C++ global_iterator::getType方法代码示例

/// computeTypeMapping - Loop over all of the linked values to compute type
/// mappings.  For example, if we link "extern Foo *x" and "Foo *x = NULL", then
/// we have two struct types 'Foo' but one got renamed when the module was
/// loaded into the same LLVMContext.
void ModuleLinker::computeTypeMapping() {
  // Incorporate globals.
  for (Module::global_iterator I = SrcM->global_begin(),
       E = SrcM->global_end(); I != E; ++I) {
    GlobalValue *DGV = getLinkedToGlobal(I);
    if (DGV == 0) continue;
    
    if (!DGV->hasAppendingLinkage() || !I->hasAppendingLinkage()) {
      TypeMap.addTypeMapping(DGV->getType(), I->getType());
      continue;      
    }
    
    // Unify the element type of appending arrays.
    ArrayType *DAT = cast<ArrayType>(DGV->getType()->getElementType());
    ArrayType *SAT = cast<ArrayType>(I->getType()->getElementType());
    TypeMap.addTypeMapping(DAT->getElementType(), SAT->getElementType());
  }
  
  // Incorporate functions.
  for (Module::iterator I = SrcM->begin(), E = SrcM->end(); I != E; ++I) {
    if (GlobalValue *DGV = getLinkedToGlobal(I))
      TypeMap.addTypeMapping(DGV->getType(), I->getType());
  }

  // Incorporate types by name, scanning all the types in the source module.
  // At this point, the destination module may have a type "%foo = { i32 }" for
  // example.  When the source module got loaded into the same LLVMContext, if
  // it had the same type, it would have been renamed to "%foo.42 = { i32 }".
  // Though it isn't required for correctness, attempt to link these up to clean
  // up the IR.
  std::vector<StructType*> SrcStructTypes;
  SrcM->findUsedStructTypes(SrcStructTypes);
  
  SmallPtrSet<StructType*, 32> SrcStructTypesSet(SrcStructTypes.begin(),
                                                 SrcStructTypes.end());
  
  for (unsigned i = 0, e = SrcStructTypes.size(); i != e; ++i) {
    StructType *ST = SrcStructTypes[i];
    if (!ST->hasName()) continue;
    
    // Check to see if there is a dot in the name followed by a digit.
    size_t DotPos = ST->getName().rfind('.');
    if (DotPos == 0 || DotPos == StringRef::npos ||
        ST->getName().back() == '.' || !isdigit(ST->getName()[DotPos+1]))
      continue;
    
    // Check to see if the destination module has a struct with the prefix name.
    if (StructType *DST = DstM->getTypeByName(ST->getName().substr(0, DotPos)))
      // Don't use it if this actually came from the source module.  They're in
      // the same LLVMContext after all.
      if (!SrcStructTypesSet.count(DST))
        TypeMap.addTypeMapping(DST, ST);
  }

  // Don't bother incorporating aliases, they aren't generally typed well.
  
  // Now that we have discovered all of the type equivalences, get a body for
  // any 'opaque' types in the dest module that are now resolved. 
  TypeMap.linkDefinedTypeBodies();
}

bool GlobalMerge::doInitialization(Module &M) {
  DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
                                                        BSSGlobals;
  const DataLayout *TD = TLI->getDataLayout();
  unsigned MaxOffset = TLI->getMaximalGlobalOffset();
  bool Changed = false;

  // Grab all non-const globals.
  for (Module::global_iterator I = M.global_begin(),
         E = M.global_end(); I != E; ++I) {
    // Merge is safe for "normal" internal globals only
    if (!I->hasLocalLinkage() || I->isThreadLocal() || I->hasSection())
      continue;

    PointerType *PT = dyn_cast<PointerType>(I->getType());
    assert(PT && "Global variable is not a pointer!");

    unsigned AddressSpace = PT->getAddressSpace();

    // Ignore fancy-aligned globals for now.
    unsigned Alignment = TD->getPreferredAlignment(I);
    Type *Ty = I->getType()->getElementType();
    if (Alignment > TD->getABITypeAlignment(Ty))
      continue;

