATA_KIND_BC) {
      LogS << "Unexpected input data kind for " << Input->Name << "\n";
      return AMD_COMGR_STATUS_ERROR;
    }

    // Parse the bitcode module
    SMDiagnostic Err;
    MemoryBufferRef BufferRef(StringRef(Input->Data, Input->Size), Input->Name);
    auto M = parseIR(BufferRef, Err, Context);
    if (!M) {
      LogS << "Failed to parse bitcode module: " << Input->Name << "\n";
      Err.print("comgr", LogS);
      return AMD_COMGR_STATUS_ERROR;
    }

    // Clone kernels for each block size
    SmallVector<Function *, 16> OriginalKernels;
    for (Function &F : *M) {
      // Check if this is a kernel function (SPIR-V kernels use SPIR_KERNEL
      // calling convention)
      if ((F.getCallingConv() == llvm::CallingConv::AMDGPU_KERNEL ||
           F.getCallingConv() == llvm::CallingConv::SPIR_KERNEL) &&
          !F.isDeclaration()) {
        OriginalKernels.push_back(&F);
      }
    }

    if (env::shouldEmitVerboseLogs()) {
      LogS << "Found " << OriginalKernels.size() << " kernel(s) in module "
           << Input->Name << "\n";
    }

    for (Function *OrigKernel : OriginalKernels) {
      // Determine the bounds of the original kernel.
      auto FlatWGSizeAttr =
          OrigKernel->getFnAttribute("amdgpu-flat-work-group-size");
      std::pair<size_t, size_t> OriginalBlockSizeLimits{1ul, 1024ul};
      if (FlatWGSizeAttr.isValid()) {
        StringRef Val = FlatWGSizeAttr.getValueAsString();
        std::pair<StringRef, StringRef> Sizes = Val.split(',');
        if (!Sizes.first.empty())
          Sizes.first.getAsInteger(10, OriginalBlockSizeLimits.first);
        if (!Sizes.second.empty())
          Sizes.second.getAsInteger(10, OriginalBlockSizeLimits.second);
      }

      std::string OriginalName = OrigKernel->getName().str();
      std::string BlockSizeLowerBound =
          std::to_string(OriginalBlockSizeLimits.first);

      for (size_t BlockSize : ActionInfo->BlockSizes) {
        if (BlockSize == OriginalBlockSizeLimits.second) {
          // Keep the original kernel with its original name and block size
          continue;
        }

        if (BlockSize < OriginalBlockSizeLimits.first) {
          if (env::shouldEmitVerboseLogs()) {
            LogS << "Cannot clone kernel for block size " << BlockSize
                 << " since it is smaller than the minimum block size of "
                 << OriginalBlockSizeLimits.first << "\n";
          }
          continue;
        }

        if (BlockSize > OriginalBlockSizeLimits.second) {
          if (env::shouldEmitVerboseLogs()) {
            LogS << "Cannot clone kernel for block size " << BlockSize
                 << " since it is larger than the maximum block size of "
                 << OriginalBlockSizeLimits.second << "\n";
          }
          continue;
        }

        if (env::shouldEmitVerboseLogs()) {
          LogS << "Cloning " << OrigKernel->getName()
               << " for block size: " << BlockSize << "\n";
        }

        // Create a clone of the kernel
        ValueToValueMapTy VMap;
        Function *ClonedKernel = CloneFunction(OrigKernel, VMap);

        // Ensure calling convention is preserved (should be AMDGPU_KERNEL)
        ClonedKernel->setCallingConv(OrigKernel->getCallingConv());

        // Rename the cloned kernel with block size suffix
        std::string NewName = OriginalName + ".bs" + std::to_string(BlockSize);
        ClonedKernel->setName(NewName);

        // Remove the old amdgpu-flat-work-group-size attribute and add the new
        // one
        ClonedKernel->removeFnAttr("amdgpu-flat-work-group-size");
        ClonedKernel->addFnAttr("amdgpu-flat-work-group-size",
                                BlockSizeLowerBound + "," +
                                    std::to_string(BlockSize));

        if (env::shouldEmitVerboseLogs()) {
          LogS << "  Cloned " << OrigKernel->getName() << " -> " << NewName
               << "\n";
        }
      }
    }

    // Write the modified module to a bitcode buffer
    SmallString<0> BitcodeBuffer;
    raw_svector_ostream OS(BitcodeBuffer);
    WriteBitcodeToFile(*M, OS);

    // Update the existing data object with the cloned bitcode
    Input->setData(BitcodeBuffer);

    if (env::shouldSaveTemps()) {
      if (auto Status = outputToFile(Input, getFilePath(Input, OutputDir))) {
        return Status;
      }
    }

    // Re-add the modified bitcode to the set
    BcSet->DataObjects.insert(Input);
  }

  return AMD_COMGR_STATUS_SUCCESS;
}

amd_comgr_status_t AMDGPUCompiler::compileSpirvToRelocatable() {
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  for (auto *Input : InSet->DataObjects) {
    if (Input->DataKind != AMD_COMGR_DATA_KIND_SPIRV)
      return AMD_COMGR_STATUS_ERROR_INVALID_ARGUMENT;
  }

  // Translate .spv to .bc
  amd_comgr_data_set_t TranslatedSpirvT;
  if (auto Status = amd_comgr_create_data_set(&TranslatedSpirvT))
    return Status;
  ScopedDataSetReleaser SDSR(TranslatedSpirvT);
  DataSet *TranslatedSpirv = DataSet::convert(TranslatedSpirvT);

  if (auto Status = translateSpirvToBitcodeImpl(InSet, TranslatedSpirv))
    return Status;

