intel-graphics-compiler/IGC/Compiler/CISACodeGen/DebugInfo.cpp

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

Copyright (C) 2018-2021 Intel Corporation

SPDX-License-Identifier: MIT

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

#include "Probe/Assertion.h"

#include "DebugInfo/ScalarVISAModule.h"
#include "DebugInfo/DwarfDebug.hpp"
#include "DebugInfo/VISADebugInfo.hpp"
#include "Compiler/CISACodeGen/DebugInfo.hpp"
#include "llvm/IR/IntrinsicInst.h"

#include "llvmWrapper/IR/IntrinsicInst.h"

using namespace llvm;
using namespace IGC;
using namespace IGC::IGCMD;
using namespace std;
// ElfReader related typedefs
using namespace CLElfLib;

char DebugInfoPass::ID = 0;
char CatchAllLineNumber::ID = 0;

// Register pass to igc-opt
#define PASS_FLAG1 "igc-debug-finalize"
#define PASS_DESCRIPTION1 "DebugInfo pass, llvmIR part(WAs)"
#define PASS_CFG_ONLY1 false
#define PASS_ANALYSIS1 false

IGC_INITIALIZE_PASS_BEGIN(DebugInfoPass, PASS_FLAG1, PASS_DESCRIPTION1, PASS_CFG_ONLY1, PASS_ANALYSIS1)
IGC_INITIALIZE_PASS_END(DebugInfoPass, PASS_FLAG1, PASS_DESCRIPTION1, PASS_CFG_ONLY1, PASS_ANALYSIS1)

// Used for opt testing, could be removed if KernelShaderMap is moved to ctx.
static CShaderProgram::KernelShaderMap KernelShaderMap;

// Default ctor used for igc-opt testing
DebugInfoPass::DebugInfoPass() : ModulePass(ID), kernels(KernelShaderMap) {}

DebugInfoPass::DebugInfoPass(CShaderProgram::KernelShaderMap &k) : ModulePass(ID), kernels(k) {
  initializeMetaDataUtilsWrapperPass(*PassRegistry::getPassRegistry());
}

DebugInfoPass::~DebugInfoPass() {}

bool DebugInfoPass::doInitialization(llvm::Module &M) { return true; }

bool DebugInfoPass::doFinalization(llvm::Module &M) { return true; }

bool DebugInfoPass::runOnModule(llvm::Module &M) {
  // This loop is just a workaround till we add support for DIArgList metadata.
  // If we implement DIArgList support, it should be deleted.
  for (auto &F : M) {
    for (auto &BB : F) {
      for (auto &I : BB) {
        if (auto *dbgInst = dyn_cast<DbgVariableIntrinsic>(&I)) {
          if (dbgInst->getNumVariableLocationOps() > 1) {
            IGCLLVM::setKillLocation(dbgInst);
          }
        }
      }
    }
  }
  // Early out
  if (kernels.empty())
    return false;

  std::vector<CShader *> units;

  auto isCandidate = [](CShaderProgram *shaderProgram, SIMDMode m,
                        ShaderDispatchMode mode = ShaderDispatchMode::NOT_APPLICABLE) {
    auto currShader = shaderProgram->GetShader(m, mode);
    if (!currShader || !currShader->GetDebugInfoData().m_pDebugEmitter)
      return (CShader *)nullptr;

    if (currShader->ProgramOutput()->m_programSize == 0)
      return (CShader *)nullptr;

    return currShader;
  };

  for (auto &k : kernels) {
    auto shaderProgram = k.second;
    auto simd8 = isCandidate(shaderProgram, SIMDMode::SIMD8);
    auto quadSimd8Dynamic = isCandidate(shaderProgram, SIMDMode::SIMD32, ShaderDispatchMode::QUAD_SIMD8_DYNAMIC);
    auto simd16 = isCandidate(shaderProgram, SIMDMode::SIMD16);
    auto simd32 = isCandidate(shaderProgram, SIMDMode::SIMD32);

    if (simd8)
      units.push_back(simd8);
    if (simd16)
      units.push_back(simd16);
    if (simd32)
      units.push_back(simd32);
    if (quadSimd8Dynamic)
      units.push_back(quadSimd8Dynamic);
  }