    // Ignore all 'special' globals.
    if (I->getName().startswith("llvm.") ||
        I->getName().startswith(".llvm."))
      continue;

    if (TD->getTypeAllocSize(Ty) < MaxOffset) {
      if (TargetLoweringObjectFile::getKindForGlobal(I, TLI->getTargetMachine())
          .isBSSLocal())
        BSSGlobals[AddressSpace].push_back(I);
      else if (I->isConstant())
        ConstGlobals[AddressSpace].push_back(I);
      else
        Globals[AddressSpace].push_back(I);
    }
  }

  for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
       I = Globals.begin(), E = Globals.end(); I != E; ++I)
    if (I->second.size() > 1)
      Changed |= doMerge(I->second, M, false, I->first);

  for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
       I = BSSGlobals.begin(), E = BSSGlobals.end(); I != E; ++I)
    if (I->second.size() > 1)
      Changed |= doMerge(I->second, M, false, I->first);

  // FIXME: This currently breaks the EH processing due to way how the
  // typeinfo detection works. We might want to detect the TIs and ignore
  // them in the future.
  // if (ConstGlobals.size() > 1)
  //  Changed |= doMerge(ConstGlobals, M, true);

  return Changed;
}

void MemoryInstrumenter::checkFeatures(Module &M) {
  // Check whether any memory allocation function can
  // potentially be pointed by function pointers.
  // Also, all intrinsic functions will be called directly,
  // i.e. not via function pointers.
  for (Module::iterator F = M.begin(); F != M.end(); ++F) {
    if (DynAAUtils::IsMalloc(F) || F->isIntrinsic()) {
      for (Value::use_iterator UI = F->use_begin(); UI != F->use_end(); ++UI) {
        User *Usr = *UI;
        assert(isa<CallInst>(Usr) || isa<InvokeInst>(Usr));
        CallSite CS(cast<Instruction>(Usr));
        for (unsigned i = 0; i < CS.arg_size(); ++i)
          assert(CS.getArgument(i) != F);
      }
    }
  }

  // Check whether memory allocation functions are captured.
  for (Module::iterator F = M.begin(); F != M.end(); ++F) {
    // 0 is the return, 1 is the first parameter.
    if (F->isDeclaration() && F->doesNotAlias(0) && !DynAAUtils::IsMalloc(F)) {
      errs().changeColor(raw_ostream::RED);
      errs() << F->getName() << "'s return value is marked noalias, ";
      errs() << "but the function is not treated as malloc.\n";
      errs().resetColor();
    }
  }

  // Sequential types except pointer types shouldn't be used as the type of
  // an instruction, a function parameter, or a global variable.
  for (Module::global_iterator GI = M.global_begin(), E = M.global_end();
       GI != E; ++GI) {
    if (isa<SequentialType>(GI->getType()))
      assert(GI->getType()->isPointerTy());
  }
  for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
    for (Function::arg_iterator AI = F->arg_begin(); AI != F->arg_end(); ++AI) {
      if (isa<SequentialType>(AI->getType()))
        assert(AI->getType()->isPointerTy());
    }
  }
  for (Module::iterator F = M.begin(); F != M.end(); ++F) {
    for (Function::iterator BB = F->begin(); BB != F->end(); ++BB) {
      for (BasicBlock::iterator Ins = BB->begin(); Ins != BB->end(); ++Ins) {
        if (isa<SequentialType>(Ins->getType()))
          assert(Ins->getType()->isPointerTy());
      }
    }
  }

  // We don't support multi-process programs for now.
  if (!HookFork)
    assert(M.getFunction("fork") == NULL);
}

void PIC16AsmPrinter::EmitRomData (Module &M)
{
  SwitchToSection(TAI->getReadOnlySection());
  IsRomData = true;
  for (Module::global_iterator I = M.global_begin(), E = M.global_end();
       I != E; ++I) {
    if (!I->hasInitializer())   // External global require no code.
      continue;

    Constant *C = I->getInitializer();
    const PointerType *PtrTy = I->getType();
    int AddrSpace = PtrTy->getAddressSpace();
    if ((!C->isNullValue()) && (AddrSpace == PIC16ISD::ROM_SPACE)) {

      if (EmitSpecialLLVMGlobal(I))
        continue;

      // Any variables reaching here with "." in its name is a local scope
      // variable and should not be printed in global data section.
      std::string name = Mang->getValueName(I);
      if (name.find(".") != std::string::npos)
        continue;

      I->setSection(TAI->getReadOnlySection()->getName());
      O << name;
      EmitGlobalConstant(C, AddrSpace);
      O << "\n";
    }
  }
  IsRomData = false;
}

