llvm/tools/llvm-opt-report/OptReport.cpp

//===------------------ llvm-opt-report/OptReport.cpp ---------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
///
/// \file
/// \brief This file implements a tool that can parse the YAML optimization
/// records and generate an optimization summary annotated source listing
/// report.
///
//===----------------------------------------------------------------------===//

#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Error.h"
#include "llvm/Support/ErrorOr.h"
#include "llvm/Support/FileSystem.h"
#include "llvm/Support/Format.h"
#include "llvm/Support/LineIterator.h"
#include "llvm/Support/FileSystem.h"
#include "llvm/Support/MemoryBuffer.h"
#include "llvm/Support/Path.h"
#include "llvm/Support/Program.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Support/Signals.h"
#include "llvm/Support/YAMLTraits.h"

using namespace llvm;
using namespace llvm::yaml;

static cl::opt<bool> Help("h", cl::desc("Alias for -help"), cl::Hidden);

// Mark all our options with this category, everything else (except for -version
// and -help) will be hidden.
static cl::OptionCategory
    OptReportCategory("llvm-opt-report options");

static cl::opt<std::string>
  InputFileName(cl::Positional, cl::desc("<input>"), cl::init("-"),
                cl::cat(OptReportCategory));

static cl::opt<std::string>
  OutputFileName("o", cl::desc("Output file"), cl::init("-"),
                 cl::cat(OptReportCategory));

static cl::opt<std::string>
  InputRelDir("r", cl::desc("Root for relative input paths"), cl::init(""),
              cl::cat(OptReportCategory));

static cl::opt<bool>
  Succinct("s", cl::desc("Don't include vectorization factors, etc."),
           cl::init(false), cl::cat(OptReportCategory));

namespace {
// For each location in the source file, the common per-transformation state
// collected.
struct OptReportLocationItemInfo {
  bool Analyzed = false;
  bool Transformed = false;

  OptReportLocationItemInfo &operator |= (
    const OptReportLocationItemInfo &RHS) {
    Analyzed |= RHS.Analyzed;
    Transformed |= RHS.Transformed;

    return *this;
  }
};

// The per-location information collected for producing an optimization report.
struct OptReportLocationInfo {
  OptReportLocationItemInfo Inlined;
  OptReportLocationItemInfo Unrolled;
  OptReportLocationItemInfo Vectorized;

  int VectorizationFactor = 1;
  int InterleaveCount = 1;
  int UnrollCount = 1;

  OptReportLocationInfo &operator |= (const OptReportLocationInfo &RHS) {
    Inlined |= RHS.Inlined;
    Unrolled |= RHS.Unrolled;
    Vectorized |= RHS.Vectorized;

    VectorizationFactor =
      std::max(VectorizationFactor, RHS.VectorizationFactor);
    InterleaveCount = std::max(InterleaveCount, RHS.InterleaveCount);
    UnrollCount = std::max(UnrollCount, RHS.UnrollCount);

    return *this;
  }
};

typedef std::map<std::string, std::map<int, std::map<int,
          OptReportLocationInfo>>> LocationInfoTy;
} // anonymous namespace

static void collectLocationInfo(yaml::Stream &Stream,
                                LocationInfoTy &LocationInfo) {
  SmallVector<char, 8> Tmp;

  // Note: We're using the YAML parser here directly, instead of using the
  // YAMLTraits implementation, because the YAMLTraits implementation does not
  // support a way to handle only a subset of the input keys (it will error out
  // if there is an input key that you don't map to your class), and
  // furthermore, it does not provide a way to handle the Args sequence of
  // key/value pairs, where the order must be captured and the 'String' key
  // might be repeated.
  for (auto &Doc : Stream) {
    auto *Root = dyn_cast<yaml::MappingNode>(Doc.getRoot());
    if (!Root)
      continue;

    bool Transformed = Root->getRawTag() == "!Passed";
    std::string Pass, File;
    int Line = 0, Column = 1;

    int VectorizationFactor = 1;
    int InterleaveCount = 1;
    int UnrollCount = 1;

    for (auto &RootChild : *Root) {
      auto *Key = dyn_cast<yaml::ScalarNode>(RootChild.getKey());
      if (!Key)
        continue;
      StringRef KeyName = Key->getValue(Tmp);
      if (KeyName == "Pass") {
        auto *Value = dyn_cast<yaml::ScalarNode>(RootChild.getValue());
        if (!Value)
          continue;
        Pass = Value->getValue(Tmp);
      } else if (KeyName == "DebugLoc") {
        auto *DebugLoc = dyn_cast<yaml::MappingNode>(RootChild.getValue());
        if (!DebugLoc)
          continue;