  DwarfDISubprogramCache DISPCache;

  for (auto &currShader : units) {
    // Look for the right CShaderProgram instance
    m_currShader = currShader;

    MetaDataUtils *pMdUtils = m_currShader->GetMetaDataUtils();
    if (!isEntryFunc(pMdUtils, m_currShader->entry))
      continue;

    bool finalize = false;
    unsigned int size = m_currShader->GetDebugInfoData().m_VISAModules.size();
    m_pDebugEmitter = m_currShader->GetDebugInfoData().m_pDebugEmitter;
    std::vector<std::pair<unsigned int, std::pair<llvm::Function *, IGC::VISAModule *>>> sortedVISAModules;

    // Sort modules in order of their placement in binary
    IGC::VISADebugInfo VisaDbgInfo(m_currShader->ProgramOutput()->m_debugDataGenISA);
    const auto &decodedDbg = VisaDbgInfo.getRawDecodedData();
    auto getGenOff = [&decodedDbg](const std::vector<std::pair<unsigned int, unsigned int>> &data,
                                   unsigned int VISAIndex) {
      unsigned retval = 0;
      for (auto &item : data) {
        if (item.first == VISAIndex) {
          retval = item.second;
        }
      }
      return retval;
    };

    auto getLastGenOff = [this, &decodedDbg, &getGenOff](IGC::VISAModule *v) {
      unsigned int genOff = 0;
      // Detect last instructions of kernel. This information is absent in
      // dbg info. So detect is as first instruction of first subroutine - 1.
      // reloc_index, first sub inst's VISA id
      std::unordered_map<uint32_t, unsigned int> firstSubVISAIndex;

      for (auto &item : decodedDbg.compiledObjs) {
        firstSubVISAIndex[item.relocOffset] = item.CISAIndexMap.back().first;
        for (auto &sub : item.subs) {
          auto subStartVISAIndex = sub.startVISAIndex;
          if (firstSubVISAIndex[item.relocOffset] > subStartVISAIndex)
            firstSubVISAIndex[item.relocOffset] = subStartVISAIndex - 1;
        }
      }

      for (auto &item : decodedDbg.compiledObjs) {
        auto &name = item.kernelName;
        auto firstInst = (v->GetInstInfoMap()->begin())->first;
        auto funcName = firstInst->getParent()->getParent()->getName();
        if (item.subs.size() == 0 && funcName.compare(name) == 0) {
          genOff = item.CISAIndexMap.back().second;
        } else {
          if (funcName.compare(name) == 0) {
            genOff = getGenOff(item.CISAIndexMap, firstSubVISAIndex[item.relocOffset]);
            break;
          }
          for (auto &sub : item.subs) {
            auto &subName = sub.name;
            if (funcName.compare(subName) == 0) {
              genOff = getGenOff(item.CISAIndexMap, sub.endVISAIndex);
              break;
            }
          }
        }

        if (genOff)
          break;
      }

      return genOff;
    };

    auto setType = [&decodedDbg](VISAModule *v) {
      auto firstInst = (v->GetInstInfoMap()->begin())->first;
      auto funcName = firstInst->getParent()->getParent()->getName();

      for (auto &item : decodedDbg.compiledObjs) {
        auto &name = item.kernelName;
        if (funcName.compare(name) == 0) {
          if (item.relocOffset == 0)
            v->SetType(VISAModule::ObjectType::KERNEL);
          else
            v->SetType(VISAModule::ObjectType::STACKCALL_FUNC);
          return;
        }
        for (auto &sub : item.subs) {
          auto &subName = sub.name;
          if (funcName.compare(subName) == 0) {
            v->SetType(VISAModule::ObjectType::SUBROUTINE);
            return;
          }
        }
      }
    };

