/*========================== begin_copyright_notice ============================

Copyright (C) 2018-2025 Intel Corporation

SPDX-License-Identifier: MIT

============================= end_copyright_notice ===========================*/

#include <sstream>
#include <iomanip>
#include "common/LLVMWarningsPush.hpp"
#include <llvm/Support/ScaledNumber.h>
#include <llvm/Demangle/Demangle.h>
#include <llvm/IR/DebugInfo.h>
#include "llvmWrapper/IR/LLVMContext.h"
#include "common/LLVMWarningsPop.hpp"
#include "Compiler/CISACodeGen/ShaderCodeGen.hpp"
#include "Compiler/CodeGenPublic.h"
#include "Probe/Assertion.h"
#include <llvm/IR/LLVMRemarkStreamer.h>
#include <PointersSettings.h>

namespace IGC {
struct RetryState {
  bool allowCodeScheduling;
  bool allowAddrArithCloning;
  bool allowLICM;
  bool allowCodeSinking;
  bool allowAddressArithmeticSinking;
  bool allowSimd32Slicing;
  bool allowPromotePrivateMemory;
  bool allowVISAPreRAScheduler;
  bool allowLargeURBWrite;
  bool allowConstantCoalescing;
  bool allowLargeGRF;
  bool allowLoadSinking;
  bool forceIndirectCallsInSyncRT;
  bool allowRaytracingSpillCompaction;
  unsigned nextState;
};

static const RetryState RetryTable[] = {
    // sched adrCl  licm codSk AdrSk  Slice  PrivM  VISAP  URBWr Coals  GRF loadSk, SyncRT, compactspills
    {false, false, true, true, false, false, true, true, true, true, false, false, false, true, 1},
    {true, true, false, true, true, true, false, false, false, false, true, true, true, false, 500}};

static constexpr size_t RetryTableSize = sizeof(RetryTable) / sizeof(RetryState);

RetryManager::RetryManager() : enabled(false), perKernel(false) {
  firstStateId = IGC_GET_FLAG_VALUE(RetryManagerFirstStateId);
  stateId = firstStateId;
  prevStateId = 500;
  shaderType = ShaderType::UNKNOWN;
  IGC_ASSERT(stateId < RetryTableSize);
}

bool RetryManager::AdvanceState() {
  if (!enabled || IGC_IS_FLAG_ENABLED(DisableRecompilation)) {
    return false;
  }
  IGC_ASSERT(stateId < RetryTableSize);
  prevStateId = stateId;
  stateId = RetryTable[stateId].nextState;
  return (stateId < RetryTableSize);
}

unsigned RetryManager::GetPerFuncRetryStateId(Function *F) const {
  if (IGC_GET_FLAG_VALUE(AllowStackCallRetry) == 2 && F != nullptr && prevStateId < RetryTableSize &&
      !PerFuncRetrySet.empty()) {
    std::string FName = StripCloneName(F->getName().str());
    return (PerFuncRetrySet.count(FName) != 0) ? stateId : prevStateId;
  }
  return stateId;
}

bool RetryManager::AllowLICM(Function *F) const {
  unsigned id = GetPerFuncRetryStateId(F);
  IGC_ASSERT(id < RetryTableSize);
  // Since we currently can't enable/disable LICM per-function, enabling LICM
  // when retrying for stackcalls seems to give better performance. So always
  // enable when recompiling with stackcalls.
  return RetryTable[id].allowLICM || (shaderType == ShaderType::OPENCL_SHADER && !PerFuncRetrySet.empty());
}

bool RetryManager::AllowAddressArithmeticSinking(Function *F) const {
  unsigned id = GetPerFuncRetryStateId(F);
  IGC_ASSERT(id < RetryTableSize);
  return RetryTable[id].allowAddressArithmeticSinking;
}

bool RetryManager::AllowPromotePrivateMemory(Function *F) const {
  unsigned id = GetPerFuncRetryStateId(F);
  IGC_ASSERT(id < RetryTableSize);
  return RetryTable[id].allowPromotePrivateMemory;
}

bool RetryManager::AllowVISAPreRAScheduler(Function *F) const {
  unsigned id = GetPerFuncRetryStateId(F);
  IGC_ASSERT(id < RetryTableSize);
  return RetryTable[id].allowVISAPreRAScheduler;
}

bool RetryManager::AllowCodeSinking(Function *F) const {
  unsigned id = GetPerFuncRetryStateId(F);
  IGC_ASSERT(id < RetryTableSize);
  return RetryTable[id].allowCodeSinking;
}