        for (auto &DLChild : *DebugLoc) {
          auto *DLKey = dyn_cast<yaml::ScalarNode>(DLChild.getKey());
          if (!DLKey)
            continue;
          StringRef DLKeyName = DLKey->getValue(Tmp);
          if (DLKeyName == "File") {
            auto *Value = dyn_cast<yaml::ScalarNode>(DLChild.getValue());
            if (!Value)
              continue;
            File = Value->getValue(Tmp);
          } else if (DLKeyName == "Line") {
            auto *Value = dyn_cast<yaml::ScalarNode>(DLChild.getValue());
            if (!Value)
              continue;
            Value->getValue(Tmp).getAsInteger(10, Line);
          } else if (DLKeyName == "Column") {
            auto *Value = dyn_cast<yaml::ScalarNode>(DLChild.getValue());
            if (!Value)
              continue;
            Value->getValue(Tmp).getAsInteger(10, Column);
          }
        }
      } else if (KeyName == "Args") {
        auto *Args = dyn_cast<yaml::SequenceNode>(RootChild.getValue());
        if (!Args)
          continue;
        for (auto &ArgChild : *Args) {
          auto *ArgMap = dyn_cast<yaml::MappingNode>(&ArgChild);
          if (!ArgMap)
            continue;
          for (auto &ArgKV : *ArgMap) {
            auto *ArgKey = dyn_cast<yaml::ScalarNode>(ArgKV.getKey());
            if (!ArgKey)
              continue;
            StringRef ArgKeyName = ArgKey->getValue(Tmp);
            if (ArgKeyName == "VectorizationFactor") {
              auto *Value = dyn_cast<yaml::ScalarNode>(ArgKV.getValue());
              if (!Value)
                continue;
              Value->getValue(Tmp).getAsInteger(10, VectorizationFactor);
            } else if (ArgKeyName == "InterleaveCount") {
              auto *Value = dyn_cast<yaml::ScalarNode>(ArgKV.getValue());
              if (!Value)
                continue;
              Value->getValue(Tmp).getAsInteger(10, InterleaveCount);
            } else if (ArgKeyName == "UnrollCount") {
              auto *Value = dyn_cast<yaml::ScalarNode>(ArgKV.getValue());
              if (!Value)
                continue;
              Value->getValue(Tmp).getAsInteger(10, UnrollCount);
            }
          }
        }
      }
    }

    if (Line < 1 || File.empty())
      continue;

    // We track information on both actual and potential transformations. This
    // way, if there are multiple possible things on a line that are, or could
    // have been transformed, we can indicate that explicitly in the output.
    auto UpdateLLII = [Transformed, VectorizationFactor,
                       InterleaveCount,
                       UnrollCount](OptReportLocationInfo &LI,
                                    OptReportLocationItemInfo &LLII) {
      LLII.Analyzed = true;
      if (Transformed) {
        LLII.Transformed = true;

        LI.VectorizationFactor = VectorizationFactor;
        LI.InterleaveCount = InterleaveCount;
        LI.UnrollCount = UnrollCount;
      }
    };

    if (Pass == "inline") {
      auto &LI = LocationInfo[File][Line][Column];
      UpdateLLII(LI, LI.Inlined);
    } else if (Pass == "loop-unroll") {
      auto &LI = LocationInfo[File][Line][Column];
      UpdateLLII(LI, LI.Unrolled);
    } else if (Pass == "loop-vectorize") {
      auto &LI = LocationInfo[File][Line][Column];
      UpdateLLII(LI, LI.Vectorized);
    }
  }
}

static bool readLocationInfo(LocationInfoTy &LocationInfo) {
  ErrorOr<std::unique_ptr<MemoryBuffer>> Buf =
      MemoryBuffer::getFileOrSTDIN(InputFileName);
  if (std::error_code EC = Buf.getError()) {
    errs() << "error: Can't open file " << InputFileName << ": " <<
              EC.message() << "\n";
    return false;
  }

  SourceMgr SM;
  yaml::Stream Stream(Buf.get()->getBuffer(), SM);
  collectLocationInfo(Stream, LocationInfo);

  return true; 
}

static bool writeReport(LocationInfoTy &LocationInfo) {
  std::error_code EC;
  llvm::raw_fd_ostream OS(OutputFileName, EC,
              llvm::sys::fs::F_Text);
  if (EC) {
    errs() << "error: Can't open file " << OutputFileName << ": " <<
              EC.message() << "\n";
    return false;
  }

  bool FirstFile = true;
  for (auto &FI : LocationInfo) {
    SmallString<128> FileName(FI.first);
    if (!InputRelDir.empty()) {
      if (std::error_code EC = sys::fs::make_absolute(InputRelDir, FileName)) {
        errs() << "error: Can't resolve file path to " << FileName << ": " <<
                  EC.message() << "\n";
        return false;
      }
    }

    const auto &FileInfo = FI.second;

    ErrorOr<std::unique_ptr<MemoryBuffer>> Buf =
        MemoryBuffer::getFile(FileName);
    if (std::error_code EC = Buf.getError()) {
      errs() << "error: Can't open file " << FileName << ": " <<
                EC.message() << "\n";
      return false;
    }

    if (FirstFile)
      FirstFile = false;
    else
      OS << "\n";