    for (auto &m : m_currShader->GetDebugInfoData().m_VISAModules) {
      setType(m.second);
      auto lastVISAId = getLastGenOff(m.second);
      // getLastGenOffset returns zero iff debug info for given function
      // was not found, skip the function in such case. This can happen,
      // when the function was optimized away but the definition is still
      // present inside the module.
      if (lastVISAId == 0)
        continue;
      sortedVISAModules.push_back(std::make_pair(lastVISAId, std::make_pair(m.first, m.second)));
    }

    std::sort(sortedVISAModules.begin(), sortedVISAModules.end(),
              [](std::pair<unsigned int, std::pair<llvm::Function *, IGC::VISAModule *>> &p1,
                 std::pair<unsigned int, std::pair<llvm::Function *, IGC::VISAModule *>> &p2) {
                return p1.first < p2.first;
              });

    m_pDebugEmitter->SetDISPCache(&DISPCache);
    for (auto &m : sortedVISAModules) {
      m_pDebugEmitter->registerVISA(m.second.second);
    }

    std::vector<llvm::Function *> functions;
    std::for_each(sortedVISAModules.begin(), sortedVISAModules.end(),
                  [&functions](auto &item) { functions.push_back(item.second.first); });

    for (auto &m : sortedVISAModules) {
      m_pDebugEmitter->setCurrentVISA(m.second.second);

      if (--size == 0)
        finalize = true;

      EmitDebugInfo(finalize, VisaDbgInfo);
    }

    // set VISA dbg info to nullptr to indicate 1-step debug is enabled
    if (currShader->ProgramOutput()->m_debugDataGenISA) {
      IGC::aligned_free(currShader->ProgramOutput()->m_debugDataGenISA);
    }
    currShader->ProgramOutput()->m_debugDataGenISASize = 0;
    currShader->ProgramOutput()->m_debugDataGenISA = nullptr;

    if (finalize) {
      IDebugEmitter::Release(m_pDebugEmitter);
    }
  }

  return false;
}

static void debugDump(const CShader *Shader, llvm::StringRef Ext, ArrayRef<char> Blob) {
  if (Blob.empty())
    return;

  auto ExtStr = Ext.str();
  IGC::Debug::DumpName DumpNameObj = IGC::Debug::GetDumpNameObj(Shader, ExtStr.c_str());
  std::string DumpName = DumpNameObj.str();
  if (IGC_IS_FLAG_ENABLED(DebugDumpNamePrefix)) {
    auto hash = ShaderHash();
    DumpNameObj = DumpNameObj.Hash(hash);
    DumpName = DumpNameObj.AbsolutePath(IGC_GET_REGKEYSTRING(DebugDumpNamePrefix));
  }

  if (DumpNameObj.allow()) {
    FILE *const DumpFile = fopen(DumpName.c_str(), "wb+");
    if (nullptr == DumpFile)
      return;

    fwrite(Blob.data(), Blob.size(), 1, DumpFile);
    fclose(DumpFile);
  }
}

void DebugInfoPass::EmitDebugInfo(bool finalize, const IGC::VISADebugInfo &VisaDbgInfo) {
  IGC_ASSERT(m_pDebugEmitter);

  std::vector<char> buffer = m_pDebugEmitter->Finalize(finalize, VisaDbgInfo);

  if (IGC_IS_FLAG_ENABLED(ShaderDumpEnable) || IGC_IS_FLAG_ENABLED(ElfDumpEnable))
    debugDump(m_currShader, "elf", {buffer.data(), buffer.size()});

  const std::string &DbgErrors = m_pDebugEmitter->getErrors();
  if (IGC_IS_FLAG_ENABLED(ShaderDumpEnable))
    debugDump(m_currShader, "dbgerr", {DbgErrors.data(), DbgErrors.size()});

  void *dbgInfo = IGC::aligned_malloc(buffer.size(), sizeof(void *));
  if (dbgInfo)
    memcpy_s(dbgInfo, buffer.size(), buffer.data(), buffer.size());

  SProgramOutput *pOutput = m_currShader->ProgramOutput();
  pOutput->m_debugData = dbgInfo;
  pOutput->m_debugDataSize = dbgInfo ? buffer.size() : 0;
}