/// AnalyzeGlobals - Scan through the users of all of the internal
/// GlobalValue's in the program.  If none of them have their "address taken"
/// (really, their address passed to something nontrivial), record this fact,
/// and record the functions that they are used directly in.
void GlobalsModRef::AnalyzeGlobals(Module &M) {
  std::vector<Function*> Readers, Writers;
  for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
    if (I->hasLocalLinkage()) {
      if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
        // Remember that we are tracking this global.
        NonAddressTakenGlobals.insert(I);
        ++NumNonAddrTakenFunctions;
      }
      Readers.clear(); Writers.clear();
    }

  for (Module::global_iterator I = M.global_begin(), E = M.global_end();
       I != E; ++I)
    if (I->hasLocalLinkage()) {
      if (!AnalyzeUsesOfPointer(I, Readers, Writers)) {
        // Remember that we are tracking this global, and the mod/ref fns
        NonAddressTakenGlobals.insert(I);

        for (unsigned i = 0, e = Readers.size(); i != e; ++i)
          FunctionInfo[Readers[i]].GlobalInfo[I] |= Ref;

        if (!I->isConstant())  // No need to keep track of writers to constants
          for (unsigned i = 0, e = Writers.size(); i != e; ++i)
            FunctionInfo[Writers[i]].GlobalInfo[I] |= Mod;
        ++NumNonAddrTakenGlobalVars;

        // If this global holds a pointer type, see if it is an indirect global.
        if (I->getType()->getElementType()->isPointerTy() &&
            AnalyzeIndirectGlobalMemory(I))
          ++NumIndirectGlobalVars;
      }
      Readers.clear(); Writers.clear();
    }
}

void MemoryInstrumenter::instrumentGlobals(Module &M) {
  TargetData &TD = getAnalysis<TargetData>();
  IDAssigner &IDA = getAnalysis<IDAssigner>();

  // Function HookGlobalsAlloc contains only one basic block.
  // The BB iterates through all global variables, and calls HookMemAlloc
  // for each of them.
  BasicBlock *BB = BasicBlock::Create(M.getContext(), "entry",
                                      GlobalsAllocHook);
  Instruction *Ret = ReturnInst::Create(M.getContext(), BB);

  for (Module::global_iterator GI = M.global_begin(), E = M.global_end();
       GI != E; ++GI) {
    // We are going to delete llvm.global_ctors.
    // Therefore, don't instrument it.
    if (GI->getName() == "llvm.global_ctors")
      continue;
    // Prevent global variables from sharing the same address, because it
    // breaks the assumption that global variables do not alias.
    // The same goes to functions.
    if (GI->hasUnnamedAddr()) {
      GI->setUnnamedAddr(false);
    }
    uint64_t TypeSize = TD.getTypeStoreSize(GI->getType()->getElementType());
    instrumentMemoryAllocation(GI,
                               ConstantInt::get(LongType, TypeSize),
                               NULL,
                               Ret);
    instrumentPointer(GI, NULL, Ret);
  }

  for (Module::iterator F = M.begin(); F != M.end(); ++F) {
    // These hooks added by us don't have a value ID.
    if (MemAllocHook == F || MainArgsAllocHook == F || TopLevelHook == F ||
        AddrTakenHook == F || CallHook == F || ReturnHook == F ||
        GlobalsAllocHook == F || MemHooksIniter == F || AfterForkHook == F ||
        BeforeForkHook == F) {
      continue;
    }
    // InvalidID: maybe this is inserted by alias checker in hybrid mode.
    if (IDA.getValueID(F) == IDAssigner::InvalidID)
      continue;
    // Ignore intrinsic functions because we cannot take the address of
    // an intrinsic. Also, no function pointers will point to instrinsic
    // functions.
    if (F->isIntrinsic())
      continue;
    // Prevent functions from sharing the same address.
    if (F->hasUnnamedAddr()) {
      F->setUnnamedAddr(false);
    }
    uint64_t TypeSize = TD.getTypeStoreSize(F->getType());
    assert(TypeSize == TD.getPointerSize());
    instrumentMemoryAllocation(F,
                               ConstantInt::get(LongType, TypeSize),
                               NULL,
                               Ret);
    instrumentPointer(F, NULL, Ret);
  }
}

