Internalize.cpp revision 263508
1//===-- Internalize.cpp - Mark functions internal -------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass loops over all of the functions and variables in the input module. 11// If the function or variable is not in the list of external names given to 12// the pass it is marked as internal. 13// 14// This transformation would not be legal in a regular compilation, but it gets 15// extra information from the linker about what is safe. 16// 17// For example: Internalizing a function with external linkage. Only if we are 18// told it is only used from within this module, it is safe to do it. 19// 20//===----------------------------------------------------------------------===// 21 22#define DEBUG_TYPE "internalize" 23#include "llvm/Transforms/IPO.h" 24#include "llvm/ADT/SmallPtrSet.h" 25#include "llvm/ADT/Statistic.h" 26#include "llvm/Analysis/CallGraph.h" 27#include "llvm/IR/Module.h" 28#include "llvm/Pass.h" 29#include "llvm/Support/CommandLine.h" 30#include "llvm/Support/Debug.h" 31#include "llvm/Support/raw_ostream.h" 32#include "llvm/Transforms/Utils/GlobalStatus.h" 33#include "llvm/Transforms/Utils/ModuleUtils.h" 34#include <fstream> 35#include <set> 36using namespace llvm; 37 38STATISTIC(NumAliases , "Number of aliases internalized"); 39STATISTIC(NumFunctions, "Number of functions internalized"); 40STATISTIC(NumGlobals , "Number of global vars internalized"); 41 42// APIFile - A file which contains a list of symbols that should not be marked 43// external. 44static cl::opt<std::string> 45APIFile("internalize-public-api-file", cl::value_desc("filename"), 46 cl::desc("A file containing list of symbol names to preserve")); 47 48// APIList - A list of symbols that should not be marked internal. 49static cl::list<std::string> 50APIList("internalize-public-api-list", cl::value_desc("list"), 51 cl::desc("A list of symbol names to preserve"), 52 cl::CommaSeparated); 53 54namespace { 55 class InternalizePass : public ModulePass { 56 std::set<std::string> ExternalNames; 57 public: 58 static char ID; // Pass identification, replacement for typeid 59 explicit InternalizePass(); 60 explicit InternalizePass(ArrayRef<const char *> ExportList); 61 void LoadFile(const char *Filename); 62 virtual bool runOnModule(Module &M); 63 64 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 65 AU.setPreservesCFG(); 66 AU.addPreserved<CallGraph>(); 67 } 68 }; 69} // end anonymous namespace 70 71char InternalizePass::ID = 0; 72INITIALIZE_PASS(InternalizePass, "internalize", 73 "Internalize Global Symbols", false, false) 74 75InternalizePass::InternalizePass() 76 : ModulePass(ID) { 77 initializeInternalizePassPass(*PassRegistry::getPassRegistry()); 78 if (!APIFile.empty()) // If a filename is specified, use it. 79 LoadFile(APIFile.c_str()); 80 ExternalNames.insert(APIList.begin(), APIList.end()); 81} 82 83InternalizePass::InternalizePass(ArrayRef<const char *> ExportList) 84 : ModulePass(ID){ 85 initializeInternalizePassPass(*PassRegistry::getPassRegistry()); 86 for(ArrayRef<const char *>::const_iterator itr = ExportList.begin(); 87 itr != ExportList.end(); itr++) { 88 ExternalNames.insert(*itr); 89 } 90} 91 92void InternalizePass::LoadFile(const char *Filename) { 93 // Load the APIFile... 94 std::ifstream In(Filename); 95 if (!In.good()) { 96 errs() << "WARNING: Internalize couldn't load file '" << Filename 97 << "'! Continuing as if it's empty.\n"; 98 return; // Just continue as if the file were empty 99 } 100 while (In) { 101 std::string Symbol; 102 In >> Symbol; 103 if (!Symbol.empty()) 104 ExternalNames.insert(Symbol); 105 } 106} 107 108static bool shouldInternalize(const GlobalValue &GV, 109 const std::set<std::string> &ExternalNames) { 110 // Function must be defined here 111 if (GV.isDeclaration()) 112 return false; 113 114 // Available externally is really just a "declaration with a body". 