bool RetryManager::AllowCloneAddressArithmetic(Function *F) const {
  unsigned id = GetPerFuncRetryStateId(F);
  IGC_ASSERT(id < RetryTableSize);
  return RetryTable[id].allowAddrArithCloning;
}

bool RetryManager::AllowCodeScheduling(Function *F) const {
  unsigned id = GetPerFuncRetryStateId(F);
  IGC_ASSERT(id < RetryTableSize);
  return RetryTable[id].allowCodeScheduling;
}

bool RetryManager::AllowSimd32Slicing(Function *F) const {
  unsigned id = GetPerFuncRetryStateId(F);
  IGC_ASSERT(id < RetryTableSize);
  return RetryTable[id].allowSimd32Slicing;
}

bool RetryManager::AllowLargeURBWrite(Function *F) const {
  unsigned id = GetPerFuncRetryStateId(F);
  IGC_ASSERT(id < RetryTableSize);
  return RetryTable[id].allowLargeURBWrite;
}

bool RetryManager::AllowConstantCoalescing(Function *F) const {
  unsigned id = GetPerFuncRetryStateId(F);
  IGC_ASSERT(id < RetryTableSize);
  return RetryTable[id].allowConstantCoalescing;
}

bool RetryManager::AllowLargeGRF(Function *F) const {
  unsigned id = GetPerFuncRetryStateId(F);
  IGC_ASSERT(id < RetryTableSize);
  return RetryTable[id].allowLargeGRF;
}

bool RetryManager::ForceIndirectCallsInSyncRT() const {
  unsigned id = GetRetryId();
  IGC_ASSERT(id < RetryTableSize);
  return RetryTable[id].forceIndirectCallsInSyncRT;
}

bool RetryManager::AllowRaytracingSpillCompaction() const {
  if (IGC_IS_FLAG_ENABLED(AllowSpillCompactionOnRetry)) // allow spill compaction on retry
    return true;

  unsigned id = GetRetryId();
  IGC_ASSERT(id < RetryTableSize);
  return RetryTable[id].allowRaytracingSpillCompaction;
}

bool RetryManager::AllowLoadSinking(Function *F) const {
  unsigned id = GetPerFuncRetryStateId(F);
  IGC_ASSERT(id < RetryTableSize);
  return RetryTable[id].allowLoadSinking;
}

void RetryManager::SetFirstStateId(int id) { firstStateId = id; }

bool RetryManager::IsFirstTry() const { return (stateId == firstStateId); }

bool RetryManager::IsLastTry() const {
  return (!enabled || IGC_IS_FLAG_ENABLED(DisableRecompilation) ||
          lastSpillSize < IGC_GET_FLAG_VALUE(AllowedSpillRegCount) ||
          (stateId < RetryTableSize && RetryTable[stateId].nextState >= RetryTableSize));
}

bool RetryManager::Trigger2xGRFRetry() const {
  return (lastSpillSize > IGC_GET_FLAG_VALUE(CSSpillThreshold2xGRFRetry));
}
unsigned RetryManager::GetRetryId() const { return stateId; }

void RetryManager::Enable(ShaderType ty) {
  enabled = true;
  shaderType = ty;
}

// Disable retry
// If DisablePerKernel is true, disable retry for all kernels.
// If DisablePerKernel is false, this is no-op if per-Kernel retry is on,
//   otherwise, disable retry for all kernels.
void RetryManager::Disable(bool DisablePerKernel) {
  if (DisablePerKernel || !perKernel) {
    enabled = false;
  }
}

void RetryManager::SetSpillSize(unsigned int spillSize) { lastSpillSize = spillSize; }

unsigned int RetryManager::GetLastSpillSize() const { return lastSpillSize; }

void RetryManager::ClearSpillParams() {
  lastSpillSize = 0;
  numInstructions = 0;
}

void RetryManager::Collect(CShaderProgram::UPtr pCurrent) {
  // Get previous kernel name
  std::string funcName = pCurrent->getLLVMFunction()->getName().str();
  // Delete/unattach from memory previous version of this kernel
  // and attach new version
  previousKernels[funcName] = std::move(pCurrent);
}

CShaderProgram *RetryManager::GetPrevious(CShaderProgram *pCurrent, bool ReleaseUPtr) {
  std::string funcName = pCurrent->getLLVMFunction()->getName().str();
  if (previousKernels.find(funcName) != previousKernels.end()) {
    if (ReleaseUPtr) {
      auto ptr = previousKernels[funcName].release();
      previousKernels.erase(funcName);
      return ptr;
    } else {
      return previousKernels[funcName].get();
    }
  }
  return nullptr;
}

bool RetryManager::IsBetterThanPrevious(CShaderProgram *pCurrent, float threshold) {
  bool isBetter = true;
  auto pPrevious = GetPrevious(pCurrent);
  if (pPrevious) {
    auto simdToAnalysis = {SIMDMode::SIMD32, SIMDMode::SIMD16, SIMDMode::SIMD8};