/// OptimizeGlobals - This method uses information taken from DSA to optimize
/// global variables.
///
bool DSOpt::OptimizeGlobals(Module &M) {
  DSGraph* GG = TD->getGlobalsGraph();
  const DSGraph::ScalarMapTy &SM = GG->getScalarMap();
  bool Changed = false;

  for (Module::global_iterator I = M.global_begin(), E = M.global_end(); I != E; ++I)
    if (!I->isDeclaration()) { // Loop over all of the non-external globals...
      // Look up the node corresponding to this global, if it exists.
      DSNode *GNode = 0;
      DSGraph::ScalarMapTy::const_iterator SMI = SM.find(I);
      if (SMI != SM.end()) GNode = SMI->second.getNode();

      if (GNode == 0 && I->hasInternalLinkage()) {
        // If there is no entry in the scalar map for this global, it was never
        // referenced in the program.  If it has internal linkage, that means we
        // can delete it.  We don't ACTUALLY want to delete the global, just
        // remove anything that references the global: later passes will take
        // care of nuking it.
        if (!I->use_empty()) {
          I->replaceAllUsesWith(ConstantPointerNull::get(I->getType()));
          ++NumGlobalsIsolated;
        }
      } else if (GNode && GNode->NodeType.isCompleteNode()) {

        // If the node has not been read or written, and it is not externally
        // visible, kill any references to it so it can be DCE'd.
        if (!GNode->NodeType.isModifiedNode() && !GNode->NodeType.isReadNode() &&I->hasInternalLinkage()){
          if (!I->use_empty()) {
            I->replaceAllUsesWith(ConstantPointerNull::get(I->getType()));
            ++NumGlobalsIsolated;
          }
        }

        // We expect that there will almost always be a node for this global.
        // If there is, and the node doesn't have the M bit set, we can set the
        // 'constant' bit on the global.
        if (!GNode->NodeType.isModifiedNode() && !I->isConstant()) {
          I->setConstant(true);
          ++NumGlobalsConstanted;
          Changed = true;
        }
      }
    }
  return Changed;
}

void MemoryInstrumenter::checkFeatures(Module &M) {
    // Check whether any memory allocation function can
    // potentially be pointed by function pointers.
    // Also, all intrinsic functions will be called directly,
    // i.e. not via function pointers.
    for (Module::iterator F = M.begin(); F != M.end(); ++F) {
        if (DynAAUtils::IsMalloc(F) || F->isIntrinsic()) {
            for (Value::use_iterator UI = F->use_begin(); UI != F->use_end(); ++UI) {
                User *Usr = *UI;
                assert(isa<CallInst>(Usr) || isa<InvokeInst>(Usr));
                CallSite CS(cast<Instruction>(Usr));
                for (unsigned i = 0; i < CS.arg_size(); ++i)
                    assert(CS.getArgument(i) != F);
            }
        }
    }

    // Check whether memory allocation functions are captured.
    for (Module::iterator F = M.begin(); F != M.end(); ++F) {
        // 0 is the return, 1 is the first parameter.
        if (F->isDeclaration() && F->doesNotAlias(0) && !DynAAUtils::IsMalloc(F)) {
            errs().changeColor(raw_ostream::RED);
            errs() << F->getName() << "'s return value is marked noalias, ";
            errs() << "but the function is not treated as malloc.\n";
            errs().resetColor();
        }
    }

    // Global variables shouldn't be of the array type.
    for (Module::global_iterator GI = M.global_begin(), E = M.global_end();
            GI != E; ++GI) {
        assert(!GI->getType()->isArrayTy());
    }
    // A function parameter or an instruction can be an array, but we don't
    // instrument such constructs for now. Issue a warning on such cases.
    for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
        for (Function::arg_iterator AI = F->arg_begin(); AI != F->arg_end(); ++AI) {
            if (AI->getType()->isArrayTy()) {
                errs().changeColor(raw_ostream::RED);
                errs() << F->getName() << ":" << *AI << " is an array\n";
                errs().resetColor();
            }
        }
    }
    for (Module::iterator F = M.begin(); F != M.end(); ++F) {
        for (Function::iterator BB = F->begin(); BB != F->end(); ++BB) {
            for (BasicBlock::iterator Ins = BB->begin(); Ins != BB->end(); ++Ins) {
                if (Ins->getType()->isArrayTy()) {
                    errs().changeColor(raw_ostream::RED);
                    errs() << F->getName() << ":" << *Ins << " is an array\n";
                    errs().resetColor();
                }
            }
        }
    }
}