// Detect instructions with an address class pattern. Then remove all opcodes of this pattern from
// this instruction's last operand (metadata of DIExpression).
// Pattern: !DIExpression(DW_OP_constu, 4, DW_OP_swap, DW_OP_xderef)
void DebugInfoData::extractAddressClass(llvm::Function &F) {
  DIBuilder di(*F.getParent());

  for (auto &bb : F) {
    for (auto &pInst : bb) {
      if (auto *DI = dyn_cast<DbgVariableIntrinsic>(&pInst)) {
        const DIExpression *DIExpr = DI->getExpression();
        llvm::SmallVector<uint64_t, 5> newElements;
        for (auto I = DIExpr->expr_op_begin(), E = DIExpr->expr_op_end(); I != E; ++I) {
          if (I->getOp() == dwarf::DW_OP_constu) {
            auto patternI = I;
            if (++patternI != E && patternI->getOp() == dwarf::DW_OP_swap && ++patternI != E &&
                patternI->getOp() == dwarf::DW_OP_xderef) {
              I = patternI;
              continue;
            }
          }
          I->appendToVector(newElements);
        }

        if (newElements.size() < DIExpr->getNumElements()) {
          DIExpression *newDIExpr = di.createExpression(newElements);
          DI->setExpression(newDIExpr);
        }
      }
    }
  }
}

void DebugInfoData::markVariableAsOutput(CShader *pShader, CVariable *pVariable) {
  IGC_ASSERT(pShader && pVariable);

  // Mark variable with "Output", to extend it's living time will be extended to the end of the function.
  pShader->GetEncoder().GetVISAKernel()->AddAttributeToVar(pVariable->visaGenVariable[0], "Output", 0, nullptr);
  if (pShader->m_State.m_dispatchSize == SIMDMode::SIMD32 && pVariable->visaGenVariable[1]) {
    pShader->GetEncoder().GetVISAKernel()->AddAttributeToVar(pVariable->visaGenVariable[1], "Output", 0, nullptr);
  }
}

void DebugInfoData::saveAndMarkPrivateMemoryVars(llvm::Function &F, CShader *pShader) {
  IGC_ASSERT_MESSAGE(pShader, "CShader is missing.");

  IDebugEmitter *pDebugEmitter = pShader->GetDebugInfoData().m_pDebugEmitter;

  ScalarVisaModule *mVISAModule = nullptr;
  if (pDebugEmitter) {
    mVISAModule = (ScalarVisaModule *)pDebugEmitter->getCurrentVISA();
    IGC_ASSERT_MESSAGE(mVISAModule, "Missing VISA module.");
  }

  // Add FP to VISA module.
  // Debug emitter will decide whether it needs to use it.
  if (mVISAModule && pShader->hasFP()) {
    mVISAModule->setFramePtr(pShader->GetFP());
  }

  for (auto &bb : F) {
    for (auto &pInst : bb) {
      // If function has "perThreadOffset" variable - add it to VISA module
      // and for mark it as output so that emitter will extend it's living time
      // to the end of the function.
      // Mark privateBase as output, because it will need it too.
      if (MDNode *perThreadOffsetMD = pInst.getMetadata("perThreadOffset")) {
        CVariable *perThreadOffset = pShader->GetSymbol(&pInst);
        CVariable *privateBase = pShader->GetPrivateBase();

        IGC_ASSERT_MESSAGE(perThreadOffset, "Missing perThreadOffset.");
        IGC_ASSERT_MESSAGE(privateBase, "Missing privateBase.");

        markVariableAsOutput(pShader, perThreadOffset);
        markVariableAsOutput(pShader, privateBase);

        if (mVISAModule) {
          mVISAModule->setPerThreadOffset(perThreadOffset);
          mVISAModule->setPrivateBase(privateBase);
        }
      }
    }
  }
}

bool IGC::DebugInfoData::hasDebugInfo(CShader *pShader) { return pShader->GetContext()->m_instrTypes.hasDebugInfo; }

