1 //===-- DeadArgumentElimination.cpp - Eliminate dead arguments ------------===// 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 deletes dead arguments from internal functions. Dead argument 11 // elimination removes arguments which are directly dead, as well as arguments 12 // only passed into function calls as dead arguments of other functions. This 13 // pass also deletes dead return values in a similar way. 14 // 15 // This pass is often useful as a cleanup pass to run after aggressive 16 // interprocedural passes, which add possibly-dead arguments or return values. 17 // 18 //===----------------------------------------------------------------------===// 19 20 #define DEBUG_TYPE "deadargelim" 21 #include "llvm/Transforms/IPO.h" 22 #include "llvm/ADT/DenseMap.h" 23 #include "llvm/ADT/SmallVector.h" 24 #include "llvm/ADT/Statistic.h" 25 #include "llvm/ADT/StringExtras.h" 26 #include "llvm/DIBuilder.h" 27 #include "llvm/DebugInfo.h" 28 #include "llvm/IR/CallingConv.h" 29 #include "llvm/IR/Constant.h" 30 #include "llvm/IR/DerivedTypes.h" 31 #include "llvm/IR/Instructions.h" 32 #include "llvm/IR/IntrinsicInst.h" 33 #include "llvm/IR/LLVMContext.h" 34 #include "llvm/IR/Module.h" 35 #include "llvm/Pass.h" 36 #include "llvm/Support/CallSite.h" 37 #include "llvm/Support/Debug.h" 38 #include "llvm/Support/raw_ostream.h" 39 #include <map> 40 #include <set> 41 using namespace llvm; 42 43 STATISTIC(NumArgumentsEliminated, "Number of unread args removed"); 44 STATISTIC(NumRetValsEliminated , "Number of unused return values removed"); 45 STATISTIC(NumArgumentsReplacedWithUndef, 46 "Number of unread args replaced with undef"); 47 namespace { 48 /// DAE - The dead argument elimination pass. 49 /// 50 class DAE : public ModulePass { 51 public: 52 53 /// Struct that represents (part of) either a return value or a function 54 /// argument. Used so that arguments and return values can be used 55 /// interchangeably. 56 struct RetOrArg { 57 RetOrArg(const Function *F, unsigned Idx, bool IsArg) : F(F), Idx(Idx), 58 IsArg(IsArg) {} 59 const Function *F; 60 unsigned Idx; 61 bool IsArg; 62 63 /// Make RetOrArg comparable, so we can put it into a map. 64 bool operator<(const RetOrArg &O) const { 65 if (F != O.F) 66 return F < O.F; 67 else if (Idx != O.Idx) 68 return Idx < O.Idx; 69 else 70 return IsArg < O.IsArg; 71 } 72 73 /// Make RetOrArg comparable, so we can easily iterate the multimap. 74 bool operator==(const RetOrArg &O) const { 75 return F == O.F && Idx == O.Idx && IsArg == O.IsArg; 76 } 77 78 std::string getDescription() const { 79 return std::string((IsArg ? "Argument #" : "Return value #")) 80 + utostr(Idx) + " of function " + F->getName().str(); 81 } 82 }; 83 84 /// Liveness enum - During our initial pass over the program, we determine 85 /// that things are either alive or maybe alive. We don't mark anything 86 /// explicitly dead (even if we know they are), since anything not alive 87 /// with no registered uses (in Uses) will never be marked alive and will 88 /// thus become dead in the end. 89 enum Liveness { Live, MaybeLive }; 90 91 /// Convenience wrapper 92 RetOrArg CreateRet(const Function *F, unsigned Idx) { 93 return RetOrArg(F, Idx, false); 94 } 95 /// Convenience wrapper 96 RetOrArg CreateArg(const Function *F, unsigned Idx) { 97 return RetOrArg(F, Idx, true); 98 } 99 100 typedef std::multimap<RetOrArg, RetOrArg> UseMap; 101 /// This maps a return value or argument to any MaybeLive return values or 102 /// arguments it uses. This allows the MaybeLive values to be marked live 103 /// when any of its users is marked live. 104 /// For example (indices are left out for clarity): 105 /// - Uses[ret F] = ret G 106 /// This means that F calls G, and F returns the value returned by G. 107 /// - Uses[arg F] = ret G 108 /// This means that some function calls G and passes its result as an 109 /// argument to F. 110 /// - Uses[ret F] = arg F 111 /// This means that F returns one of its own arguments. 112 /// - Uses[arg F] = arg G 113 /// This means that G calls F and passes one of its own (G's) arguments 114 /// directly to F. 115 UseMap Uses; 116 117 typedef std::set<RetOrArg> LiveSet; 118 typedef std::set<const Function*> LiveFuncSet; 119 120 /// This set contains all values that have been determined to be live. 121 LiveSet LiveValues; 122 /// This set contains all values that are cannot be changed in any way. 123 LiveFuncSet LiveFunctions; 124 125 typedef SmallVector<RetOrArg, 5> UseVector; 126 127 // Map each LLVM function to corresponding metadata with debug info. If 128 // the function is replaced with another one, we should patch the pointer 129 // to LLVM function in metadata. 130 // As the code generation for module is finished (and DIBuilder is 131 // finalized) we assume that subprogram descriptors won't be changed, and 132 // they are stored in map for short duration anyway. 