Remove "localize global" optimization
[oota-llvm.git] / tools / llvm-prof / llvm-prof.cpp
index f807234f82ae0ef41c7244763d923cc1a3fb51c1..6c340b89c65c2bc49f921ccffee9a2496d3abe9b 100644 (file)
 //
 //===----------------------------------------------------------------------===//
 
-#include "llvm/InstrTypes.h"
-#include "llvm/LLVMContext.h"
-#include "llvm/Module.h"
-#include "llvm/PassManager.h"
-#include "llvm/Assembly/AsmAnnotationWriter.h"
+#include "llvm/IR/LLVMContext.h"
+#include "llvm/Analysis/Passes.h"
 #include "llvm/Analysis/ProfileInfo.h"
 #include "llvm/Analysis/ProfileInfoLoader.h"
+#include "llvm/Assembly/AssemblyAnnotationWriter.h"
 #include "llvm/Bitcode/ReaderWriter.h"
+#include "llvm/IR/InstrTypes.h"
+#include "llvm/IR/Module.h"
+#include "llvm/PassManager.h"
 #include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Format.h"
+#include "llvm/Support/FormattedStream.h"
 #include "llvm/Support/ManagedStatic.h"
 #include "llvm/Support/MemoryBuffer.h"
 #include "llvm/Support/PrettyStackTrace.h"
+#include "llvm/Support/Signals.h"
 #include "llvm/Support/raw_ostream.h"
-#include "llvm/System/Signals.h"
+#include "llvm/Support/system_error.h"
 #include <algorithm>
-#include <iostream>
 #include <iomanip>
 #include <map>
 #include <set>
@@ -57,58 +60,62 @@ namespace {
 // PairSecondSort - A sorting predicate to sort by the second element of a pair.
 template<class T>
 struct PairSecondSortReverse
-  : public std::binary_function<std::pair<T, unsigned>,
-                                std::pair<T, unsigned>, bool> {
-  bool operator()(const std::pair<T, unsigned> &LHS,
-                  const std::pair<T, unsigned> &RHS) const {
+  : public std::binary_function<std::pair<T, double>,
+                                std::pair<T, double>, bool> {
+  bool operator()(const std::pair<T, double> &LHS,
+                  const std::pair<T, double> &RHS) const {
     return LHS.second > RHS.second;
   }
 };
 
+static double ignoreMissing(double w) {
+  if (w == ProfileInfo::MissingValue) return 0;
+  return w;
+}
+
 namespace {
   class ProfileAnnotator : public AssemblyAnnotationWriter {
-    std::map<const Function  *, unsigned> &FuncFreqs;
-    std::map<const BasicBlock*, unsigned> &BlockFreqs;
-    std::map<ProfileInfoLoader::Edge, unsigned> &EdgeFreqs;
+    ProfileInfo &PI;
   public:
-    ProfileAnnotator(std::map<const Function  *, unsigned> &FF,
-                     std::map<const BasicBlock*, unsigned> &BF,
-                     std::map<ProfileInfoLoader::Edge, unsigned> &EF)
-      : FuncFreqs(FF), BlockFreqs(BF), EdgeFreqs(EF) {}
-
-    virtual void emitFunctionAnnot(const Function *F, raw_ostream &OS) {
-      OS << ";;; %" << F->getName() << " called " << FuncFreqs[F]
-         << " times.\n;;;\n";
+    ProfileAnnotator(ProfileInfo &pi) : PI(pi) {}
+
+    virtual void emitFunctionAnnot(const Function *F,
+                                   formatted_raw_ostream &OS) {
+      double w = PI.getExecutionCount(F);
+      if (w != ProfileInfo::MissingValue) {
+        OS << ";;; %" << F->getName() << " called "<<(unsigned)w
+           <<" times.\n;;;\n";
+      }
     }
     virtual void emitBasicBlockStartAnnot(const BasicBlock *BB,
-                                          raw_ostream &OS) {
-      if (BlockFreqs.empty()) return;
-      std::map<const BasicBlock *, unsigned>::const_iterator I =
-        BlockFreqs.find(BB);
-      if (I != BlockFreqs.end())
-        OS << "\t;;; Basic block executed " << I->second << " times.\n";
-      else
-        OS << "\t;;; Never executed!\n";
+                                          formatted_raw_ostream &OS) {
+      double w = PI.getExecutionCount(BB);
+      if (w != ProfileInfo::MissingValue) {
+        if (w != 0) {
+          OS << "\t;;; Basic block executed " << (unsigned)w << " times.\n";
+        } else {
+          OS << "\t;;; Never executed!\n";
+        }
+      }
     }
 