    CShader *previousShader = pPrevious->GetShaderIfAny();
    CShader *currentShader = pCurrent->GetShaderIfAny();

    IGC_ASSERT(currentShader);
    IGC_ASSERT(previousShader);

       // basically a small work around, if we have high spilling kernel on our hands, we are not afraid to use
       // pre-retry shader, when we have less spills than retry one has
    unsigned int Threshold = IGC_GET_FLAG_VALUE(RetryRevertExcessiveSpillingKernelThreshold);
    bool IsExcessiveSpillKernel = currentShader->m_spillSize >= Threshold;
    float SpillCoefficient = float(IGC_GET_FLAG_VALUE(RetryRevertExcessiveSpillingKernelCoefficient)) / 100.0f;
    if (IsExcessiveSpillKernel)
      threshold = SpillCoefficient;

    // Check if current shader spill is larger than previous shader spill
    // Threshold flag controls comparison tolerance - i.e. A threshold of 2.0 means that the
    // current shader spill must be 2x larger than previous spill to be considered "better".
    bool spillSizeBigger = currentShader->m_spillSize > (unsigned int)(previousShader->m_spillSize * threshold);

    if (spillSizeBigger) {
      // The previous shader was better, ignore any future retry compilation
      isBetter = false;
    }
  }
  return isBetter;
}

// save entry for given SIMD mode, to avoid recompile for next retry.
void RetryManager::SaveSIMDEntry(SIMDMode simdMode, CShader *shader) {
  auto entry = GetCacheEntry(simdMode);
  IGC_ASSERT(entry);
  if (entry) {
    entry->shader = shader;
  }
}

CShader *RetryManager::GetSIMDEntry(SIMDMode simdMode) {
  auto entry = GetCacheEntry(simdMode);
  IGC_ASSERT(entry);
  return entry ? entry->shader : nullptr;
}

RetryManager::~RetryManager() {
  for (auto &it : cache) {
    if (it.shader) {
      delete it.shader;
    }
  }
}

bool RetryManager::AnyKernelSpills() const {
  return std::any_of(std::begin(cache), std::end(cache),
                     [](const CacheEntry &entry) { return entry.shader && entry.shader->m_spillCost > 0.0; });
}

bool RetryManager::PickupKernels(CodeGenContext *cgCtx) {
  {
    IGC_ASSERT_MESSAGE(0, "TODO for other shader types");
    return true;
  }
}


RetryManager::CacheEntry *RetryManager::GetCacheEntry(SIMDMode simdMode) {
  auto result = std::find_if(std::begin(cache), std::end(cache),
                             [&simdMode](const CacheEntry &entry) { return entry.simdMode == simdMode; });
  return result != std::end(cache) ? result : nullptr;
}

LLVMContextWrapper::LLVMContextWrapper(bool createResourceDimTypes) : m_UserAddrSpaceMD(this) {
  if (createResourceDimTypes) {
    CreateResourceDimensionTypes(*this);
  }
  auto *basePtr = static_cast<llvm::LLVMContext *>(this);
  // By default, LLVM 16+ operates on opaque pointers and older LLVM versions on typed pointers. The flag
  // EnableOpaquePointersBackend allows enabling opaque pointers on older LLVM versions and then any incoming
  // pointer types are dropped. The flag ForceTypedPointers enables us to force typed pointers on LLVM 16+
  // for maintenance purposes. It is not possible to ForceTypedPointers when any opaque pointers are present
  // in an LLVM IR module, then opaque pointer mode is force enabled. EnableOpaquePointersBackend does not
  // perform automatic conversion of builtin types which should be represented using TargetExtTy.
  // TODO: For transition purposes, consider introducing an IGC internal option to tweak typed/opaque pointers
  // with a precedence over the environment flag.

  // TODO: Remove/Re-evaluate once fully moved to the LLVM 16 opaque ptrs.
  // This WA_OpaquePointersCL flag is related to the same flag in LLVM itself
  // We're using it here to have consistent behaviour.
  // https://github.com/llvm/llvm-project/blob/release/16.x/llvm/lib/IR/LLVMContextImpl.cpp#L50
  auto WA_OpaquePointersCL = cl::getRegisteredOptions()["opaque-pointers"];
  if (WA_OpaquePointersCL && WA_OpaquePointersCL->getNumOccurrences() > 0) {
    IGC_IsPointerModeAlreadySet = true;
  }
  
  if (IGC::canOverwriteLLVMCtxPtrMode(basePtr, IGC_IsPointerModeAlreadySet)) {
    bool enableOpaquePointers = AreOpaquePointersEnabled();
    IGCLLVM::setOpaquePointers(basePtr, enableOpaquePointers);
    IGC_IsPointerModeAlreadySet = true;
  }
  // TODO: end
}

void LLVMContextWrapper::AddRef() { refCount++; }

void LLVMContextWrapper::Release() {
  refCount--;
  if (refCount == 0) {
    delete this;
  }
}

void CodeGenContext::print(llvm::raw_ostream &stream) const {
#define PRINT_CTX_MEMBER(member) stream << "\n" << #member << ": " << member;

