ExtractFunction.cpp revision 276479
1185573Srwatson//===- ExtractFunction.cpp - Extract a function from Program --------------===// 2185573Srwatson// 3155131Srwatson// The LLVM Compiler Infrastructure 4155131Srwatson// 5155131Srwatson// This file is distributed under the University of Illinois Open Source 6155131Srwatson// License. See LICENSE.TXT for details. 7155131Srwatson// 8155131Srwatson//===----------------------------------------------------------------------===// 9155131Srwatson// 10155131Srwatson// This file implements several methods that are used to extract functions, 11155131Srwatson// loops, or portions of a module from the rest of the module. 12155131Srwatson// 13155131Srwatson//===----------------------------------------------------------------------===// 14185573Srwatson 15155131Srwatson#include "BugDriver.h" 16155131Srwatson#include "llvm/IR/Constants.h" 17155131Srwatson#include "llvm/IR/DataLayout.h" 18155131Srwatson#include "llvm/IR/DerivedTypes.h" 19155131Srwatson#include "llvm/IR/LLVMContext.h" 20155131Srwatson#include "llvm/IR/Module.h" 21155131Srwatson#include "llvm/IR/Verifier.h" 22155131Srwatson#include "llvm/Pass.h" 23155131Srwatson#include "llvm/PassManager.h" 24155131Srwatson#include "llvm/Support/CommandLine.h" 25155131Srwatson#include "llvm/Support/Debug.h" 26155131Srwatson#include "llvm/Support/FileUtilities.h" 27155131Srwatson#include "llvm/Support/Path.h" 28155131Srwatson#include "llvm/Support/Signals.h" 29155131Srwatson#include "llvm/Support/ToolOutputFile.h" 30186647Srwatson#include "llvm/Transforms/IPO.h" 31155131Srwatson#include "llvm/Transforms/Scalar.h" 32155131Srwatson#include "llvm/Transforms/Utils/Cloning.h" 33155131Srwatson#include "llvm/Transforms/Utils/CodeExtractor.h" 34156283Srwatson#include <set> 35156283Srwatsonusing namespace llvm; 36156283Srwatson 37155131Srwatson#define DEBUG_TYPE "bugpoint" 38156283Srwatson 39156283Srwatsonnamespace llvm { 40156283Srwatson bool DisableSimplifyCFG = false; 41155131Srwatson extern cl::opt<std::string> OutputPrefix; 42155131Srwatson} // End llvm namespace 43155131Srwatson 44186647Srwatsonnamespace { 45155131Srwatson cl::opt<bool> 46186647Srwatson NoDCE ("disable-dce", 47155131Srwatson cl::desc("Do not use the -dce pass to reduce testcases")); 48155131Srwatson cl::opt<bool, true> 49155131Srwatson NoSCFG("disable-simplifycfg", cl::location(DisableSimplifyCFG), 50155131Srwatson cl::desc("Do not use the -simplifycfg pass to reduce testcases")); 51186647Srwatson 52155131Srwatson Function* globalInitUsesExternalBA(GlobalVariable* GV) { 53186647Srwatson if (!GV->hasInitializer()) 54155131Srwatson return nullptr; 55155131Srwatson 56155131Srwatson Constant *I = GV->getInitializer(); 57155131Srwatson 58155131Srwatson // walk the values used by the initializer 59155131Srwatson // (and recurse into things like ConstantExpr) 60155131Srwatson std::vector<Constant*> Todo; 61155131Srwatson std::set<Constant*> Done; 62155131Srwatson Todo.push_back(I); 63155131Srwatson 64155131Srwatson while (!Todo.empty()) { 65155131Srwatson Constant* V = Todo.back(); 66155131Srwatson Todo.pop_back(); 67155131Srwatson Done.insert(V); 68155131Srwatson 69155131Srwatson if (BlockAddress *BA = dyn_cast<BlockAddress>(V)) { 70155131Srwatson Function *F = BA->getFunction(); 71155131Srwatson if (F->isDeclaration()) 72155131Srwatson return F; 73155131Srwatson } 74155131Srwatson 75155131Srwatson for (User::op_iterator i = V->op_begin(), e = V->op_end(); i != e; ++i) { 76155131Srwatson Constant *C = dyn_cast<Constant>(*i); 77155131Srwatson if (C && !isa<GlobalValue>(C) && !Done.count(C)) 78155131Srwatson Todo.push_back(C); 79155131Srwatson } 80155131Srwatson } 81155131Srwatson return nullptr; 82155131Srwatson } 83155131Srwatson} // end anonymous namespace 84155131Srwatson 85155131Srwatson/// deleteInstructionFromProgram - This method clones the current Program and 86155131Srwatson/// deletes the specified instruction from the cloned module. It then runs a 87155131Srwatson/// series of cleanup passes (ADCE and SimplifyCFG) to eliminate any code which 88155131Srwatson/// depends on the value. The modified module is then returned. 