void CaptureConstraints::identify_fixed_integers(Module &M) {
	ExecOnce &EO = getAnalysis<ExecOnce>();
	
	fixed_integers.clear();
	// Global variables. 
	for (Module::global_iterator gi = M.global_begin();
			gi != M.global_end(); ++gi) {
		if (isa<IntegerType>(gi->getType()) || isa<PointerType>(gi->getType())) {
			fixed_integers.insert(gi);
			if (gi->hasInitializer())
				extract_from_consts(gi->getInitializer());
		}
	}
	
	// Instructions and their constant operands. 
	forallinst(M, ii) {
		if (EO.not_executed(ii))
			continue;
		if (!EO.executed_once(ii))
			continue;
		if (isa<IntegerType>(ii->getType()) || isa<PointerType>(ii->getType())) {
			fixed_integers.insert(ii);
		}
		// No matter reachable or not, capture its constant operands. 
		for (unsigned i = 0; i < ii->getNumOperands(); ++i) {
			if (Constant *c = dyn_cast<Constant>(ii->getOperand(i)))
				extract_from_consts(c);
		}
	}
	
	// Function parameters. 
	forallfunc(M, f) {
		if (EO.not_executed(f))
			continue;
		if (!EO.executed_once(f))
			continue;
		for (Function::arg_iterator ai = f->arg_begin();
				ai != f->arg_end(); ++ai) {
			if (isa<IntegerType>(ai->getType()) || isa<PointerType>(ai->getType()))
				fixed_integers.insert(ai);
		}
	}
}

bool ARMGlobalMerge::doInitialization(Module &M) {
  SmallVector<GlobalVariable*, 16> Globals, ConstGlobals, BSSGlobals;
  const TargetData *TD = TLI->getTargetData();
  unsigned MaxOffset = TLI->getMaximalGlobalOffset();
  bool Changed = false;

  // Disable this pass on darwin. The debugger is not yet ready to extract
  // variable's  info from a merged global.
  if (TLI->getTargetMachine().getSubtarget<ARMSubtarget>().isTargetDarwin())
    return false;

  // Grab all non-const globals.
  for (Module::global_iterator I = M.global_begin(),
         E = M.global_end(); I != E; ++I) {
    // Merge is safe for "normal" internal globals only
    if (!I->hasLocalLinkage() || I->isThreadLocal() || I->hasSection())
      continue;

    // Ignore fancy-aligned globals for now.
    if (I->getAlignment() != 0)
      continue;

    // Ignore all 'special' globals.
    if (I->getName().startswith("llvm.") ||
        I->getName().startswith(".llvm."))
      continue;

    if (TD->getTypeAllocSize(I->getType()->getElementType()) < MaxOffset) {
      const TargetLoweringObjectFile &TLOF = TLI->getObjFileLowering();
      if (TLOF.getKindForGlobal(I, TLI->getTargetMachine()).isBSSLocal())
        BSSGlobals.push_back(I);
      else if (I->isConstant())
        ConstGlobals.push_back(I);
      else
        Globals.push_back(I);
    }
  }

  if (Globals.size() > 1)
    Changed |= doMerge(Globals, M, false);
  if (BSSGlobals.size() > 1)
    Changed |= doMerge(BSSGlobals, M, false);

  // FIXME: This currently breaks the EH processing due to way how the 
  // typeinfo detection works. We might want to detect the TIs and ignore 
  // them in the future.
  // if (ConstGlobals.size() > 1)
  //  Changed |= doMerge(ConstGlobals, M, true);

  return Changed;
}

// Create shadow information for all global variables.
void LLPEAnalysisPass::initShadowGlobals(Module& M, uint32_t extraSlots) {

  uint32_t i = 0;
  uint32_t nGlobals = std::distance(M.global_begin(), M.global_end());
  // extraSlots are reserved for new globals we know will be introduced between now and specialisation start.
  nGlobals += extraSlots;
  shadowGlobals = new ShadowGV[nGlobals];

  // Assign them all numbers before computing initialisers, because the initialiser can
  // reference another global, and getValPB will then lookup in shadowGlobalsIdx.