133 typedef DenseMap<Function*, DISubprogram> FunctionDIMap; 134 FunctionDIMap FunctionDIs; 135 136 protected: 137 // DAH uses this to specify a different ID. 138 explicit DAE(char &ID) : ModulePass(ID) {} 139 140 public: 141 static char ID; // Pass identification, replacement for typeid 142 DAE() : ModulePass(ID) { 143 initializeDAEPass(*PassRegistry::getPassRegistry()); 144 } 145 146 bool runOnModule(Module &M); 147 148 virtual bool ShouldHackArguments() const { return false; } 149 150 private: 151 Liveness MarkIfNotLive(RetOrArg Use, UseVector &MaybeLiveUses); 152 Liveness SurveyUse(Value::const_use_iterator U, UseVector &MaybeLiveUses, 153 unsigned RetValNum = 0); 154 Liveness SurveyUses(const Value *V, UseVector &MaybeLiveUses); 155 156 void CollectFunctionDIs(Module &M); 157 void SurveyFunction(const Function &F); 158 void MarkValue(const RetOrArg &RA, Liveness L, 159 const UseVector &MaybeLiveUses); 160 void MarkLive(const RetOrArg &RA); 161 void MarkLive(const Function &F); 162 void PropagateLiveness(const RetOrArg &RA); 163 bool RemoveDeadStuffFromFunction(Function *F); 164 bool DeleteDeadVarargs(Function &Fn); 165 bool RemoveDeadArgumentsFromCallers(Function &Fn); 166 }; 167 } 168 169 170 char DAE::ID = 0; 171 INITIALIZE_PASS(DAE, "deadargelim", "Dead Argument Elimination", false, false) 172 173 namespace { 174 /// DAH - DeadArgumentHacking pass - Same as dead argument elimination, but 175 /// deletes arguments to functions which are external. This is only for use 176 /// by bugpoint. 177 struct DAH : public DAE { 178 static char ID; 179 DAH() : DAE(ID) {} 180 181 virtual bool ShouldHackArguments() const { return true; } 182 }; 183 } 184 185 char DAH::ID = 0; 186 INITIALIZE_PASS(DAH, "deadarghaX0r", 187 "Dead Argument Hacking (BUGPOINT USE ONLY; DO NOT USE)", 188 false, false) 189 190 /// createDeadArgEliminationPass - This pass removes arguments from functions 191 /// which are not used by the body of the function. 192 /// 193 ModulePass *llvm::createDeadArgEliminationPass() { return new DAE(); } 194 ModulePass *llvm::createDeadArgHackingPass() { return new DAH(); } 195 196 /// CollectFunctionDIs - Map each function in the module to its debug info 197 /// descriptor. 198 void DAE::CollectFunctionDIs(Module &M) { 199 FunctionDIs.clear(); 200 201 for (Module::named_metadata_iterator I = M.named_metadata_begin(), 202 E = M.named_metadata_end(); I != E; ++I) { 203 NamedMDNode &NMD = *I; 204 for (unsigned MDIndex = 0, MDNum = NMD.getNumOperands(); 205 MDIndex < MDNum; ++MDIndex) { 206 MDNode *Node = NMD.getOperand(MDIndex); 207 if (!DIDescriptor(Node).isCompileUnit()) 208 continue; 209 DICompileUnit CU(Node); 210 const DIArray &SPs = CU.getSubprograms(); 211 for (unsigned SPIndex = 0, SPNum = SPs.getNumElements(); 212 SPIndex < SPNum; ++SPIndex) { 213 DISubprogram SP(SPs.getElement(SPIndex)); 214 assert((!SP || SP.isSubprogram()) && 215 "A MDNode in subprograms of a CU should be null or a DISubprogram."); 216 if (!SP) 217 continue; 218 if (Function *F = SP.getFunction()) 219 FunctionDIs[F] = SP; 220 } 221 } 222 } 223 } 224 225 /// DeleteDeadVarargs - If this is an function that takes a ... list, and if 226 /// llvm.vastart is never called, the varargs list is dead for the function. 227 bool DAE::DeleteDeadVarargs(Function &Fn) { 228 assert(Fn.getFunctionType()->isVarArg() && "Function isn't varargs!"); 229 if (Fn.isDeclaration() || !Fn.hasLocalLinkage()) return false; 230 231 // Ensure that the function is only directly called. 232 if (Fn.hasAddressTaken()) 233 return false; 234 235 // Okay, we know we can transform this function if safe. Scan its body 236 // looking for calls to llvm.vastart. 237 for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) { 238 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 239 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 240 if (II->getIntrinsicID() == Intrinsic::vastart) 241 return false; 242 } 243 } 244 } 245 246 // If we get here, there are no calls to llvm.vastart in the function body, 247 // remove the "..." and adjust all the calls. 248 249 // Start by computing a new prototype for the function, which is the same as 250 // the old function, but doesn't have isVarArg set. 251 FunctionType *FTy = Fn.getFunctionType(); 252 253 std::vector<Type*> Params(FTy->param_begin(), FTy->param_end()); 254 FunctionType *NFTy = FunctionType::get(FTy->getReturnType(), 255 Params, false); 256 unsigned NumArgs = Params.size(); 257 258 // Create the new function body and insert it into the module... 259 Function *NF = Function::Create(NFTy, Fn.getLinkage()); 260 NF->copyAttributesFrom(&Fn); 261 Fn.getParent()->getFunctionList().insert(&Fn, NF); 262 NF->takeName(&Fn); 263 264 // Loop over all of the callers of the function, transforming the call sites 265 // to pass in a smaller number of arguments into the new function. 266 // 267 std::vector<Value*> Args; 268 for (Value::use_iterator I = Fn.use_begin(), E = Fn.use_end(); I != E; ) { 269 CallSite CS(*I++); 270 if (!CS) 271 continue; 272 Instruction *Call = CS.getInstruction(); 273 274 // Pass all the same arguments. 275 Args.assign(CS.arg_begin(), CS.arg_begin() + NumArgs); 276 277 // Drop any attributes that were on the vararg arguments. 