89155131Srwatson/// 90155131SrwatsonModule *BugDriver::deleteInstructionFromProgram(const Instruction *I, 91155131Srwatson unsigned Simplification) { 92155131Srwatson // FIXME, use vmap? 93155131Srwatson Module *Clone = CloneModule(Program); 94155131Srwatson 95155131Srwatson const BasicBlock *PBB = I->getParent(); 96155131Srwatson const Function *PF = PBB->getParent(); 97155131Srwatson 98155131Srwatson Module::iterator RFI = Clone->begin(); // Get iterator to corresponding fn 99155131Srwatson std::advance(RFI, std::distance(PF->getParent()->begin(), 100155131Srwatson Module::const_iterator(PF))); 101155131Srwatson 102155131Srwatson Function::iterator RBI = RFI->begin(); // Get iterator to corresponding BB 103155131Srwatson std::advance(RBI, std::distance(PF->begin(), Function::const_iterator(PBB))); 104155131Srwatson 105155131Srwatson BasicBlock::iterator RI = RBI->begin(); // Get iterator to corresponding inst 106155131Srwatson std::advance(RI, std::distance(PBB->begin(), BasicBlock::const_iterator(I))); 107155131Srwatson Instruction *TheInst = RI; // Got the corresponding instruction! 108155131Srwatson 109155131Srwatson // If this instruction produces a value, replace any users with null values 110155131Srwatson if (!TheInst->getType()->isVoidTy()) 111155131Srwatson TheInst->replaceAllUsesWith(Constant::getNullValue(TheInst->getType())); 112155131Srwatson 113155131Srwatson // Remove the instruction from the program. 114155131Srwatson TheInst->getParent()->getInstList().erase(TheInst); 115155131Srwatson 116155131Srwatson // Spiff up the output a little bit. 117155131Srwatson std::vector<std::string> Passes; 118155131Srwatson 119155131Srwatson /// Can we get rid of the -disable-* options? 120155131Srwatson if (Simplification > 1 && !NoDCE) 121155131Srwatson Passes.push_back("dce"); 122155131Srwatson if (Simplification && !DisableSimplifyCFG) 123155131Srwatson Passes.push_back("simplifycfg"); // Delete dead control flow 124155131Srwatson 125155131Srwatson Passes.push_back("verify"); 126155131Srwatson Module *New = runPassesOn(Clone, Passes); 127155131Srwatson delete Clone; 128155131Srwatson if (!New) { 129155131Srwatson errs() << "Instruction removal failed. Sorry. :( Please report a bug!\n"; 130155131Srwatson exit(1); 131155131Srwatson } 132155131Srwatson return New; 133155131Srwatson} 134155131Srwatson 135155131Srwatson/// performFinalCleanups - This method clones the current Program and performs 136155131Srwatson/// a series of cleanups intended to get rid of extra cruft on the module 137155131Srwatson/// before handing it to the user. 138155131Srwatson/// 139155131SrwatsonModule *BugDriver::performFinalCleanups(Module *M, bool MayModifySemantics) { 140155131Srwatson // Make all functions external, so GlobalDCE doesn't delete them... 141155131Srwatson for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) 142155131Srwatson I->setLinkage(GlobalValue::ExternalLinkage); 143155131Srwatson 144155131Srwatson std::vector<std::string> CleanupPasses; 145155131Srwatson CleanupPasses.push_back("globaldce"); 146155131Srwatson 147155131Srwatson if (MayModifySemantics) 148155131Srwatson CleanupPasses.push_back("deadarghaX0r"); 149155131Srwatson else 150155131Srwatson CleanupPasses.push_back("deadargelim"); 151155131Srwatson 152155131Srwatson Module *New = runPassesOn(M, CleanupPasses); 153155131Srwatson if (!New) { 154155131Srwatson errs() << "Final cleanups failed. Sorry. :( Please report a bug!