  // TODO: Automate, see comment for CodeGenContext::print's declaration.
  PRINT_CTX_MEMBER(HdcEnableIndexSize);
  PRINT_CTX_MEMBER(LtoUsedMask);
  PRINT_CTX_MEMBER(m_checkFastFlagPerInstructionInCustomUnsafeOptPass);
  PRINT_CTX_MEMBER(m_ConstantBufferCount);
  PRINT_CTX_MEMBER(m_ConstantBufferReplaceShaderPatternsSize);
  PRINT_CTX_MEMBER(m_ConstantBufferReplaceSize);
  PRINT_CTX_MEMBER(m_ConstantBufferUsageMask);
  PRINT_CTX_MEMBER(m_constantPayloadNextAvailableGRFOffset);
  PRINT_CTX_MEMBER(m_disableICBPromotion);
  PRINT_CTX_MEMBER(m_dxbcCount);
  PRINT_CTX_MEMBER(m_enableFunctionPointer);
  PRINT_CTX_MEMBER(m_enableSampleMultiversioning);
  PRINT_CTX_MEMBER(m_enableSimdVariantCompilation);
  PRINT_CTX_MEMBER(m_enableSubroutine);
  PRINT_CTX_MEMBER(m_ForceEarlyZMathCheck);
  PRINT_CTX_MEMBER(m_hasDPDivSqrtEmu);
  PRINT_CTX_MEMBER(m_hasDPEmu);
  PRINT_CTX_MEMBER(m_hasEmu64BitInsts);
  PRINT_CTX_MEMBER(m_hasGlobalInPrivateAddressSpace);
  PRINT_CTX_MEMBER(m_hasLegacyDebugInfo);
  PRINT_CTX_MEMBER(m_hasStackCalls);
  PRINT_CTX_MEMBER(m_hasVendorExtension);
  PRINT_CTX_MEMBER(m_highPsRegisterPressure);
  PRINT_CTX_MEMBER(m_inputCount);
  PRINT_CTX_MEMBER(m_numGradientSinked);
  PRINT_CTX_MEMBER(m_NumGRFPerThread);
  PRINT_CTX_MEMBER(m_numPasses);
  PRINT_CTX_MEMBER(m_sampler);
  PRINT_CTX_MEMBER(m_SIMDInfo.simd8);
  PRINT_CTX_MEMBER(m_SIMDInfo.simd16);
  PRINT_CTX_MEMBER(m_SIMDInfo.simd32);
  PRINT_CTX_MEMBER(m_SIMDInfo.dual_simd8);
  PRINT_CTX_MEMBER(m_SIMDInfo.quad_simd8_dynamic);
  PRINT_CTX_MEMBER(m_src1RemovedForBlendOpt);
  PRINT_CTX_MEMBER(m_tempCount);
  PRINT_CTX_MEMBER(m_threadCombiningOptDone);

  stream << "\n";
  for (auto k : m_kernelsWithForcedRetry) {
    stream << "\nKernel with forced retry: " << k->getName().str();
  }

  stream << "\n\n";
}

void CodeGenContext::initLLVMContextWrapper(bool createResourceDimTypes) {
  llvmCtxWrapper = new LLVMContextWrapper(createResourceDimTypes);
  llvmCtxWrapper->AddRef();
}

llvm::LLVMContext *CodeGenContext::getLLVMContext() const { return llvmCtxWrapper; }

IGC::IGCMD::MetaDataUtils *CodeGenContext::getMetaDataUtils() const {
  IGC_ASSERT_MESSAGE(nullptr != m_pMdUtils, "Metadata Utils is not initialized");
  return m_pMdUtils;
}

IGCLLVM::Module *CodeGenContext::getModule() const { return module; }
std::vector<std::string> CodeGenContext::getEntryNames() const { return entry_names; }

static void initCompOptionFromRegkey(CodeGenContext *ctx) {
  SetCurrentDebugHash(ctx->hash);
  SetCurrentEntryPoints(ctx->entry_names);