void DebugInfoData::transferMappings(const llvm::Function &F) {
  auto cacheMapping = [this](llvm::DenseMap<llvm::Value *, CVariable *> &Map) {
    for (auto &mapping : Map) {
      auto CVar = mapping.second;
      if (CVar->visaGenVariable[0]) {
        unsigned int lower16Channels =
            (unsigned int)m_pShader->GetEncoder().GetVISAKernel()->getDeclarationID(CVar->visaGenVariable[0]);
        unsigned int higher16Channels = 0;
        if (numLanes(m_pShader->m_State.m_dispatchSize) == 32 && !CVar->IsUniform() && CVar->visaGenVariable[1])
          higher16Channels = m_pShader->GetEncoder().GetVISAKernel()->getDeclarationID(CVar->visaGenVariable[1]);
        CVarToVISADclId[CVar] = std::make_pair(lower16Channels, higher16Channels);
      }
    }
  };

  // Store llvm::Value->CVariable mappings from CShader.
  // CShader clears these mappings before compiling a new function.
  // Debug info is computed after all functions are compiled.
  // This instance stores mappings per llvm::Function so debug
  // info generation can emit variable locations correctly.
  auto &SymbolMapping = m_pShader->GetSymbolMapping();
  m_FunctionSymbols[&F] = SymbolMapping;

  // VISA builder gets destroyed at end of EmitVISAPass.
  // Debug info pass is invoked later. We need a way to
  // preserve mapping of CVariable -> VISA reg# so that
  // we can emit location information in debug info. This
  // code below iterates over all CVariable instances and
  // retrieves and stored their VISA reg# in a map. This
  // map is later queried by debug info pass.
  cacheMapping(SymbolMapping);

  auto &GlobalSymbolMapping = m_pShader->GetGlobalMapping();
  cacheMapping(GlobalSymbolMapping);

  if (m_pShader->hasFP()) {
    auto FP = m_pShader->GetFP();
    auto VISADclIdx = m_pShader->GetEncoder().GetVISAKernel()->getDeclarationID(FP->visaGenVariable[0]);
    CVarToVISADclId[FP] = std::make_pair(VISADclIdx, 0);
  }
}

CVariable *DebugInfoData::getMapping(const llvm::Function &F, const llvm::Value *V) {
  auto &Data = m_FunctionSymbols[&F];
  auto Iter = Data.find(V);
  if (Iter != Data.end())
    return (*Iter).second;
  return nullptr;
}

// Register pass to igc-opt
#define PASS_FLAG "igc-catch-all-linenum"
#define PASS_DESCRIPTION "CatchAllLineNumber pass"
#define PASS_CFG_ONLY false
#define PASS_ANALYSIS false
IGC_INITIALIZE_PASS_BEGIN(CatchAllLineNumber, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)
IGC_INITIALIZE_PASS_END(CatchAllLineNumber, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)

CatchAllLineNumber::CatchAllLineNumber() : FunctionPass(ID) {
  initializeMetaDataUtilsWrapperPass(*PassRegistry::getPassRegistry());
}

CatchAllLineNumber::~CatchAllLineNumber() {}

bool CatchAllLineNumber::runOnFunction(llvm::Function &F) {
  // Insert placeholder intrinsic instruction in each kernel/stack call function.
  if (!F.getSubprogram() || F.isDeclaration() || IGC_IS_FLAG_ENABLED(NoCatchAllDebugLine))
    return false;

  if (F.getCallingConv() != llvm::CallingConv::SPIR_KERNEL && !F.hasFnAttribute("visaStackCall"))
    return false;

  llvm::IRBuilder<> Builder(F.getParent()->getContext());
  DIBuilder di(*F.getParent());
  Function *lineNumPlaceholder =
      GenISAIntrinsic::getDeclaration(F.getParent(), GenISAIntrinsic::ID::GenISA_CatchAllDebugLine);
  auto intCall = Builder.CreateCall(lineNumPlaceholder);

  auto line = F.getSubprogram()->getLine();
  auto scope = F.getSubprogram();

  auto dbg = DILocation::get(F.getParent()->getContext(), line, 0, scope);

  intCall->setDebugLoc(dbg);

  intCall->insertBefore(&*F.getEntryBlock().getFirstInsertionPt());

  return true;
}