    OS << "< " << FileName << "\n";

    // Figure out how many characters we need for the vectorization factors
    // and similar.
    OptReportLocationInfo MaxLI;
    for (auto &FI : FileInfo)
      for (auto &LI : FI.second)
        MaxLI |= LI.second;

    unsigned VFDigits = llvm::utostr(MaxLI.VectorizationFactor).size();
    unsigned ICDigits = llvm::utostr(MaxLI.InterleaveCount).size();
    unsigned UCDigits = llvm::utostr(MaxLI.UnrollCount).size();

    // Figure out how many characters we need for the line numbers.
    int64_t NumLines = 0;
    for (line_iterator LI(*Buf.get(), false); LI != line_iterator(); ++LI)
      ++NumLines;

    unsigned LNDigits = llvm::utostr(NumLines).size();

    for (line_iterator LI(*Buf.get(), false); LI != line_iterator(); ++LI) {
      int64_t L = LI.line_number();
      OptReportLocationInfo LLI;

      std::map<int, OptReportLocationInfo> ColsInfo;
      unsigned InlinedCols = 0, UnrolledCols = 0, VectorizedCols = 0;

      auto LII = FileInfo.find(L);
      if (LII != FileInfo.end()) {
        const auto &LineInfo = LII->second;

        for (auto &CI : LineInfo) {
          int Col = CI.first;
          ColsInfo[Col] = CI.second;
          InlinedCols += CI.second.Inlined.Analyzed;
          UnrolledCols += CI.second.Unrolled.Analyzed;
          VectorizedCols += CI.second.Vectorized.Analyzed;
          LLI |= CI.second;
        }
      }

      // We try to keep the output as concise as possible. If only one thing on
      // a given line could have been inlined, vectorized, etc. then we can put
      // the marker on the source line itself. If there are multiple options
      // then we want to distinguish them by placing the marker for each
      // transformation on a separate line following the source line. When we
      // do this, we use a '^' character to point to the appropriate column in
      // the source line.

      std::string USpaces(Succinct ? 0 : UCDigits, ' ');
      std::string VSpaces(Succinct ? 0 : VFDigits + ICDigits + 1, ' ');

      auto UStr = [UCDigits](OptReportLocationInfo &LLI) {
        std::string R;
        raw_string_ostream RS(R);
        if (!Succinct)
          RS << llvm::format_decimal(LLI.UnrollCount, UCDigits);
        return RS.str();
      };

      auto VStr = [VFDigits,
                   ICDigits](OptReportLocationInfo &LLI) -> std::string {
        std::string R;
        raw_string_ostream RS(R);
        if (!Succinct)
         RS << llvm::format_decimal(LLI.VectorizationFactor, VFDigits) <<
                 "," << llvm::format_decimal(LLI.InterleaveCount, ICDigits);
        return RS.str();
      };

      OS << llvm::format_decimal(L + 1, LNDigits) << " ";
      OS << (LLI.Inlined.Transformed && InlinedCols < 2 ? "I" : " ");
      OS << (LLI.Unrolled.Transformed && UnrolledCols < 2 ?
              "U" + UStr(LLI) : " " + USpaces);
      OS << (LLI.Vectorized.Transformed && VectorizedCols < 2 ?
              "V" + VStr(LLI) : " " + VSpaces);

      OS << " | " << *LI << "\n";

      for (auto &J : ColsInfo) {
        if ((J.second.Inlined.Transformed && InlinedCols > 1) ||
            (J.second.Unrolled.Transformed && UnrolledCols > 1) ||
            (J.second.Vectorized.Transformed && VectorizedCols > 1)) {
          OS << std::string(LNDigits + 1, ' ');
          OS << (J.second.Inlined.Transformed &&
                 InlinedCols > 1 ? "I" : " ");
          OS << (J.second.Unrolled.Transformed &&
                 UnrolledCols > 1 ? "U" + UStr(J.second) : " " + USpaces);
          OS << (J.second.Vectorized.Transformed &&
                 VectorizedCols > 1 ? "V" + VStr(J.second) : " " + VSpaces);

          OS << " | " << std::string(J.first - 1, ' ') << "^\n";
        }
      }
    }
  }

  return true;
}

int main(int argc, const char **argv) {
  sys::PrintStackTraceOnErrorSignal(argv[0]);

  cl::HideUnrelatedOptions(OptReportCategory);
  cl::ParseCommandLineOptions(
      argc, argv,
      "A tool to generate an optimization report from YAML optimization"
      " record files.\n");

  if (Help)
    cl::PrintHelpMessage();

  LocationInfoTy LocationInfo;
  if (!readLocationInfo(LocationInfo))
    return 1;
  if (!writeReport(LocationInfo))
    return 1; 

  return 0;
}