278 AttributeSet PAL = CS.getAttributes(); 279 if (!PAL.isEmpty() && PAL.getSlotIndex(PAL.getNumSlots() - 1) > NumArgs) { 280 SmallVector<AttributeSet, 8> AttributesVec; 281 for (unsigned i = 0; PAL.getSlotIndex(i) <= NumArgs; ++i) 282 AttributesVec.push_back(PAL.getSlotAttributes(i)); 283 if (PAL.hasAttributes(AttributeSet::FunctionIndex)) 284 AttributesVec.push_back(AttributeSet::get(Fn.getContext(), 285 PAL.getFnAttributes())); 286 PAL = AttributeSet::get(Fn.getContext(), AttributesVec); 287 } 288 289 Instruction *New; 290 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 291 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 292 Args, "", Call); 293 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 294 cast<InvokeInst>(New)->setAttributes(PAL); 295 } else { 296 New = CallInst::Create(NF, Args, "", Call); 297 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 298 cast<CallInst>(New)->setAttributes(PAL); 299 if (cast<CallInst>(Call)->isTailCall()) 300 cast<CallInst>(New)->setTailCall(); 301 } 302 New->setDebugLoc(Call->getDebugLoc()); 303 304 Args.clear(); 305 306 if (!Call->use_empty()) 307 Call->replaceAllUsesWith(New); 308 309 New->takeName(Call); 310 311 // Finally, remove the old call from the program, reducing the use-count of 312 // F. 313 Call->eraseFromParent(); 314 } 315 316 // Since we have now created the new function, splice the body of the old 317 // function right into the new function, leaving the old rotting hulk of the 318 // function empty. 319 NF->getBasicBlockList().splice(NF->begin(), Fn.getBasicBlockList()); 320 321 // Loop over the argument list, transferring uses of the old arguments over to 322 // the new arguments, also transferring over the names as well. While we're at 323 // it, remove the dead arguments from the DeadArguments list. 324 // 325 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(), 326 I2 = NF->arg_begin(); I != E; ++I, ++I2) { 327 // Move the name and users over to the new version. 328 I->replaceAllUsesWith(I2); 329 I2->takeName(I); 330 } 331 332 // Patch the pointer to LLVM function in debug info descriptor. 333 FunctionDIMap::iterator DI = FunctionDIs.find(&Fn); 334 if (DI != FunctionDIs.end()) 335 DI->second.replaceFunction(NF); 336 337 // Fix up any BlockAddresses that refer to the function. 338 Fn.replaceAllUsesWith(ConstantExpr::getBitCast(NF, Fn.getType())); 339 // Delete the bitcast that we just created, so that NF does not 340 // appear to be address-taken. 341 NF->removeDeadConstantUsers(); 342 // Finally, nuke the old function. 343 Fn.eraseFromParent(); 344 return true; 345 } 346 347 /// RemoveDeadArgumentsFromCallers - Checks if the given function has any 348 /// arguments that are unused, and changes the caller parameters to be undefined 349 /// instead. 350 bool DAE::RemoveDeadArgumentsFromCallers(Function &Fn) 351 { 352 if (Fn.isDeclaration() || Fn.mayBeOverridden()) 353 return false; 354 355 // Functions with local linkage should already have been handled, except the 356 // fragile (variadic) ones which we can improve here. 357 if (Fn.hasLocalLinkage() && !Fn.getFunctionType()->isVarArg()) 358 return false; 359 360 if (Fn.use_empty()) 361 return false; 362 363 SmallVector<unsigned, 8> UnusedArgs; 364 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(); 365 I != E; ++I) { 366 Argument *Arg = I; 367 368 if (Arg->use_empty() && !Arg->hasByValAttr()) 369 UnusedArgs.push_back(Arg->getArgNo()); 370 } 371 372 if (UnusedArgs.empty()) 373 return false; 374 375 bool Changed = false; 376 377 for (Function::use_iterator I = Fn.use_begin(), E = Fn.use_end(); 378 I != E; ++I) { 379 CallSite CS(*I); 380 if (!CS || !CS.isCallee(I)) 381 continue; 382 383 // Now go through all unused args and replace them with "undef". 384 for (unsigned I = 0, E = UnusedArgs.size(); I != E; ++I) { 385 unsigned ArgNo = UnusedArgs[I]; 386 387 Value *Arg = CS.getArgument(ArgNo); 388 CS.setArgument(ArgNo, UndefValue::get(Arg->getType())); 389 ++NumArgumentsReplacedWithUndef; 390 Changed = true; 391 } 392 } 393 394 return Changed; 395 } 396 397 /// Convenience function that returns the number of return values. It returns 0 398 /// for void functions and 1 for functions not returning a struct. It returns 399 /// the number of struct elements for functions returning a struct. 400 static unsigned NumRetVals(const Function *F) { 401 if (F->getReturnType()->isVoidTy()) 402 return 0; 403 else if (StructType *STy = dyn_cast<StructType>(F->getReturnType())) 404 return STy->getNumElements(); 405 else 406 return 1; 407 } 408 409 /// MarkIfNotLive - This checks Use for liveness in LiveValues. If Use is not 410 /// live, it adds Use to the MaybeLiveUses argument. Returns the determined 411 /// liveness of Use. 412 DAE::Liveness DAE::MarkIfNotLive(RetOrArg Use, UseVector &MaybeLiveUses) { 413 // We're live if our use or its Function is already marked as live. 414 if (LiveFunctions.count(Use.F) || LiveValues.count(Use)) 415 return Live; 416 417 // We're maybe live otherwise, but remember that we must become live if 418 // Use becomes live. 419 MaybeLiveUses.push_back(Use); 420 return MaybeLive; 421 } 422 423 424 /// SurveyUse - This looks at a single use of an argument or return value 425 /// and determines if it should be alive or not. Adds this use to MaybeLiveUses 426 /// if it causes the used value to become MaybeLive. 427 /// 428 /// RetValNum is the return value number to use when this use is used in a 429 /// return instruction. This is used in the recursion, you should always leave 430 /// it at 0. 