-    virtual void emitBasicBlockEndAnnot(const BasicBlock *BB, raw_ostream &OS) {
-      if (EdgeFreqs.empty()) return;
-
+    virtual void emitBasicBlockEndAnnot(const BasicBlock *BB,
+                                        formatted_raw_ostream &OS) {
       // Figure out how many times each successor executed.
-      std::vector<std::pair<const BasicBlock*, unsigned> > SuccCounts;
-      const TerminatorInst *TI = BB->getTerminator();
+      std::vector<std::pair<ProfileInfo::Edge, double> > SuccCounts;
 
-      std::map<ProfileInfoLoader::Edge, unsigned>::iterator I =
-        EdgeFreqs.lower_bound(std::make_pair(const_cast<BasicBlock*>(BB), 0U));
-      for (; I != EdgeFreqs.end() && I->first.first == BB; ++I)
-        if (I->second)
-          SuccCounts.push_back(std::make_pair(TI->getSuccessor(I->first.second),
-                                              I->second));
+      const TerminatorInst *TI = BB->getTerminator();
+      for (unsigned s = 0, e = TI->getNumSuccessors(); s != e; ++s) {
+        BasicBlock* Succ = TI->getSuccessor(s);
+        double w = ignoreMissing(PI.getEdgeWeight(std::make_pair(BB, Succ)));
+        if (w != 0)
+          SuccCounts.push_back(std::make_pair(std::make_pair(BB, Succ), w));
+      }
       if (!SuccCounts.empty()) {
         OS << "\t;;; Out-edge counts:";
         for (unsigned i = 0, e = SuccCounts.size(); i != e; ++i)
-          OS << " [" << SuccCounts[i].second << " -> "
-             << SuccCounts[i].first->getName() << "]";
+          OS << " [" << (SuccCounts[i]).second << " -> "
+             << (SuccCounts[i]).first.second->getName() << "]";
         OS << "\n";
       }
     }
@@ -124,8 +131,8 @@ namespace {
     ProfileInfoLoader &PIL;
   public:
     static char ID; // Class identification, replacement for typeinfo.
-    explicit ProfileInfoPrinterPass(ProfileInfoLoader &_PIL) 
-      : ModulePass(&ID), PIL(_PIL) {}
+    explicit ProfileInfoPrinterPass(ProfileInfoLoader &_PIL)
+      : ModulePass(ID), PIL(_PIL) {}
 
     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
       AU.setPreservesAll();
@@ -139,113 +146,104 @@ namespace {
 char ProfileInfoPrinterPass::ID = 0;
 
 bool ProfileInfoPrinterPass::runOnModule(Module &M) {
-  std::map<const Function  *, unsigned> FuncFreqs;
-  std::map<const BasicBlock*, unsigned> BlockFreqs;
-  std::map<ProfileInfoLoader::Edge, unsigned> EdgeFreqs;
+  ProfileInfo &PI = getAnalysis<ProfileInfo>();
 
   // Output a report. Eventually, there will be multiple reports selectable on
   // the command line, for now, just keep things simple.
 
   // Emit the most frequent function table...
-  std::vector<std::pair<Function*, unsigned> > FunctionCounts;
-  PIL.getFunctionCounts(FunctionCounts);
-  FuncFreqs.insert(FunctionCounts.begin(), FunctionCounts.end());
+  std::vector<std::pair<Function*, double> > FunctionCounts;
+  std::vector<std::pair<BasicBlock*, double> > Counts;
+  for (Module::iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI) {
+    if (FI->isDeclaration()) continue;
+    double w = ignoreMissing(PI.getExecutionCount(FI));
+    FunctionCounts.push_back(std::make_pair(FI, w));
+    for (Function::iterator BB = FI->begin(), BBE = FI->end();
+         BB != BBE; ++BB) {
+      double w = ignoreMissing(PI.getExecutionCount(BB));
+      Counts.push_back(std::make_pair(BB, w));
+    }
+  }
 
   // Sort by the frequency, backwards.
   sort(FunctionCounts.begin(), FunctionCounts.end(),
             PairSecondSortReverse<Function*>());
 
-  uint64_t TotalExecutions = 0;
+  double TotalExecutions = 0;
   for (unsigned i = 0, e = FunctionCounts.size(); i != e; ++i)
     TotalExecutions += FunctionCounts[i].second;
 