\n"; 155155131Srwatson return M; 156155131Srwatson } 157155131Srwatson delete M; 158155131Srwatson return New; 159155131Srwatson} 160155131Srwatson 161155131Srwatson 162155131Srwatson/// ExtractLoop - Given a module, extract up to one loop from it into a new 163155131Srwatson/// function. This returns null if there are no extractable loops in the 164155131Srwatson/// program or if the loop extractor crashes. 165155131SrwatsonModule *BugDriver::ExtractLoop(Module *M) { 166155131Srwatson std::vector<std::string> LoopExtractPasses; 167155131Srwatson LoopExtractPasses.push_back("loop-extract-single"); 168155131Srwatson 169186647Srwatson Module *NewM = runPassesOn(M, LoopExtractPasses); 170155131Srwatson if (!NewM) { 171186647Srwatson outs() << "*** Loop extraction failed: "; 172155131Srwatson EmitProgressBitcode(M, "loopextraction", true); 173155131Srwatson outs() << "*** Sorry. :( Please report a bug!\n"; 174155131Srwatson return nullptr; 175186647Srwatson } 176155131Srwatson 177186647Srwatson // Check to see if we created any new functions. If not, no loops were 178155131Srwatson // extracted and we should return null. Limit the number of loops we extract 179155131Srwatson // to avoid taking forever. 180155131Srwatson static unsigned NumExtracted = 32; 181155131Srwatson if (M->size() == NewM->size() || --NumExtracted == 0) { 182155131Srwatson delete NewM; 183155131Srwatson return nullptr; 184155131Srwatson } else { 185186647Srwatson assert(M->size() < NewM->size() && "Loop extract removed functions?"); 186155131Srwatson Module::iterator MI = NewM->begin(); 187186647Srwatson for (unsigned i = 0, e = M->size(); i != e; ++i) 188155131Srwatson ++MI; 189155131Srwatson } 190155131Srwatson 191155131Srwatson return NewM; 192155131Srwatson} 193155131Srwatson 194155131Srwatson 195155131Srwatson// DeleteFunctionBody - "Remove" the function by deleting all of its basic 196155131Srwatson// blocks, making it external. 197155131Srwatson// 198155131Srwatsonvoid llvm::DeleteFunctionBody(Function *F) { 199186647Srwatson // delete the body of the function... 200155131Srwatson F->deleteBody(); 201186647Srwatson assert(F->isDeclaration() && "This didn't make the function external!"); 202155131Srwatson} 203155131Srwatson 204155131Srwatson/// GetTorInit - Given a list of entries for static ctors/dtors, return them 205155131Srwatson/// as a constant array. 206155131Srwatsonstatic Constant *GetTorInit(std::vector<std::pair<Function*, int> > &TorList) { 207155131Srwatson assert(!TorList.empty() && "Don't create empty tor list!"); 208186647Srwatson std::vector<Constant*> ArrayElts; 209155131Srwatson Type *Int32Ty = Type::getInt32Ty(TorList[0].first->getContext()); 210186647Srwatson 211155131Srwatson StructType *STy = 212155131Srwatson StructType::get(Int32Ty, TorList[0].first->getType(), NULL); 213155131Srwatson for (unsigned i = 0, e = TorList.size(); i != e; ++i) { 214155131Srwatson Constant *Elts[] = { 215 ConstantInt::get(Int32Ty, TorList[i].second), 216 TorList[i].first 217 }; 218 ArrayElts.push_back(ConstantStruct::get(STy, Elts)); 219 } 220 return ConstantArray::get(ArrayType::get(ArrayElts[0]->getType(), 221 ArrayElts.size()), 222 ArrayElts); 223} 224 225/// SplitStaticCtorDtor - A module was recently split