  for(Module::global_iterator it = M.global_begin(), itend = M.global_end(); it != itend; ++it, ++i) {

    shadowGlobals[i].G = it;
    shadowGlobalsIdx[it] = i;

  }

  i = 0;
  for(Module::global_iterator it = M.global_begin(), itend = M.global_end(); it != itend; ++it, ++i) {

    // getTypeStoreSize can be expensive, so do it once here.
    if(it->isConstant()) {
      shadowGlobals[i].storeSize = GlobalAA->getTypeStoreSize(shadowGlobals[i].G->getType());
      continue;
    }

    // Non-constant global -- assign it a heap slot.
    shadowGlobals[i].allocIdx = (int32_t)heap.size();
    
    heap.push_back(AllocData());
    AllocData& AD = heap.back();
    AD.allocIdx = heap.size() - 1;
    AD.storeSize = GlobalAA->getTypeStoreSize(it->getType()->getElementType());
    AD.isCommitted = true;
    // This usually points to a malloc instruction -- here the global itself.
    AD.allocValue = ShadowValue(&(shadowGlobals[i]));
    AD.allocType = shadowGlobals[i].G->getType();

    //errs() << "Init store for " << *it << " -> ";
    //printPB(errs(), *Init);
    //errs() << "\n";

    shadowGlobals[i].storeSize = AD.storeSize;

  }

}

/**
 * Print the field declarations.
 */
void JVMWriter::printFields() {
    out << "; Fields\n";
    for(Module::global_iterator i = module->global_begin(),
                                e = module->global_end(); i != e; i++) {
        if(i->isDeclaration()) {
            out << ".extern field ";
            externRefs.insert(i);
        } else
            out << ".field "
                << (i->hasLocalLinkage() ? "private " : "public ")
                << "static final ";
        out << getValueName(i) << ' ' << getTypeDescriptor(i->getType());
        if(debug >= 3)
            out << " ; " << *i;
        else
            out << '\n';
    }
    out << '\n';
}

bool ARMGlobalMerge::doInitialization(Module &M) {
  SmallVector<GlobalVariable*, 16> Globals, ConstGlobals;
  const TargetData *TD = TLI->getTargetData();
  unsigned MaxOffset = TLI->getMaximalGlobalOffset();
  bool Changed = false;

  // Grab all non-const globals.
  for (Module::global_iterator I = M.global_begin(),
         E = M.global_end(); I != E; ++I) {
    // Merge is safe for "normal" internal globals only
    if (!I->hasLocalLinkage() || I->isThreadLocal() || I->hasSection())
      continue;

    // Ignore fancy-aligned globals for now.
    if (I->getAlignment() != 0)
      continue;

    // Ignore all 'special' globals.
    if (I->getName().startswith("llvm.") ||
        I->getName().startswith(".llvm."))
      continue;

    if (TD->getTypeAllocSize(I->getType()) < MaxOffset) {
      if (I->isConstant())
        ConstGlobals.push_back(I);
      else
        Globals.push_back(I);
    }
  }

  if (Globals.size() > 1)
    Changed |= doMerge(Globals, M, false);
  // FIXME: This currently breaks the EH processing due to way how the 
  // typeinfo detection works. We might want to detect the TIs and ignore 
  // them in the future.
  
  // if (ConstGlobals.size() > 1)
  //  Changed |= doMerge(ConstGlobals, M, true);

  return Changed;
}

bool GlobalMerge::doInitialization(Module &M) {
  if (!EnableGlobalMerge)
    return false;

  auto &DL = M.getDataLayout();
  DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
                                                        BSSGlobals;
  bool Changed = false;
  setMustKeepGlobalVariables(M);

  // Grab all non-const globals.
  for (Module::global_iterator I = M.global_begin(),
         E = M.global_end(); I != E; ++I) {
    // Merge is safe for "normal" internal or external globals only
    if (I->isDeclaration() || I->isThreadLocal() || I->hasSection())
      continue;

    if (!(EnableGlobalMergeOnExternal && I->hasExternalLinkage()) &&
        !I->hasInternalLinkage())
      continue;

    PointerType *PT = dyn_cast<PointerType>(I->getType());
    assert(PT && "Global variable is not a pointer!");

    unsigned AddressSpace = PT->getAddressSpace();