431 DAE::Liveness DAE::SurveyUse(Value::const_use_iterator U, 432 UseVector &MaybeLiveUses, unsigned RetValNum) { 433 const User *V = *U; 434 if (const ReturnInst *RI = dyn_cast<ReturnInst>(V)) { 435 // The value is returned from a function. It's only live when the 436 // function's return value is live. We use RetValNum here, for the case 437 // that U is really a use of an insertvalue instruction that uses the 438 // original Use. 439 RetOrArg Use = CreateRet(RI->getParent()->getParent(), RetValNum); 440 // We might be live, depending on the liveness of Use. 441 return MarkIfNotLive(Use, MaybeLiveUses); 442 } 443 if (const InsertValueInst *IV = dyn_cast<InsertValueInst>(V)) { 444 if (U.getOperandNo() != InsertValueInst::getAggregateOperandIndex() 445 && IV->hasIndices()) 446 // The use we are examining is inserted into an aggregate. Our liveness 447 // depends on all uses of that aggregate, but if it is used as a return 448 // value, only index at which we were inserted counts. 449 RetValNum = *IV->idx_begin(); 450 451 // Note that if we are used as the aggregate operand to the insertvalue, 452 // we don't change RetValNum, but do survey all our uses. 453 454 Liveness Result = MaybeLive; 455 for (Value::const_use_iterator I = IV->use_begin(), 456 E = V->use_end(); I != E; ++I) { 457 Result = SurveyUse(I, MaybeLiveUses, RetValNum); 458 if (Result == Live) 459 break; 460 } 461 return Result; 462 } 463 464 if (ImmutableCallSite CS = V) { 465 const Function *F = CS.getCalledFunction(); 466 if (F) { 467 // Used in a direct call. 468 469 // Find the argument number. We know for sure that this use is an 470 // argument, since if it was the function argument this would be an 471 // indirect call and the we know can't be looking at a value of the 472 // label type (for the invoke instruction). 473 unsigned ArgNo = CS.getArgumentNo(U); 474 475 if (ArgNo >= F->getFunctionType()->getNumParams()) 476 // The value is passed in through a vararg! Must be live. 477 return Live; 478 479 assert(CS.getArgument(ArgNo) 480 == CS->getOperand(U.getOperandNo()) 481 && "Argument is not where we expected it"); 482 483 // Value passed to a normal call. It's only live when the corresponding 484 // argument to the called function turns out live. 485 RetOrArg Use = CreateArg(F, ArgNo); 486 return MarkIfNotLive(Use, MaybeLiveUses); 487 } 488 } 489 // Used in any other way? Value must be live. 490 return Live; 491 } 492 493 /// SurveyUses - This looks at all the uses of the given value 494 /// Returns the Liveness deduced from the uses of this value. 495 /// 496 /// Adds all uses that cause the result to be MaybeLive to MaybeLiveRetUses. If 497 /// the result is Live, MaybeLiveUses might be modified but its content should 498 /// be ignored (since it might not be complete). 499 DAE::Liveness DAE::SurveyUses(const Value *V, UseVector &MaybeLiveUses) { 500 // Assume it's dead (which will only hold if there are no uses at all..). 501 Liveness Result = MaybeLive; 502 // Check each use. 503 for (Value::const_use_iterator I = V->use_begin(), 504 E = V->use_end(); I != E; ++I) { 505 Result = SurveyUse(I, MaybeLiveUses); 506 if (Result == Live) 507 break; 508 } 509 return Result; 510 } 511 512 // SurveyFunction - This performs the initial survey of the specified function, 513 // checking out whether or not it uses any of its incoming arguments or whether 514 // any callers use the return value. This fills in the LiveValues set and Uses 515 // map. 516 // 517 // We consider arguments of non-internal functions to be intrinsically alive as 518 // well as arguments to functions which have their "address taken". 519 // 520 void DAE::SurveyFunction(const Function &F) { 521 unsigned RetCount = NumRetVals(&F); 522 // Assume all return values are dead 523 typedef SmallVector<Liveness, 5> RetVals; 524 RetVals RetValLiveness(RetCount, MaybeLive); 525 526 typedef SmallVector<UseVector, 5> RetUses; 527 // These vectors map each return value to the uses that make it MaybeLive, so 528 // we can add those to the Uses map if the return value really turns out to be 529 // MaybeLive. Initialized to a list of RetCount empty lists. 530 RetUses MaybeLiveRetUses(RetCount); 531 532 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 533 if (const ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) 534 if (RI->getNumOperands() != 0 && RI->getOperand(0)->getType() 535 != F.getFunctionType()->getReturnType()) { 536 // We don't support old style multiple return values. 537 MarkLive(F); 538 return; 539 } 540 541 if (!F.hasLocalLinkage() && (!ShouldHackArguments() || F.isIntrinsic())) { 542 MarkLive(F); 543 return; 544 } 545 546 DEBUG(dbgs() << "DAE - Inspecting callers for fn: " << F.getName() << "\n"); 547 // Keep track of the number of live retvals, so we can skip checks once all 548 // of them turn out to be live. 549 unsigned NumLiveRetVals = 0; 550 Type *STy = dyn_cast<StructType>(F.getReturnType()); 551 // Loop all uses of the function. 552 for (Value::const_use_iterator I = F.use_begin(), E = F.use_end(); 553 I != E; ++I) { 554 // If the function is PASSED IN as an argument, its address has been 555 // taken. 