into two parts, M1/M2, and 226/// M1 has all of the global variables. If M2 contains any functions that are 227/// static ctors/dtors, we need to add an llvm.global_[cd]tors global to M2, and 228/// prune appropriate entries out of M1s list. 229static void SplitStaticCtorDtor(const char *GlobalName, Module *M1, Module *M2, 230 ValueToValueMapTy &VMap) { 231 GlobalVariable *GV = M1->getNamedGlobal(GlobalName); 232 if (!GV || GV->isDeclaration() || GV->hasLocalLinkage() || 233 !GV->use_empty()) return; 234 235 std::vector<std::pair<Function*, int> > M1Tors, M2Tors; 236 ConstantArray *InitList = dyn_cast<ConstantArray>(GV->getInitializer()); 237 if (!InitList) return; 238 239 for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i) { 240 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(InitList->getOperand(i))){ 241 if (CS->getNumOperands() != 2) return; // Not array of 2-element structs. 242 243 if (CS->getOperand(1)->isNullValue()) 244 break; // Found a null terminator, stop here. 245 246 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0)); 247 int Priority = CI ? CI->getSExtValue() : 0; 248 249 Constant *FP = CS->getOperand(1); 250 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(FP)) 251 if (CE->isCast()) 252 FP = CE->getOperand(0); 253 if (Function *F = dyn_cast<Function>(FP)) { 254 if (!F->isDeclaration()) 255 M1Tors.push_back(std::make_pair(F, Priority)); 256 else { 257 // Map to M2's version of the function. 258 F = cast<Function>(VMap[F]); 259 M2Tors.push_back(std::make_pair(F, Priority)); 260 } 261 } 262 } 263 } 264 265 GV->eraseFromParent(); 266 if (!M1Tors.empty()) { 267 Constant *M1Init = GetTorInit(M1Tors); 268 new GlobalVariable(*M1, M1Init->getType(), false, 269 GlobalValue::AppendingLinkage, 270 M1Init, GlobalName); 271 } 272 273 GV = M2->getNamedGlobal(GlobalName); 274 assert(GV && "Not a clone of M1?"); 275 assert(GV->use_empty() && "llvm.ctors shouldn't have uses!"); 276 277 GV->eraseFromParent(); 278 if (!M2Tors.empty()) { 279 Constant *M2Init = GetTorInit(M2Tors); 280 new GlobalVariable(*M2, M2Init->getType(), false, 281 GlobalValue::AppendingLinkage, 282 M2Init, GlobalName); 283 } 284} 285 286 287/// SplitFunctionsOutOfModule - Given a module and a list of functions in the 288/// module, split the functions OUT of the specified module, and place them in 289/// the new module. 290Module * 291llvm::SplitFunctionsOutOfModule(Module *M, 292 const std::vector<Function*> &F, 293 ValueToValueMapTy &VMap) { 294 // Make sure functions & globals are all external so that linkage 295 // between the two modules will work. 296 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) 297 I->setLinkage(GlobalValue::ExternalLinkage); 298 for (Module::global_iterator I = M->global_begin(), E = M->global_end(); 299 I != E; ++I) { 300 if (I->hasName() && I->getName()[0] == '\01') 301 I->setName(I->getName().substr(1)); 302 I->setLinkage(GlobalValue::ExternalLinkage); 303 } 304 305 ValueToValueMapTy NewVMap; 306 Module *New = CloneModule(M, NewVMap); 307 308 // Remove the Test functions from the Safe module 309 std::set<Function *> TestFunctions; 310 for (unsigned i = 0, e = F.size(); i != e; ++i) { 311 Function *TNOF = cast<Function>(VMap[F[i]]); 312 DEBUG(errs() << "Removing function "); 313 DEBUG(TNOF->printAsOperand(errs(), false)); 314 DEBUG(errs() << "\n"); 315 TestFunctions.insert(cast<Function>(NewVMap[TNOF])); 316 DeleteFunctionBody(TNOF); // Function is now external in this module! 