    // Ignore fancy-aligned globals for now.
    unsigned Alignment = DL.getPreferredAlignment(I);
    Type *Ty = I->getType()->getElementType();
    if (Alignment > DL.getABITypeAlignment(Ty))
      continue;

    // Ignore all 'special' globals.
    if (I->getName().startswith("llvm.") ||
        I->getName().startswith(".llvm."))
      continue;

    // Ignore all "required" globals:
    if (isMustKeepGlobalVariable(I))
      continue;

    if (DL.getTypeAllocSize(Ty) < MaxOffset) {
      if (TargetLoweringObjectFile::getKindForGlobal(I, *TM).isBSSLocal())
        BSSGlobals[AddressSpace].push_back(I);
      else if (I->isConstant())
        ConstGlobals[AddressSpace].push_back(I);
      else
        Globals[AddressSpace].push_back(I);
    }
  }

  for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
       I = Globals.begin(), E = Globals.end(); I != E; ++I)
    if (I->second.size() > 1)
      Changed |= doMerge(I->second, M, false, I->first);

  for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
       I = BSSGlobals.begin(), E = BSSGlobals.end(); I != E; ++I)
    if (I->second.size() > 1)
      Changed |= doMerge(I->second, M, false, I->first);

  if (EnableGlobalMergeOnConst)
    for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
         I = ConstGlobals.begin(), E = ConstGlobals.end(); I != E; ++I)
      if (I->second.size() > 1)
        Changed |= doMerge(I->second, M, true, I->first);

  return Changed;
}

bool AllocateDataSegment::runOnModule(Module &M) {
  DataLayout DL(&M);
  Type *I8 = Type::getInt8Ty(M.getContext());
  Type *I32 = Type::getInt32Ty(M.getContext());
  Type *IntPtrType = DL.getIntPtrType(M.getContext());

  // First, we do a pass over the global variables, in which we compute
  // the amount of required padding between them and consequently their
  // addresses relative to the memory base of the sandbox. References to each
  // global are then replaced with direct memory pointers.
  uint32_t VarOffset = 0;
  DenseMap<GlobalVariable*, uint32_t> VarPadding;
  for (Module::global_iterator GV = M.global_begin(), E = M.global_end();
       GV != E; ++GV) {
    assert(GV->hasInitializer());

    uint32_t Padding = getPadding(VarOffset, GV, DL);
    VarPadding[GV] = Padding;
    VarOffset += Padding;

    GV->replaceAllUsesWith(
        ConstantExpr::getIntToPtr(
            ConstantInt::get(IntPtrType,
                             DataSegmentBaseAddress + VarOffset),
            GV->getType()));

    VarOffset += DL.getTypeStoreSize(GV->getType()->getPointerElementType());
  }

  // Using the offsets computed above, we prepare the layout and the contents
  // of the desired data structure. After the type and initializer of each
  // global is copied, the global is not needed any more and it is erased.
  SmallVector<Type*, 10> TemplateLayout;
  SmallVector<Constant*, 10> TemplateData;
  for (Module::global_iterator It = M.global_begin(), E = M.global_end();
       It != E; ) {
    GlobalVariable *GV = It++;

    uint32_t Padding = VarPadding[GV];
    if (Padding > 0) {
      Type *PaddingType = ArrayType::get(I8, Padding);
      TemplateLayout.push_back(PaddingType);
      TemplateData.push_back(ConstantAggregateZero::get(PaddingType));
    }

    TemplateLayout.push_back(GV->getType()->getPointerElementType());
    TemplateData.push_back(GV->getInitializer());

    GV->eraseFromParent();
  }

  // Finally, we create the struct and size global variables.
  StructType *TemplateType =
      StructType::create(M.getContext(), ExternalSymName_DataSegment);
  TemplateType->setBody(TemplateLayout, /*isPacked=*/true);

  Constant *Template = ConstantStruct::get(TemplateType, TemplateData);
  new GlobalVariable(M, Template->getType(), /*isConstant=*/true,
                     GlobalVariable::ExternalLinkage, Template,
                     ExternalSymName_DataSegment);

  Constant *TemplateSize =
      ConstantInt::get(I32, DL.getTypeAllocSize(TemplateType));
  new GlobalVariable(M, TemplateSize->getType(), /*isConstant=*/true,
                     GlobalVariable::ExternalLinkage, TemplateSize,
                     ExternalSymName_DataSegmentSize);

  return true;
}