556 ImmutableCallSite CS(*I); 557 if (!CS || !CS.isCallee(I)) { 558 MarkLive(F); 559 return; 560 } 561 562 // If this use is anything other than a call site, the function is alive. 563 const Instruction *TheCall = CS.getInstruction(); 564 if (!TheCall) { // Not a direct call site? 565 MarkLive(F); 566 return; 567 } 568 569 // If we end up here, we are looking at a direct call to our function. 570 571 // Now, check how our return value(s) is/are used in this caller. Don't 572 // bother checking return values if all of them are live already. 573 if (NumLiveRetVals != RetCount) { 574 if (STy) { 575 // Check all uses of the return value. 576 for (Value::const_use_iterator I = TheCall->use_begin(), 577 E = TheCall->use_end(); I != E; ++I) { 578 const ExtractValueInst *Ext = dyn_cast<ExtractValueInst>(*I); 579 if (Ext && Ext->hasIndices()) { 580 // This use uses a part of our return value, survey the uses of 581 // that part and store the results for this index only. 582 unsigned Idx = *Ext->idx_begin(); 583 if (RetValLiveness[Idx] != Live) { 584 RetValLiveness[Idx] = SurveyUses(Ext, MaybeLiveRetUses[Idx]); 585 if (RetValLiveness[Idx] == Live) 586 NumLiveRetVals++; 587 } 588 } else { 589 // Used by something else than extractvalue. Mark all return 590 // values as live. 591 for (unsigned i = 0; i != RetCount; ++i ) 592 RetValLiveness[i] = Live; 593 NumLiveRetVals = RetCount; 594 break; 595 } 596 } 597 } else { 598 // Single return value 599 RetValLiveness[0] = SurveyUses(TheCall, MaybeLiveRetUses[0]); 600 if (RetValLiveness[0] == Live) 601 NumLiveRetVals = RetCount; 602 } 603 } 604 } 605 606 // Now we've inspected all callers, record the liveness of our return values. 607 for (unsigned i = 0; i != RetCount; ++i) 608 MarkValue(CreateRet(&F, i), RetValLiveness[i], MaybeLiveRetUses[i]); 609 610 DEBUG(dbgs() << "DAE - Inspecting args for fn: " << F.getName() << "\n"); 611 612 // Now, check all of our arguments. 613 unsigned i = 0; 614 UseVector MaybeLiveArgUses; 615 for (Function::const_arg_iterator AI = F.arg_begin(), 616 E = F.arg_end(); AI != E; ++AI, ++i) { 617 Liveness Result; 618 if (F.getFunctionType()->isVarArg()) { 619 // Variadic functions will already have a va_arg function expanded inside 620 // them, making them potentially very sensitive to ABI changes resulting 621 // from removing arguments entirely, so don't. For example AArch64 handles 622 // register and stack HFAs very differently, and this is reflected in the 623 // IR which has already been generated. 624 Result = Live; 625 } else { 626 // See what the effect of this use is (recording any uses that cause 627 // MaybeLive in MaybeLiveArgUses). 628 Result = SurveyUses(AI, MaybeLiveArgUses); 629 } 630 631 // Mark the result. 632 MarkValue(CreateArg(&F, i), Result, MaybeLiveArgUses); 633 // Clear the vector again for the next iteration. 634 MaybeLiveArgUses.clear(); 635 } 636 } 637 638 /// MarkValue - This function marks the liveness of RA depending on L. If L is 639 /// MaybeLive, it also takes all uses in MaybeLiveUses and records them in Uses, 640 /// such that RA will be marked live if any use in MaybeLiveUses gets marked 641 /// live later on. 642 void DAE::MarkValue(const RetOrArg &RA, Liveness L, 643 const UseVector &MaybeLiveUses) { 644 switch (L) { 645 case Live: MarkLive(RA); break; 646 case MaybeLive: 647 { 648 // Note any uses of this value, so this return value can be 649 // marked live whenever one of the uses becomes live. 650 for (UseVector::const_iterator UI = MaybeLiveUses.begin(), 651 UE = MaybeLiveUses.end(); UI != UE; ++UI) 652 Uses.insert(std::make_pair(*UI, RA)); 653 break; 654 } 655 } 656 } 657 658 /// MarkLive - Mark the given Function as alive, meaning that it cannot be 659 /// changed in any way. Additionally, 660 /// mark any values that are used as this function's parameters or by its return 661 /// values (according to Uses) live as well. 662 void DAE::MarkLive(const Function &F) { 663 DEBUG(dbgs() << "DAE - Intrinsically live fn: " << F.getName() << "\n"); 664 // Mark the function as live. 665 LiveFunctions.insert(&F); 666 // Mark all arguments as live. 667 for (unsigned i = 0, e = F.arg_size(); i != e; ++i) 668 PropagateLiveness(CreateArg(&F, i)); 669 // Mark all return values as live. 670 for (unsigned i = 0, e = NumRetVals(&F); i != e; ++i) 671 PropagateLiveness(CreateRet(&F, i)); 672 } 673 674 /// MarkLive - Mark the given return value or argument as live. Additionally, 675 /// mark any values that are used by this value (according to Uses) live as 676 /// well. 677 void DAE::MarkLive(const RetOrArg &RA) { 678 if (LiveFunctions.count(RA.F)) 679 return; // Function was already marked Live. 680 681 if (!LiveValues.insert(RA).second) 682 return; // We were already marked Live. 683 684 DEBUG(dbgs() << "DAE - Marking " << RA.getDescription() << " live\n"); 685 PropagateLiveness(RA); 686 } 687 688 /// PropagateLiveness - Given that RA is a live value, propagate it's liveness 689 /// to any other values it uses (according to Uses). 690 void DAE::PropagateLiveness(const RetOrArg &RA) { 691 // We don't use upper_bound (or equal_range) here, because our recursive call 692 // to ourselves is likely to cause the upper_bound (which is the first value 693 // not belonging to RA) to become erased and the iterator invalidated. 