path: root/tools/bugpoint/Miscompilation.cpp

//===- Miscompilation.cpp - Debug program miscompilations -----------------===//
//
// This file implements program miscompilation debugging support.
//
//===----------------------------------------------------------------------===//

#include "BugDriver.h"
#include "SystemUtils.h"
#include "ListReducer.h"
#include "llvm/Pass.h"
#include "llvm/Module.h"
#include "llvm/Transforms/Utils/Cloning.h"
#include "llvm/Transforms/Utils/Linker.h"
#include "Support/CommandLine.h"

// Anonymous namespace to define command line options for miscompilation
// debugging.
//
namespace {
  // Output - The user can specify a file containing the expected output of the
  // program.  If this filename is set, it is used as the reference diff source,
  // otherwise the raw input run through an interpreter is used as the reference
  // source.
  //
  cl::opt<std::string> 
  Output("output", cl::desc("Specify a reference program output "
			    "(for miscompilation detection)"));
}

class ReduceMiscompilingPasses : public ListReducer<const PassInfo*> {
  BugDriver &BD;
public:
  ReduceMiscompilingPasses(BugDriver &bd) : BD(bd) {}

  virtual TestResult doTest(std::vector<const PassInfo*> &Prefix,
                            std::vector<const PassInfo*> &Kept);
};

ReduceMiscompilingPasses::TestResult
ReduceMiscompilingPasses::doTest(std::vector<const PassInfo*> &Prefix,
                                 std::vector<const PassInfo*> &Kept) {
  // First, run the program with just the Kept passes.  If it is still broken
  // with JUST the kept passes, discard the prefix passes.
  std::cout << "Checking to see if '" << getPassesString(Kept)
            << "' compile correctly: ";

  std::string BytecodeResult;
  if (BD.runPasses(Kept, BytecodeResult, false/*delete*/, true/*quiet*/)) {
    std::cerr << BD.getToolName() << ": Error running this sequence of passes"
              << " on the input program!\n";
    exit(1);
  }

  // Check to see if the finished program matches the reference output...
  if (BD.diffProgram(Output, BytecodeResult, true /*delete bytecode*/)) {
    std::cout << "nope.\n";
    return KeepSuffix;        // Miscompilation detected!
  }
  std::cout << "yup.\n";      // No miscompilation!

  if (Prefix.empty()) return NoFailure;

  // First, run the program with just the Kept passes.  If it is still broken
  // with JUST the kept passes, discard the prefix passes.
  std::cout << "Checking to see if '" << getPassesString(Prefix)
            << "' compile correctly: ";

  // If it is not broken with the kept passes, it's possible that the prefix
  // passes must be run before the kept passes to break it.  If the program
  // WORKS after the prefix passes, but then fails if running the prefix AND
  // kept passes, we can update our bytecode file to include the result of the
  // prefix passes, then discard the prefix passes.
  //
  if (BD.runPasses(Prefix, BytecodeResult, false/*delete*/, true/*quiet*/)) {
    std::cerr << BD.getToolName() << ": Error running this sequence of passes"
              << " on the input program!\n";
    exit(1);
  }

  // If the prefix maintains the predicate by itself, only keep the prefix!
  if (BD.diffProgram(Output, BytecodeResult)) {
    std::cout << "nope.\n";
    removeFile(BytecodeResult);
    return KeepPrefix;
  }
  std::cout << "yup.\n";      // No miscompilation!

  // Ok, so now we know that the prefix passes work, try running the suffix
  // passes on the result of the prefix passes.
  //
  Module *PrefixOutput = BD.ParseInputFile(BytecodeResult);
  if (PrefixOutput == 0) {
    std::cerr << BD.getToolName() << ": Error reading bytecode file '"
              << BytecodeResult << "'!\n";
    exit(1);
  }
  removeFile(BytecodeResult);  // No longer need the file on disk
    
  std::cout << "Checking to see if '" << getPassesString(Kept)
            << "' passes compile correctly after the '"
            << getPassesString(Prefix) << "' passes: ";

  Module *OriginalInput = BD.Program;
  BD.Program = PrefixOutput;
  if (BD.runPasses(Kept, BytecodeResult, false/*delete*/, true/*quiet*/)) {
    std::cerr << BD.getToolName() << ": Error running this sequence of passes"
              << " on the input program!\n";
    exit(1);
  }