115 if (GV.hasAvailableExternallyLinkage()) 116 return false; 117 118 // Already has internal linkage 119 if (GV.hasLocalLinkage()) 120 return false; 121 122 // Marked to keep external? 123 if (ExternalNames.count(GV.getName())) 124 return false; 125 126 return true; 127} 128 129bool InternalizePass::runOnModule(Module &M) { 130 CallGraph *CG = getAnalysisIfAvailable<CallGraph>(); 131 CallGraphNode *ExternalNode = CG ? CG->getExternalCallingNode() : 0; 132 bool Changed = false; 133 134 SmallPtrSet<GlobalValue *, 8> Used; 135 collectUsedGlobalVariables(M, Used, false); 136 137 // We must assume that globals in llvm.used have a reference that not even 138 // the linker can see, so we don't internalize them. 139 // For llvm.compiler.used the situation is a bit fuzzy. The assembler and 140 // linker can drop those symbols. If this pass is running as part of LTO, 141 // one might think that it could just drop llvm.compiler.used. The problem 142 // is that even in LTO llvm doesn't see every reference. For example, 143 // we don't see references from function local inline assembly. To be 144 // conservative, we internalize symbols in llvm.compiler.used, but we 145 // keep llvm.compiler.used so that the symbol is not deleted by llvm. 146 for (SmallPtrSet<GlobalValue *, 8>::iterator I = Used.begin(), E = Used.end(); 147 I != E; ++I) { 148 GlobalValue *V = *I; 149 ExternalNames.insert(V->getName()); 150 } 151 152 // Mark all functions not in the api as internal. 153 // FIXME: maybe use private linkage? 154 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) { 155 if (!shouldInternalize(*I, ExternalNames)) 156 continue; 157 158 I->setLinkage(GlobalValue::InternalLinkage); 159 160 if (ExternalNode) 161 // Remove a callgraph edge from the external node to this function. 162 ExternalNode->removeOneAbstractEdgeTo((*CG)[I]); 163 164 Changed = true; 165 ++NumFunctions; 166 DEBUG(dbgs() << "Internalizing func " << I->getName() << "\n"); 167 } 168 169 // Never internalize the llvm.used symbol. It is used to implement 170 // attribute((used)). 171 // FIXME: Shouldn't this just filter on llvm.metadata section?? 172 ExternalNames.insert("llvm.used"); 173 ExternalNames.insert("llvm.compiler.used"); 174 175 // Never internalize anchors used by the machine module info, else the info 176 // won't find them. (see MachineModuleInfo.) 177 ExternalNames.insert("llvm.global_ctors"); 178 ExternalNames.insert("llvm.global_dtors"); 179 ExternalNames.insert("llvm.global.annotations"); 180 181 // Never internalize symbols code-gen inserts. 182 // FIXME: We should probably add this (and the __stack_chk_guard) via some 183 // type of call-back in CodeGen. 184 ExternalNames.insert("__stack_chk_fail"); 185 ExternalNames.insert("__stack_chk_guard"); 186 187 // Mark all global variables with initializers that are not in the api as 188 // internal as well. 189 // FIXME: maybe use private linkage? 190 for (Module::global_iterator I = M.global_begin(), E = M.global_end(); 191 I != E; ++I) { 192 if (!shouldInternalize(*I, ExternalNames)) 193 continue; 194 195 I->setLinkage(GlobalValue::InternalLinkage); 196 Changed = true; 197 ++NumGlobals; 198 DEBUG(dbgs() << "Internalized gvar " << I->getName() << "\n"); 199 } 200 201 // Mark all aliases that are not in the api as internal as well. 202 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end(); 203 I != E; ++I) { 204 if (!shouldInternalize(*I, ExternalNames)) 205 continue; 206 207 I->setLinkage(GlobalValue::InternalLinkage); 208 Changed = true; 209 ++NumAliases; 210 DEBUG(dbgs() << "Internalized alias " << I->getName() << "\n"); 211 } 212 213 return Changed; 214} 215 216ModulePass *llvm::createInternalizePass() { 217 return new InternalizePass(); 218} 219 220ModulePass *llvm::createInternalizePass(ArrayRef<const char *> ExportList) { 221 return new InternalizePass(ExportList); 222} 223