694 UseMap::iterator Begin = Uses.lower_bound(RA); 695 UseMap::iterator E = Uses.end(); 696 UseMap::iterator I; 697 for (I = Begin; I != E && I->first == RA; ++I) 698 MarkLive(I->second); 699 700 // Erase RA from the Uses map (from the lower bound to wherever we ended up 701 // after the loop). 702 Uses.erase(Begin, I); 703 } 704 705 // RemoveDeadStuffFromFunction - Remove any arguments and return values from F 706 // that are not in LiveValues. Transform the function and all of the callees of 707 // the function to not have these arguments and return values. 708 // 709 bool DAE::RemoveDeadStuffFromFunction(Function *F) { 710 // Don't modify fully live functions 711 if (LiveFunctions.count(F)) 712 return false; 713 714 // Start by computing a new prototype for the function, which is the same as 715 // the old function, but has fewer arguments and a different return type. 716 FunctionType *FTy = F->getFunctionType(); 717 std::vector<Type*> Params; 718 719 // Keep track of if we have a live 'returned' argument 720 bool HasLiveReturnedArg = false; 721 722 // Set up to build a new list of parameter attributes. 723 SmallVector<AttributeSet, 8> AttributesVec; 724 const AttributeSet &PAL = F->getAttributes(); 725 726 // Remember which arguments are still alive. 727 SmallVector<bool, 10> ArgAlive(FTy->getNumParams(), false); 728 // Construct the new parameter list from non-dead arguments. Also construct 729 // a new set of parameter attributes to correspond. Skip the first parameter 730 // attribute, since that belongs to the return value. 731 unsigned i = 0; 732 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 733 I != E; ++I, ++i) { 734 RetOrArg Arg = CreateArg(F, i); 735 if (LiveValues.erase(Arg)) { 736 Params.push_back(I->getType()); 737 ArgAlive[i] = true; 738 739 // Get the original parameter attributes (skipping the first one, that is 740 // for the return value. 741 if (PAL.hasAttributes(i + 1)) { 742 AttrBuilder B(PAL, i + 1); 743 if (B.contains(Attribute::Returned)) 744 HasLiveReturnedArg = true; 745 AttributesVec. 746 push_back(AttributeSet::get(F->getContext(), Params.size(), B)); 747 } 748 } else { 749 ++NumArgumentsEliminated; 750 DEBUG(dbgs() << "DAE - Removing argument " << i << " (" << I->getName() 751 << ") from " << F->getName() << "\n"); 752 } 753 } 754 755 // Find out the new return value. 756 Type *RetTy = FTy->getReturnType(); 757 Type *NRetTy = NULL; 758 unsigned RetCount = NumRetVals(F); 759 760 // -1 means unused, other numbers are the new index 761 SmallVector<int, 5> NewRetIdxs(RetCount, -1); 762 std::vector<Type*> RetTypes; 763 764 // If there is a function with a live 'returned' argument but a dead return 765 // value, then there are two possible actions: 766 // 1) Eliminate the return value and take off the 'returned' attribute on the 767 // argument. 768 // 2) Retain the 'returned' attribute and treat the return value (but not the 769 // entire function) as live so that it is not eliminated. 770 // 771 // It's not clear in the general case which option is more profitable because, 772 // even in the absence of explicit uses of the return value, code generation 773 // is free to use the 'returned' attribute to do things like eliding 774 // save/restores of registers across calls. Whether or not this happens is 775 // target and ABI-specific as well as depending on the amount of register 776 // pressure, so there's no good way for an IR-level pass to figure this out. 777 // 778 // Fortunately, the only places where 'returned' is currently generated by 779 // the FE are places where 'returned' is basically free and almost always a 780 // performance win, so the second option can just be used always for now. 781 // 782 // This should be revisited if 'returned' is ever applied more liberally. 783 if (RetTy->isVoidTy() || HasLiveReturnedArg) { 784 NRetTy = RetTy; 785 } else { 786 StructType *STy = dyn_cast<StructType>(RetTy); 787 if (STy) 788 // Look at each of the original return values individually. 789 for (unsigned i = 0; i != RetCount; ++i) { 790 RetOrArg Ret = CreateRet(F, i); 791 if (LiveValues.erase(Ret)) { 792 RetTypes.push_back(STy->getElementType(i)); 793 NewRetIdxs[i] = RetTypes.size() - 1; 794 } else { 795 ++NumRetValsEliminated; 796 DEBUG(dbgs() << "DAE - Removing return value " << i << " from " 797 << F->getName() << "\n"); 798 } 799 } 800 else 801 // We used to return a single value. 802 if (LiveValues.erase(CreateRet(F, 0))) { 803 RetTypes.push_back(RetTy); 804 NewRetIdxs[0] = 0; 805 } else { 806 DEBUG(dbgs() << "DAE - Removing return value from " << F->getName() 807 << "\n"); 808 ++NumRetValsEliminated; 809 } 810 if (RetTypes.size() > 1) 811 // More than one return type? Return a struct with them. Also, if we used 812 // to return a struct and didn't change the number of return values, 813 // return a struct again. This prevents changing {something} into 814 // something and {} into void. 815 // Make the new struct packed if we used to return a packed struct 816 // already. 