  // Extract relevant -cc1 flags from @llvm.cmdline
  if (auto Status = extractSpirvFlags(TranslatedSpirv))
    return Status;

  // Compile bitcode to relocatable
  if (ActionInfo->IsaName) {
    if (auto Status = addTargetIdentifierFlags(ActionInfo->IsaName)) {
      return Status;
    }
  }

  if (ActionInfo->ShouldLinkDeviceLibs) {
    if (auto Status = addDeviceLibraries()) {
      return Status;
    }
  }

  Args.push_back("-c");

  Args.push_back("-mllvm");
  Args.push_back("-amdgpu-internalize-symbols");

  return processFiles(AMD_COMGR_DATA_KIND_RELOCATABLE, ".o", TranslatedSpirv);
}

amd_comgr_status_t AMDGPUCompiler::compileSourceToSpirv() {
  if (auto Status = createTmpDirs()) {
    return Status;
  }

  if (ActionInfo->Language != AMD_COMGR_LANGUAGE_HIP) {
    return AMD_COMGR_STATUS_ERROR_INVALID_ARGUMENT;
  }

  if (auto Status = addIncludeFlags()) {
    return Status;
  }

  if (auto Status = addCompilationFlags()) {
    return Status;
  }

  // Add SPIRV-specific compilation flags
  Args.push_back("--offload-arch=amdgcnspirv");
  Args.push_back("--no-gpu-bundle-output");

#if _WIN32
  Args.push_back("-fshort-wchar");
#endif

  if (ActionInfo->ShouldLinkDeviceLibs) {
    if (auto Status = addDeviceLibraries()) {
      return Status;
    }
  }

  return processFiles(AMD_COMGR_DATA_KIND_SPIRV, ".spv");
}

AMDGPUCompiler::AMDGPUCompiler(DataAction *ActionInfo, DataSet *InSet,
                               DataSet *OutSet, raw_ostream &LogS)
    : ActionInfo(ActionInfo), InSet(InSet), OutSetT(DataSet::convert(OutSet)),
      LogS(LogS) {
  initializeCommandLineArgs(Args);

  // Initialize OverlayFS with the real file system which helps redirect
  // non-VFS reads and writes.
  OverlayFS = new vfs::OverlayFileSystem(vfs::getRealFileSystem());

  std::optional<bool> VFSStatus = env::shouldUseVFS();
  if ((VFSStatus.has_value() && *VFSStatus) ||
      (!VFSStatus.has_value() && ActionInfo->ShouldUseVFS)) {
    if (env::shouldEmitVerboseLogs()) {
      LogS << "\t File System: VFS\n";
    }
    UseVFS = true;
    InMemoryFS = new vfs::InMemoryFileSystem;
    OverlayFS->pushOverlay(InMemoryFS);
  } else {
    if (env::shouldEmitVerboseLogs()) {
      LogS << "\t File System: Real\n";
    }
  }

  // Add embedded libc++ headers to VFS for HIP compilations.
  if (getLanguage() == AMD_COMGR_LANGUAGE_HIP && hasEmbeddedLibcxxHeaders()) {
    if (!InMemoryFS) {
      InMemoryFS = new vfs::InMemoryFileSystem;
      OverlayFS->pushOverlay(InMemoryFS);
    }

    SmallString<256> ClangBinaryPath(env::getLLVMPath());
    sys::path::append(ClangBinaryPath, "bin", "clang");
    std::string ResourceDir = GetResourcesPath(ClangBinaryPath);

    // libc++ headers → <install>/include/c++/v1/<relative-path>
    SmallString<256> LibcxxBase(env::getLLVMPath());
    sys::path::append(LibcxxBase, "include", "c++", "v1");

    for (const auto &Entry : getLibcxxHeaderFiles()) {
      SmallString<128> Path(LibcxxBase);
      path::append(Path, Entry.RelativePath);
      InMemoryFS->addFile(Path, 0,
                          MemoryBuffer::getMemBuffer(Entry.FileContent, Path,
                                                     false));
    }

    // Clang builtin headers → <resource-dir>/include/<relative-path>
    SmallString<256> ClangBuiltinBase(ResourceDir);
    sys::path::append(ClangBuiltinBase, "include");

    for (const auto &Entry : getClangBuiltinHeaderFiles()) {
      SmallString<128> Path(ClangBuiltinBase);
      path::append(Path, Entry.RelativePath);
      InMemoryFS->addFile(Path, 0,
                          MemoryBuffer::getMemBuffer(Entry.FileContent, Path,
                                                     false));
    }

    HasEmbeddedHeaders = true;
    if (env::shouldEmitVerboseLogs()) {
      LogS << "\t Embedded " << getLibcxxHeaderFiles().size()
           << " libc++ headers at: " << LibcxxBase << "\n";
      LogS << "\t Embedded " << getClangBuiltinHeaderFiles().size()
           << " clang builtins at: " << ClangBuiltinBase << "\n";
    }
  }
}

AMDGPUCompiler::~AMDGPUCompiler() {
  // LLVM temps get saved in the same directory as regular temps. We can only
  // call `removeTmpDirs` if none of the env vars is enabled.
  if (!env::shouldSaveTemps() && !env::shouldSaveLLVMTemps()) {
    removeTmpDirs();
  }
}

} // namespace COMGR