  CompOptions &opt = ctx->getModuleMetaData()->compOpt;

  opt.pixelShaderDoNotAbortOnSpill = IGC_IS_FLAG_ENABLED(PixelShaderDoNotAbortOnSpill);
  opt.forcePixelShaderSIMDMode = IGC_GET_FLAG_VALUE(ForcePixelShaderSIMDMode);
}

void CodeGenContext::setModule(llvm::Module *m) {
  module = (IGCLLVM::Module *)m;
  this->setEntryNames(module);
  m_pMdUtils = new IGC::IGCMD::MetaDataUtils(m);
  modMD = new IGC::ModuleMetaData();
  initCompOptionFromRegkey(this);
}

// get the entry point names from the root entrypoint node
void CodeGenContext::setEntryNames(llvm::Module *m) {
  for (auto &func : *m) {
    if (func.isDeclaration())
      continue;

    const auto funcName = func.getName().str();
    if (!funcName.empty()) {
      this->entry_names.emplace_back(funcName);
    }
  }
}
void CodeGenContext::clearEntryNames() { this->entry_names.clear(); }
// Several clients explicitly delete module without resetting module to null.
// This causes the issue later when the dtor is invoked (trying to delete a
// dangling pointer again). This function is used to replace any explicit
// delete in order to prevent deleting dangling pointers happening.
void CodeGenContext::deleteModule() {
  delete m_pMdUtils;
  delete modMD;
  delete module;
  m_pMdUtils = nullptr;
  modMD = nullptr;
  module = nullptr;
  delete annotater;
  annotater = nullptr;
}

IGC::ModuleMetaData *CodeGenContext::getModuleMetaData() const { return modMD; }

unsigned int CodeGenContext::getRegisterPointerSizeInBits(unsigned int AS) const {
  unsigned int pointerSizeInRegister = 32;
  switch (AS) {
  case ADDRESS_SPACE_GLOBAL:
  case ADDRESS_SPACE_CONSTANT:
  case ADDRESS_SPACE_GENERIC:
  case ADDRESS_SPACE_GLOBAL_OR_PRIVATE:
    pointerSizeInRegister = getModule()->getDataLayout().getPointerSizeInBits(AS);
    break;
  case ADDRESS_SPACE_LOCAL:
  case ADDRESS_SPACE_THREAD_ARG:
    pointerSizeInRegister = 32;
    break;
  case ADDRESS_SPACE_PRIVATE:
    if (getModuleMetaData()->compOpt.UseScratchSpacePrivateMemory) {
      pointerSizeInRegister = 32;
    } else {
      pointerSizeInRegister =
          ((type == ShaderType::OPENCL_SHADER) ? getModule()->getDataLayout().getPointerSizeInBits(AS) : 64);
    }
    break;
  default:
    {
      pointerSizeInRegister = 32;
    }
    break;
  }
  return pointerSizeInRegister;
}

bool CodeGenContext::enableFunctionCall() const { return (m_enableSubroutine || m_enableFunctionPointer); }

/// Check for user functions in the module and enable the m_enableSubroutine flag if exists
void CodeGenContext::CheckEnableSubroutine(llvm::Module &M) {
  bool EnableSubroutine = false;
  bool EnableStackFuncs = false;
  for (auto &F : M) {
    if (F.isDeclaration() || F.use_empty() || isEntryFunc(getMetaDataUtils(), &F)) {
      continue;
    }

    if (F.hasFnAttribute("KMPLOCK") || F.hasFnAttribute(llvm::Attribute::NoInline) ||
        !F.hasFnAttribute(llvm::Attribute::AlwaysInline)) {
      EnableSubroutine = true;
      if (F.hasFnAttribute("visaStackCall") && !F.user_empty()) {
        EnableStackFuncs = true;
      }
    }
  }
  m_enableSubroutine = EnableSubroutine;
  m_hasStackCalls |= EnableStackFuncs;
}

// check if DP emu is required
void CodeGenContext::checkDPEmulationEnabled() {
  // TODO: the method should also check DivSqrt Emulation Mode
  if ((IGC_IS_FLAG_ENABLED(ForceDPEmulation) ||
       (m_DriverInfo.NeedFP64(platform.getPlatformInfo().eProductFamily) && platform.hasNoFP64Inst())) ||
      (getCompilerOption().FP64GenEmulationEnabled && platform.emulateFP64ForPlatformsWithoutHWSupport())) {
    m_hasDPEmu = true;
  }

  if (getCompilerOption().FP64GenConvEmulationEnabled && platform.emulateFP64ForPlatformsWithoutHWSupport()) {
    m_hasDPConvEmu = true;
  }
}