-  std::cout << "===" << std::string(73, '-') << "===\n"
-            << "LLVM profiling output for execution";
-  if (PIL.getNumExecutions() != 1) std::cout << "s";
-  std::cout << ":\n";
+  outs() << "===" << std::string(73, '-') << "===\n"
+         << "LLVM profiling output for execution";
+  if (PIL.getNumExecutions() != 1) outs() << "s";
+  outs() << ":\n";
 
   for (unsigned i = 0, e = PIL.getNumExecutions(); i != e; ++i) {
-    std::cout << "  ";
-    if (e != 1) std::cout << i+1 << ". ";
-    std::cout << PIL.getExecution(i) << "\n";
+    outs() << "  ";
+    if (e != 1) outs() << i+1 << ". ";
+    outs() << PIL.getExecution(i) << "\n";
   }
 
-  std::cout << "\n===" << std::string(73, '-') << "===\n";
-  std::cout << "Function execution frequencies:\n\n";
+  outs() << "\n===" << std::string(73, '-') << "===\n";
+  outs() << "Function execution frequencies:\n\n";
 
   // Print out the function frequencies...
-  std::cout << " ##   Frequency\n";
+  outs() << " ##   Frequency\n";
   for (unsigned i = 0, e = FunctionCounts.size(); i != e; ++i) {
     if (FunctionCounts[i].second == 0) {
-      std::cout << "\n  NOTE: " << e-i << " function" <<
-             (e-i-1 ? "s were" : " was") << " never executed!\n";
+      outs() << "\n  NOTE: " << e-i << " function"
+        << (e-i-1 ? "s were" : " was") << " never executed!\n";
       break;
     }
 
-    std::cout << std::setw(3) << i+1 << ". " 
-      << std::setw(5) << FunctionCounts[i].second << "/"
-      << TotalExecutions << " "
-      << FunctionCounts[i].first->getName() << "\n";
+    outs() << format("%3d", i+1) << ". "
+           << format("%5.2g", FunctionCounts[i].second) << "/"
+           << format("%g", TotalExecutions) << " "
+           << FunctionCounts[i].first->getName() << "\n";
   }
 
   std::set<Function*> FunctionsToPrint;
 
-  // If we have block count information, print out the LLVM module with
-  // frequency annotations.
-  if (PIL.hasAccurateBlockCounts()) {
-    std::vector<std::pair<BasicBlock*, unsigned> > Counts;
-    PIL.getBlockCounts(Counts);
-
-    TotalExecutions = 0;
-    for (unsigned i = 0, e = Counts.size(); i != e; ++i)
-      TotalExecutions += Counts[i].second;
-
-    // Sort by the frequency, backwards.
-    sort(Counts.begin(), Counts.end(),
-              PairSecondSortReverse<BasicBlock*>());
-
-    std::cout << "\n===" << std::string(73, '-') << "===\n";
-    std::cout << "Top 20 most frequently executed basic blocks:\n\n";
-
-    // Print out the function frequencies...
-    std::cout <<" ##      %% \tFrequency\n";
-    unsigned BlocksToPrint = Counts.size();
-    if (BlocksToPrint > 20) BlocksToPrint = 20;
-    for (unsigned i = 0; i != BlocksToPrint; ++i) {
-      if (Counts[i].second == 0) break;
-      Function *F = Counts[i].first->getParent();
-      std::cout << std::setw(3) << i+1 << ". " 
-        << std::setw(5) << std::setprecision(2) 
-        << Counts[i].second/(double)TotalExecutions*100 << "% "
-        << std::setw(5) << Counts[i].second << "/"
-        << TotalExecutions << "\t"
-        << F->getName() << "() - "
-        << Counts[i].first->getName() << "\n";
-      FunctionsToPrint.insert(F);
-    }
+  TotalExecutions = 0;
+  for (unsigned i = 0, e = Counts.size(); i != e; ++i)
+    TotalExecutions += Counts[i].second;
 
-    BlockFreqs.insert(Counts.begin(), Counts.end());
-  }
+  // Sort by the frequency, backwards.
+  sort(Counts.begin(), Counts.end(),
+       PairSecondSortReverse<BasicBlock*>());
+
+  outs() << "\n===" << std::string(73, '-') << "===\n";
+  outs() << "Top 20 most frequently executed basic blocks:\n\n";
 