817 NRetTy = StructType::get(STy->getContext(), RetTypes, STy->isPacked()); 818 else if (RetTypes.size() == 1) 819 // One return type? Just a simple value then, but only if we didn't use to 820 // return a struct with that simple value before. 821 NRetTy = RetTypes.front(); 822 else if (RetTypes.size() == 0) 823 // No return types? Make it void, but only if we didn't use to return {}. 824 NRetTy = Type::getVoidTy(F->getContext()); 825 } 826 827 assert(NRetTy && "No new return type found?"); 828 829 // The existing function return attributes. 830 AttributeSet RAttrs = PAL.getRetAttributes(); 831 832 // Remove any incompatible attributes, but only if we removed all return 833 // values. Otherwise, ensure that we don't have any conflicting attributes 834 // here. Currently, this should not be possible, but special handling might be 835 // required when new return value attributes are added. 836 if (NRetTy->isVoidTy()) 837 RAttrs = 838 AttributeSet::get(NRetTy->getContext(), AttributeSet::ReturnIndex, 839 AttrBuilder(RAttrs, AttributeSet::ReturnIndex). 840 removeAttributes(AttributeFuncs:: 841 typeIncompatible(NRetTy, AttributeSet::ReturnIndex), 842 AttributeSet::ReturnIndex)); 843 else 844 assert(!AttrBuilder(RAttrs, AttributeSet::ReturnIndex). 845 hasAttributes(AttributeFuncs:: 846 typeIncompatible(NRetTy, AttributeSet::ReturnIndex), 847 AttributeSet::ReturnIndex) && 848 "Return attributes no longer compatible?"); 849 850 if (RAttrs.hasAttributes(AttributeSet::ReturnIndex)) 851 AttributesVec.push_back(AttributeSet::get(NRetTy->getContext(), RAttrs)); 852 853 if (PAL.hasAttributes(AttributeSet::FunctionIndex)) 854 AttributesVec.push_back(AttributeSet::get(F->getContext(), 855 PAL.getFnAttributes())); 856 857 // Reconstruct the AttributesList based on the vector we constructed. 858 AttributeSet NewPAL = AttributeSet::get(F->getContext(), AttributesVec); 859 860 // Create the new function type based on the recomputed parameters. 861 FunctionType *NFTy = FunctionType::get(NRetTy, Params, FTy->isVarArg()); 862 863 // No change? 864 if (NFTy == FTy) 865 return false; 866 867 // Create the new function body and insert it into the module... 868 Function *NF = Function::Create(NFTy, F->getLinkage()); 869 NF->copyAttributesFrom(F); 870 NF->setAttributes(NewPAL); 871 // Insert the new function before the old function, so we won't be processing 872 // it again. 873 F->getParent()->getFunctionList().insert(F, NF); 874 NF->takeName(F); 875 876 // Loop over all of the callers of the function, transforming the call sites 877 // to pass in a smaller number of arguments into the new function. 878 // 879 std::vector<Value*> Args; 880 while (!F->use_empty()) { 881 CallSite CS(F->use_back()); 882 Instruction *Call = CS.getInstruction(); 883 884 AttributesVec.clear(); 885 const AttributeSet &CallPAL = CS.getAttributes(); 886 887 // The call return attributes. 888 AttributeSet RAttrs = CallPAL.getRetAttributes(); 889 890 // Adjust in case the function was changed to return void. 891 RAttrs = 892 AttributeSet::get(NF->getContext(), AttributeSet::ReturnIndex, 893 AttrBuilder(RAttrs, AttributeSet::ReturnIndex). 894 removeAttributes(AttributeFuncs:: 895 typeIncompatible(NF->getReturnType(), 896 AttributeSet::ReturnIndex), 897 AttributeSet::ReturnIndex)); 898 if (RAttrs.hasAttributes(AttributeSet::ReturnIndex)) 899 AttributesVec.push_back(AttributeSet::get(NF->getContext(), RAttrs)); 900 901 // Declare these outside of the loops, so we can reuse them for the second 902 // loop, which loops the varargs. 903 CallSite::arg_iterator I = CS.arg_begin(); 904 unsigned i = 0; 905 // Loop over those operands, corresponding to the normal arguments to the 906 // original function, and add those that are still alive. 907 for (unsigned e = FTy->getNumParams(); i != e; ++I, ++i) 908 if (ArgAlive[i]) { 909 Args.push_back(*I); 910 // Get original parameter attributes, but skip return attributes. 911 if (CallPAL.hasAttributes(i + 1)) { 912 AttrBuilder B(CallPAL, i + 1); 913 // If the return type has changed, then get rid of 'returned' on the 914 // call site. The alternative is to make all 'returned' attributes on 915 // call sites keep the return value alive just like 'returned' 916 // attributes on function declaration but it's less clearly a win 917 // and this is not an expected case anyway 918 if (NRetTy != RetTy && B.contains(Attribute::Returned)) 919 B.removeAttribute(Attribute::Returned); 920 AttributesVec. 921 push_back(AttributeSet::get(F->getContext(), Args.size(), B)); 922 } 923 } 924 925 // Push any varargs arguments on the list. Don't forget their attributes. 926 for (CallSite::arg_iterator E = CS.arg_end(); I != E; ++I, ++i) { 927 Args.push_back(*I); 928 if (CallPAL.hasAttributes(i + 1)) { 929 AttrBuilder B(CallPAL, i + 1); 930 AttributesVec. 931 push_back(AttributeSet::get(F->getContext(), Args.size(), B)); 932 } 933 } 934 935 if (CallPAL.hasAttributes(AttributeSet::FunctionIndex)) 936 AttributesVec.push_back(AttributeSet::get(Call->getContext(), 937 CallPAL.getFnAttributes())); 938 939 // Reconstruct the AttributesList based on the vector we constructed. 