317 } 318 319 320 // Remove the Safe functions from the Test module 321 for (Module::iterator I = New->begin(), E = New->end(); I != E; ++I) 322 if (!TestFunctions.count(I)) 323 DeleteFunctionBody(I); 324 325 326 // Try to split the global initializers evenly 327 for (Module::global_iterator I = M->global_begin(), E = M->global_end(); 328 I != E; ++I) { 329 GlobalVariable *GV = cast<GlobalVariable>(NewVMap[I]); 330 if (Function *TestFn = globalInitUsesExternalBA(I)) { 331 if (Function *SafeFn = globalInitUsesExternalBA(GV)) { 332 errs() << "*** Error: when reducing functions, encountered " 333 "the global '"; 334 GV->printAsOperand(errs(), false); 335 errs() << "' with an initializer that references blockaddresses " 336 "from safe function '" << SafeFn->getName() 337 << "' and from test function '" << TestFn->getName() << "'.\n"; 338 exit(1); 339 } 340 I->setInitializer(nullptr); // Delete the initializer to make it external 341 } else { 342 // If we keep it in the safe module, then delete it in the test module 343 GV->setInitializer(nullptr); 344 } 345 } 346 347 // Make sure that there is a global ctor/dtor array in both halves of the 348 // module if they both have static ctor/dtor functions. 349 SplitStaticCtorDtor("llvm.global_ctors", M, New, NewVMap); 350 SplitStaticCtorDtor("llvm.global_dtors", M, New, NewVMap); 351 352 return New; 353} 354 355//===----------------------------------------------------------------------===// 356// Basic Block Extraction Code 357//===----------------------------------------------------------------------===// 358 359/// ExtractMappedBlocksFromModule - Extract all but the specified basic blocks 360/// into their own functions. The only detail is that M is actually a module 361/// cloned from the one the BBs are in, so some mapping needs to be performed. 362/// If this operation fails for some reason (ie the implementation is buggy), 363/// this function should return null, otherwise it returns a new Module. 364Module *BugDriver::ExtractMappedBlocksFromModule(const 365 std::vector<BasicBlock*> &BBs, 366 Module *M) { 367 SmallString<128> Filename; 368 int FD; 369 std::error_code EC = sys::fs::createUniqueFile( 370 OutputPrefix + "-extractblocks%%%%%%%", FD, Filename); 371 if (EC) { 372 outs() << "*** Basic Block extraction failed!\n"; 373 errs() << "Error creating temporary file: " << EC.message() << "\n"; 374 EmitProgressBitcode(M, "basicblockextractfail", true); 375 return nullptr; 376 } 377 sys::RemoveFileOnSignal(Filename); 378 379 tool_output_file BlocksToNotExtractFile(Filename.c_str(), FD); 380 for (std::vector<BasicBlock*>::const_iterator I = BBs.begin(), E = BBs.end(); 381 I != E; ++I) { 382 BasicBlock *BB = *I; 383 // If the BB doesn't have a name, give it one so we have something to key 384 // off of. 385 if (!BB->hasName()) BB->setName("tmpbb"); 386 BlocksToNotExtractFile.os() << BB->getParent()->getName() << " " 387 << BB->getName() << "\n"; 388 } 389 BlocksToNotExtractFile.os().close(); 390 if (BlocksToNotExtractFile.os().has_error()) { 391 errs() << "Error writing list of blocks to not extract\n"; 392 EmitProgressBitcode(M, "basicblockextractfail", true); 393 BlocksToNotExtractFile.os().clear_error(); 394 return nullptr; 395 } 396 BlocksToNotExtractFile.keep(); 397 398 std::string uniqueFN = "--extract-blocks-file="; 399 uniqueFN += Filename.str(); 400 const char *ExtraArg = uniqueFN.c_str(); 401 402 std::vector<std::string> PI; 403 PI.push_back("extract-blocks"); 404 Module *Ret = runPassesOn(M, PI, false, 1, &ExtraArg); 405 406 sys::fs::remove(Filename.c_str()); 407 408 if (!Ret) { 409 outs() << "*** Basic Block extraction failed, please report a bug!\n"; 410 EmitProgressBitcode(M, "basicblockextractfail", true); 411 } 412 return Ret; 413} 414