-  if (PIL.hasAccurateEdgeCounts()) {
-    std::vector<std::pair<ProfileInfoLoader::Edge, unsigned> > Counts;
-    PIL.getEdgeCounts(Counts);
-    EdgeFreqs.insert(Counts.begin(), Counts.end());
+  // Print out the function frequencies...
+  outs() <<" ##      %% \tFrequency\n";
+  unsigned BlocksToPrint = Counts.size();
+  if (BlocksToPrint > 20) BlocksToPrint = 20;
+  for (unsigned i = 0; i != BlocksToPrint; ++i) {
+    if (Counts[i].second == 0) break;
+    Function *F = Counts[i].first->getParent();
+    outs() << format("%3d", i+1) << ". "
+           << format("%5g", Counts[i].second/(double)TotalExecutions*100)<<"% "
+           << format("%5.0f", Counts[i].second) << "/"
+           << format("%g", TotalExecutions) << "\t"
+           << F->getName() << "() - "
+           << Counts[i].first->getName() << "\n";
+    FunctionsToPrint.insert(F);
   }
 
   if (PrintAnnotatedLLVM || PrintAllCode) {
-    std::cout << "\n===" << std::string(73, '-') << "===\n";
-    std::cout << "Annotated LLVM code for the module:\n\n";
+    outs() << "\n===" << std::string(73, '-') << "===\n";
+    outs() << "Annotated LLVM code for the module:\n\n";
 
-    ProfileAnnotator PA(FuncFreqs, BlockFreqs, EdgeFreqs);
+    ProfileAnnotator PA(PI);
 
     if (FunctionsToPrint.empty() || PrintAllCode)
-      M.print(std::cout, &PA);
+      M.print(outs(), &PA);
     else
       // Print just a subset of the functions.
       for (std::set<Function*>::iterator I = FunctionsToPrint.begin(),
              E = FunctionsToPrint.end(); I != E; ++I)
-        (*I)->print(std::cout, &PA);
+        (*I)->print(outs(), &PA);
   }
 
   return false;
@@ -258,41 +256,35 @@ int main(int argc, char **argv) {
 
   LLVMContext &Context = getGlobalContext();
   llvm_shutdown_obj Y;  // Call llvm_shutdown() on exit.
-  try {
-    cl::ParseCommandLineOptions(argc, argv, "llvm profile dump decoder\n");
-
-    // Read in the bitcode file...
-    std::string ErrorMessage;
-    Module *M = 0;
-    if (MemoryBuffer *Buffer = MemoryBuffer::getFileOrSTDIN(BitcodeFile,
-                                                            &ErrorMessage)) {
-      M = ParseBitcodeFile(Buffer, Context, &ErrorMessage);
-      delete Buffer;
-    }
-    if (M == 0) {
-      errs() << argv[0] << ": " << BitcodeFile << ": "
-        << ErrorMessage << "\n";
-      return 1;
-    }
 
-    // Read the profiling information. This is redundant since we load it again
-    // using the standard profile info provider pass, but for now this gives us
-    // access to additional information not exposed via the ProfileInfo
-    // interface.
-    ProfileInfoLoader PIL(argv[0], ProfileDataFile, *M);
-
-    // Run the printer pass.
-    PassManager PassMgr;
-    PassMgr.add(createProfileLoaderPass(ProfileDataFile));
-    PassMgr.add(new ProfileInfoPrinterPass(PIL));
-    PassMgr.run(*M);
-
-    return 0;
-  } catch (const std::string& msg) {
-    errs() << argv[0] << ": " << msg << "\n";
-  } catch (...) {
-    errs() << argv[0] << ": Unexpected unknown exception occurred.\n";
+  cl::ParseCommandLineOptions(argc, argv, "llvm profile dump decoder\n");
+
+  // Read in the bitcode file...
+  std::string ErrorMessage;
+  OwningPtr<MemoryBuffer> Buffer;
+  error_code ec;
+  Module *M = 0;
+  if (!(ec = MemoryBuffer::getFileOrSTDIN(BitcodeFile, Buffer))) {
+    M = ParseBitcodeFile(Buffer.get(), Context, &ErrorMessage);
+  } else
+    ErrorMessage = ec.message();
+  if (M == 0) {
+    errs() << argv[0] << ": " << BitcodeFile << ": "
+      << ErrorMessage << "\n";
+    return 1;
   }
-  
-  return 1;
+
+  // Read the profiling information. This is redundant since we load it again
+  // using the standard profile info provider pass, but for now this gives us
+  // access to additional information not exposed via the ProfileInfo
+  // interface.
+  ProfileInfoLoader PIL(argv[0], ProfileDataFile);
+
+  // Run the printer pass.
+  PassManager PassMgr;
+  PassMgr.add(createProfileLoaderPass(ProfileDataFile));
+  PassMgr.add(new ProfileInfoPrinterPass(PIL));
+  PassMgr.run(*M);
+
+  return 0;
 }