void CodeGenContext::InitVarMetaData() {}

CodeGenContext::~CodeGenContext() { clear(); }

void CodeGenContext::clear() {
  m_enableSubroutine = false;
  m_enableFunctionPointer = false;

  delete modMD;
  delete m_pMdUtils;
  modMD = nullptr;
  m_pMdUtils = nullptr;

  delete module;
  llvmCtxWrapper->Release();
  module = nullptr;
  llvmCtxWrapper = nullptr;
}

void CodeGenContext::clearMD() {
  delete modMD;
  delete m_pMdUtils;
  modMD = nullptr;
  m_pMdUtils = nullptr;
}

static const llvm::Function *getRelatedFunction(const llvm::Value *value) {
  if (value == nullptr)
    return nullptr;

  if (const llvm::Function *F = llvm::dyn_cast<llvm::Function>(value)) {
    return F;
  }
  if (const llvm::Argument *A = llvm::dyn_cast<llvm::Argument>(value)) {
    return A->getParent();
  }
  if (const llvm::BasicBlock *BB = llvm::dyn_cast<llvm::BasicBlock>(value)) {
    return BB->getParent();
  }
  if (const llvm::Instruction *I = llvm::dyn_cast<llvm::Instruction>(value)) {
    return I->getParent()->getParent();
  }

  return nullptr;
}

static bool isEntryPoint(const CodeGenContext *ctx, const llvm::Function *F) {
  if (F == nullptr) {
    return false;
  }

  auto &FuncMD = ctx->getModuleMetaData()->FuncMD;
  auto FuncInfo = FuncMD.find(const_cast<llvm::Function *>(F));
  if (FuncInfo == FuncMD.end()) {
    return false;
  }

  const FunctionMetaData *MD = &FuncInfo->second;
  return MD->functionType == KernelFunction;
}

static void findCallingKernels(const CodeGenContext *ctx, const llvm::Function *F,
                               llvm::SmallPtrSetImpl<const llvm::Function *> &kernels,
                               SmallPtrSet<const llvm::Function *, 32> &visited) {
  if (F == nullptr || kernels.count(F))
    return;

  // Check if function was already visited during search
  if (visited.find(F) != visited.end())
    return;
  visited.insert(F);

  for (const llvm::User *U : F->users()) {
    auto *CI = llvm::dyn_cast<llvm::CallInst>(U);
    if (CI == nullptr)
      continue;

    if (CI->getCalledFunction() != F)
      continue;

    const llvm::Function *caller = getRelatedFunction(CI);
    if (isEntryPoint(ctx, caller)) {
      kernels.insert(caller);
      continue;
    }
    // Caller is not a kernel, try to check which kerneles might
    // be calling it:
    findCallingKernels(ctx, caller, kernels, visited);
  }
}

static bool handleOpenMPDemangling(const std::string &name, std::string *strippedName) {
  // OpenMP mangled names have following structure:
  //
  // __omp_offloading_DD_FFFF_PP_lBB
  //
  // where DD_FFFF is an ID unique to the file (device and file IDs), PP is the
  // mangled name of the function that encloses the target region and BB is the
  // line number of the target region.
  if (name.rfind("__omp_offloading_", 0) != 0) {
    return false;
  }
  size_t offset = sizeof "__omp_offloading_";
  offset = name.find('_', offset + 1); // Find end of DD.
  if (offset == std::string::npos)
    return false;
  offset = name.find('_', offset + 1); // Find end of FFFF.
  if (offset == std::string::npos)
    return false;

  const size_t start = offset + 1;
  const size_t end = name.rfind('_'); // Find beginning of lBB.
  if (end == std::string::npos)
    return false;

  *strippedName = name.substr(start, end - start);
  return true;
}

static std::string demangleFuncName(const std::string &rawName) {
  // OpenMP adds additional prefix and suffix to the mangling scheme,
  // remove it if present.
  std::string name;
  if (!handleOpenMPDemangling(rawName, &name)) {
    // If OpenMP demangling didn't succeed just proceed with received
    // symbol name
    name = rawName;
  }
  return llvm::demangle(name);
}

void CodeGenContext::EmitMessage(std::ostream &OS, const char *messagestr, const llvm::Value *context) const {
  OS << messagestr;
  // Try to get debug location to print out the relevant info.
  if (const llvm::Instruction *I = llvm::dyn_cast_or_null<llvm::Instruction>(context)) {
    if (const llvm::DILocation *DL = I->getDebugLoc()) {
      OS << "\nin file: " << DL->getFilename().str() << ":" << DL->getLine() << "\n";
    }
  }
  // Try to find function related to given context
  // to print more informative error message.
  if (const llvm::Function *F = getRelatedFunction(context)) {
    // If the function is a kernel just print the kernel name.
    if (isEntryPoint(this, F)) {
      OS << "\nin kernel: '" << demangleFuncName(std::string(F->getName())) << "'";
      // If the function is not a kernel try to print all kernels that
      // might be using this function.
    } else {
      llvm::SmallPtrSet<const llvm::Function *, 16> kernels;
      llvm::SmallPtrSet<const llvm::Function *, 32> visited;
      findCallingKernels(this, F, kernels, visited);