  // Run the result...
  if (BD.diffProgram(Output, BytecodeResult, true/*delete bytecode*/)) {
    std::cout << "nope.\n";
    delete OriginalInput;     // We pruned down the original input...
    return KeepPrefix;
  }

  // Otherwise, we must not be running the bad pass anymore.
  std::cout << "yup.\n";      // No miscompilation!
  BD.Program = OriginalInput; // Restore original program
  delete PrefixOutput;        // Free experiment
  return NoFailure;
}

static void PrintFunctionList(const std::vector<Function*> &Funcs) {
  for (unsigned i = 0, e = Funcs.size(); i != e; ++i) {
    if (i) std::cout << ", ";
    std::cout << Funcs[i]->getName();
  }
}


class ReduceMiscompilingFunctions : public ListReducer<Function*> {
  BugDriver &BD;
public:
  ReduceMiscompilingFunctions(BugDriver &bd) : BD(bd) {}

  virtual TestResult doTest(std::vector<Function*> &Prefix,
                            std::vector<Function*> &Kept) {
    if (TestFuncs(Kept, false))
      return KeepSuffix;
    if (!Prefix.empty() && TestFuncs(Prefix, false))
      return KeepPrefix;
    return NoFailure;
  }
  
  bool TestFuncs(const std::vector<Function*> &Prefix, bool EmitBytecode);
};

bool ReduceMiscompilingFunctions::TestFuncs(const std::vector<Function*> &Funcs,
                                            bool EmitBytecode) {
  // Test to see if the function is misoptimized if we ONLY run it on the
  // functions listed in Funcs.
  if (!EmitBytecode) {
    std::cout << "Checking to see if the program is misoptimized when these "
              << "functions are run\nthrough the passes: ";
    PrintFunctionList(Funcs);
    std::cout << "\n";
  } else {
    std::cout <<"Outputting reduced bytecode files which expose the problem:\n";
  }

  // First step: clone the module for the two halves of the program we want.
  Module *ToOptimize = CloneModule(BD.Program);

  // Second step: Make sure functions & globals are all external so that linkage
  // between the two modules will work.
  for (Module::iterator I = ToOptimize->begin(), E = ToOptimize->end();I!=E;++I)
    I->setLinkage(GlobalValue::ExternalLinkage);
  for (Module::giterator I = ToOptimize->gbegin(), E = ToOptimize->gend();
       I != E; ++I)
    I->setLinkage(GlobalValue::ExternalLinkage);

  // Third step: make a clone of the externalized program for the non-optimized
  // part.
  Module *ToNotOptimize = CloneModule(ToOptimize);

  // Fourth step: Remove the test functions from the ToNotOptimize module, and
  // all of the global variables.
  for (unsigned i = 0, e = Funcs.size(); i != e; ++i) {
    Function *TNOF = ToNotOptimize->getFunction(Funcs[i]->getName(),
                                                Funcs[i]->getFunctionType());
    assert(TNOF && "Function doesn't exist in module!");
    DeleteFunctionBody(TNOF);       // Function is now external in this module!
  }
  for (Module::giterator I = ToNotOptimize->gbegin(), E = ToNotOptimize->gend();
       I != E; ++I)
    I->setInitializer(0);  // Delete the initializer to make it external

  if (EmitBytecode) {
    std::cout << "  Non-optimized portion: ";
    std::swap(BD.Program, ToNotOptimize);
    BD.EmitProgressBytecode("tonotoptimize", true);
    std::swap(BD.Program, ToNotOptimize);
  }

  // Fifth step: Remove all functions from the ToOptimize module EXCEPT for the
  // ones specified in Funcs.  We know which ones these are because they are
  // non-external in ToOptimize, but external in ToNotOptimize.
  //
  for (Module::iterator I = ToOptimize->begin(), E = ToOptimize->end();I!=E;++I)
    if (!I->isExternal()) {
      Function *TNOF = ToNotOptimize->getFunction(I->getName(),
                                                  I->getFunctionType());
      assert(TNOF && "Function doesn't exist in ToNotOptimize module??");
      if (!TNOF->isExternal())
        DeleteFunctionBody(I);
    }

  if (EmitBytecode) {
    std::cout << "  Portion that is input to optimizer: ";
    std::swap(BD.Program, ToOptimize);
    BD.EmitProgressBytecode("tooptimize");
    std::swap(BD.Program, ToOptimize);
  }

  // Sixth step: Run the optimization passes on ToOptimize, producing a
  // transformed version of the functions being tested.
  Module *OldPr