      const size_t kernelsCount = kernels.size();
      OS << "\nin function: '" << demangleFuncName(std::string(F->getName())) << "' ";
      if (kernelsCount == 0) {
        OS << "called indirectly by at least one of the kernels.\n";
      } else if (kernelsCount == 1) {
        const llvm::Function *kernel = *kernels.begin();
        OS << "called by kernel: '" << demangleFuncName(std::string(kernel->getName())) << "'\n";
      } else {
        OS << "called by kernels:\n";
        for (const llvm::Function *kernel : kernels) {
          OS << "  - '" << demangleFuncName(std::string(kernel->getName())) << "'\n";
        }
      }
    }
  }
}

void CodeGenContext::EmitError(const char *errorstr, const llvm::Value *context) {
  this->oclErrorMessage << "\nerror: ";
  EmitMessage(this->oclErrorMessage, errorstr, context);
  this->oclErrorMessage << "\nerror: backend compiler failed build.\n";
}

void CodeGenContext::EmitWarning(const char *warningstr, const llvm::Value *context) {
  if (IGC_IS_FLAG_ENABLED(DisableWarnings))
    return;

  this->oclWarningMessage << "\nwarning: ";
  EmitMessage(this->oclWarningMessage, warningstr, context);
  this->oclWarningMessage << "\n";
}

CompOptions &CodeGenContext::getCompilerOption() { return getModuleMetaData()->compOpt; }

void CodeGenContext::resetOnRetry() { m_tempCount = 0; }

int32_t CodeGenContext::getNumThreadsPerEU() const { return -1; }

uint32_t CodeGenContext::getExpGRFSize() const { return 0; }

/// parameter "returnDefault" controls what to return when
/// there is no user-forced setting
uint32_t CodeGenContext::getNumGRFPerThread(bool returnDefault) {
  if (m_NumGRFPerThread)
    return m_NumGRFPerThread;

  if (IGC_GET_FLAG_VALUE(TotalGRFNum) != 0) {
    m_NumGRFPerThread = IGC_GET_FLAG_VALUE(TotalGRFNum);
    return m_NumGRFPerThread;
  }

  if (hasSyncRTCalls()) {
    // read value from CompOptions first
    DWORD GRFNum4RQToUse = getModuleMetaData()->compOpt.ForceLargeGRFNum4RQ ? 256 : 0;

    // override if reg key value is set
    GRFNum4RQToUse = IGC_IS_FLAG_ENABLED(TotalGRFNum4RQ) ? IGC_GET_FLAG_VALUE(TotalGRFNum4RQ) : GRFNum4RQToUse;
    if (GRFNum4RQToUse != 0) {
      m_NumGRFPerThread = GRFNum4RQToUse;
      return m_NumGRFPerThread;
    }
  }
  if (this->type == ShaderType::COMPUTE_SHADER && IGC_GET_FLAG_VALUE(TotalGRFNum4CS) != 0) {
    m_NumGRFPerThread = IGC_GET_FLAG_VALUE(TotalGRFNum4CS);
    return m_NumGRFPerThread;
  }
  // AIL check after reg check
  if (this->type == ShaderType::COMPUTE_SHADER && getModuleMetaData()->csInfo.forceTotalGRFNum != 0) {
    m_NumGRFPerThread = getModuleMetaData()->csInfo.forceTotalGRFNum;
    return m_NumGRFPerThread;
  }
  if (getModuleMetaData()->compOpt.forceTotalGRFNum != 0) {
    m_NumGRFPerThread = getModuleMetaData()->compOpt.forceTotalGRFNum;
    return m_NumGRFPerThread;
  }

  return (returnDefault ? DEFAULT_TOTAL_GRF_NUM : 0);
}

bool CodeGenContext::forceGlobalMemoryAllocation() const { return false; }

bool CodeGenContext::allocatePrivateAsGlobalBuffer() const { return false; }

bool CodeGenContext::noLocalToGenericOptionEnabled() const { return false; }

bool CodeGenContext::mustDistinguishBetweenPrivateAndGlobalPtr() const { return false; }

bool CodeGenContext::enableTakeGlobalAddress() const { return false; }

int16_t CodeGenContext::getVectorCoalescingControl() const { return 0; }

uint32_t CodeGenContext::getPrivateMemoryMinimalSizePerThread() const { return 0; }

uint32_t CodeGenContext::getIntelScratchSpacePrivateMemoryMinimalSizePerThread() const { return 0; }

bool CodeGenContext::isPOSH() const {
  return this->getModule()->getModuleFlag("IGC::PositionOnlyVertexShader") != nullptr;
}

bool CodeGenContext::isSWSubTriangleOpacityCullingEmulationEnabled() const {
  // STOC level emulation is enabled only if:
  // 1. Platform is Xe3.
  // 2. 64bit RayTracing structures are used.
  // 3. STOC level emulation is requested by UMD.
  // 4. STOC level emulation routine binary is delivered by the UMD.
  return platform.isSWSubTriangleOpacityCullingEmulationEnabled() && bvhInfo.uses64Bit &&
         m_enableSubTriangleOpacityEmulation && (m_numBifModules != 0) && (m_bifModules != nullptr);
}


bool CodeGenContext::isBufferBoundsChecking() const { return false; }

void CodeGenContext::setFlagsPerCtx() {}

// Returns the SIMD mode of a kernel based on a platform and settings flags.
// VS, DS, HS, GS currently supported only.
SIMDMode CodeGenContext::GetSIMDMode() const {
  SIMDMode simdMode = SIMDMode::UNKNOWN;

  bool isGeomFF = (type == ShaderType::VERTEX_SHADER) || (type == ShaderType::HULL_SHADER) ||
                  (type == ShaderType::DOMAIN_SHADER) || (type == ShaderType::GEOMETRY_SHADER);

  if (isGeomFF) {
    // Step 1: get platform-dependent default mode.
    {
      simdMode = platform.getMinDispatchMode();
    }
  } else {
    IGC_ASSERT_MESSAGE(0, "Incorrect shader type");
  }

  IGC_ASSERT_MESSAGE(!hasSyncRTCalls() || (simdMode <= platform.getPreferredRayQuerySIMDSize(type)),
                     "Unsupported SIMD mode for RayQuery");

  return simdMode;
}

uint64_t CodeGenContext::getMinimumValidAddress() const { return 0; }

// [used by shader dump] create unqiue id, starting from 1, for each
// entry function.
// Each entry function has 1-1 map b/w its name and its dump name.
void CodeGenContext::createFunctionIDs() {
  if (IGC_IS_FLAG_ENABLED(DumpUseShorterName) && m_functionIDs.empty() && module != nullptr && m_pMdUtils != nullptr) {
    int id = 0;
    for (auto i = m_pMdUtils->begin_FunctionsInfo(), e = m_pMdUtils->end_FunctionsInfo(); i != e; ++i) {
      Function *pFunc = i->first;
      // Skip non-entry functions.
      if (!isEntryFunc(m_pMdUtils, pFunc)) {
        continue;
      }

      // Use kernel name so that it is created once and will work for both the first and retry.
      StringRef kernelName = pFunc->getName();
      if (!kernelName.empty()) {
        // entry without name will not be in the map (getFunctionID() will
        // return 0. Thus, ID here starts from 1.
        m_functionIDs[kernelName.str()] = ++id;
      }
    }
    m_enableDumpUseShorterName = true;
  }
}

// getFunctionDumpName() is invoked if DumpUseShorterName is set.
// Expect to be used for any shader (aka entry function).
std::string CodeGenContext::getFunctionDumpName(int functionId) {
  IGC_ASSERT(IGC_IS_FLAG_ENABLED(DumpUseShorterName));
  std::stringstream ss;
  ss << "entry_" << std::setfill('0') << std::setw(4) << functionId;
  return ss.str();
}

int CodeGenContext::getFunctionID(Function *F) {
  StringRef kernelName = F->getName();
  auto MI = m_functionIDs.find(kernelName.str());
  if (MI == m_functionIDs.end()) {
    return 0;
  }
  return MI->second;
}
void CodeGenContext::initializeRemarkEmitter(const ShaderHash &hash) {
  // setting up optimization remark emitter
  if (IGC_IS_FLAG_ENABLED(EnableRemarks)) {
    std::string remark_file_name = IGC::Debug::DumpName("Remark_").Type(this->type).Hash(hash).Extension("yaml").str();
    llvm::Expected<std::unique_ptr<llvm::ToolOutputFile>> RemarksFileOrErr =
        setupLLVMOptimizationRemarks(*this->getLLVMContext(), remark_file_name, "", "yaml", false, 0);
    this->RemarksFile = std::move(*RemarksFileOrErr);
    this->RemarksFile->keep();
  }
}
} // namespace IGC