940 AttributeSet NewCallPAL = AttributeSet::get(F->getContext(), AttributesVec); 941 942 Instruction *New; 943 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 944 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 945 Args, "", Call); 946 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 947 cast<InvokeInst>(New)->setAttributes(NewCallPAL); 948 } else { 949 New = CallInst::Create(NF, Args, "", Call); 950 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 951 cast<CallInst>(New)->setAttributes(NewCallPAL); 952 if (cast<CallInst>(Call)->isTailCall()) 953 cast<CallInst>(New)->setTailCall(); 954 } 955 New->setDebugLoc(Call->getDebugLoc()); 956 957 Args.clear(); 958 959 if (!Call->use_empty()) { 960 if (New->getType() == Call->getType()) { 961 // Return type not changed? Just replace users then. 962 Call->replaceAllUsesWith(New); 963 New->takeName(Call); 964 } else if (New->getType()->isVoidTy()) { 965 // Our return value has uses, but they will get removed later on. 966 // Replace by null for now. 967 if (!Call->getType()->isX86_MMXTy()) 968 Call->replaceAllUsesWith(Constant::getNullValue(Call->getType())); 969 } else { 970 assert(RetTy->isStructTy() && 971 "Return type changed, but not into a void. The old return type" 972 " must have been a struct!"); 973 Instruction *InsertPt = Call; 974 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 975 BasicBlock::iterator IP = II->getNormalDest()->begin(); 976 while (isa<PHINode>(IP)) ++IP; 977 InsertPt = IP; 978 } 979 980 // We used to return a struct. Instead of doing smart stuff with all the 981 // uses of this struct, we will just rebuild it using 982 // extract/insertvalue chaining and let instcombine clean that up. 983 // 984 // Start out building up our return value from undef 985 Value *RetVal = UndefValue::get(RetTy); 986 for (unsigned i = 0; i != RetCount; ++i) 987 if (NewRetIdxs[i] != -1) { 988 Value *V; 989 if (RetTypes.size() > 1) 990 // We are still returning a struct, so extract the value from our 991 // return value 992 V = ExtractValueInst::Create(New, NewRetIdxs[i], "newret", 993 InsertPt); 994 else 995 // We are now returning a single element, so just insert that 996 V = New; 997 // Insert the value at the old position 998 RetVal = InsertValueInst::Create(RetVal, V, i, "oldret", InsertPt); 999 } 1000 // Now, replace all uses of the old call instruction with the return 1001 // struct we built 1002 Call->replaceAllUsesWith(RetVal); 1003 New->takeName(Call); 1004 } 1005 } 1006 1007 // Finally, remove the old call from the program, reducing the use-count of 1008 // F. 1009 Call->eraseFromParent(); 1010 } 1011 1012 // Since we have now created the new function, splice the body of the old 1013 // function right into the new function, leaving the old rotting hulk of the 1014 // function empty. 1015 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 1016 1017 // Loop over the argument list, transferring uses of the old arguments over to 1018 // the new arguments, also transferring over the names as well. 1019 i = 0; 1020 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), 1021 I2 = NF->arg_begin(); I != E; ++I, ++i) 1022 if (ArgAlive[i]) { 1023 // If this is a live argument, move the name and users over to the new 1024 // version. 1025 I->replaceAllUsesWith(I2); 1026 I2->takeName(I); 1027 ++I2; 1028 } else { 1029 // If this argument is dead, replace any uses of it with null constants 1030 // (these are guaranteed to become unused later on). 1031 if (!I->getType()->isX86_MMXTy()) 1032 I->replaceAllUsesWith(Constant::getNullValue(I->getType())); 1033 } 1034 1035 // If we change the return value of the function we must rewrite any return 1036 // instructions. Check this now. 1037 if (F->getReturnType() != NF->getReturnType()) 1038 for (Function::iterator BB = NF->begin(), E = NF->end(); BB != E; ++BB) 1039 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 1040 Value *RetVal; 1041 1042 if (NFTy->getReturnType()->isVoidTy()) { 1043 RetVal = 0; 1044 } else { 1045 assert (RetTy->isStructTy()); 1046 // The original return value was a struct, insert 1047 // extractvalue/insertvalue chains to extract only the values we need 1048 // to return and insert them into our new result. 1049 // This does generate messy code, but we'll let it to instcombine to 1050 // clean that up. 1051 Value *OldRet = RI->getOperand(0); 1052 // Start out building up our return value from undef 1053 RetVal = UndefValue::get(NRetTy); 1054 for (unsigned i = 0; i != RetCount; ++i) 1055 if (NewRetIdxs[i] != -1) { 1056 ExtractValueInst *EV = ExtractValueInst::Create(OldRet, i, 1057 "oldret", RI); 1058 if (RetTypes.size() > 1) { 1059 // We're still returning a struct, so reinsert the value into 1060 // our new return value at the new index 1061 1062 RetVal = InsertValueInst::Create(RetVal, EV, NewRetIdxs[i], 1063 "newret", RI); 1064 } else { 1065 // We are now only returning a simple value, so just return the 1066 // extracted value. 1067 RetVal = EV; 1068 } 1069 } 1070 } 1071 // Replace the return instruction with one returning the new return 1072 // value (possibly 0 if we became void). 1073 ReturnInst::Create(F->getContext(), RetVal, RI); 1074 BB->getInstList().erase(RI); 1075 } 1076 1077 // Patch the pointer to LLVM function in debug info descriptor. 1078 FunctionDIMap::iterator DI = FunctionDIs.find(F); 1079 if (DI != FunctionDIs.end()) 1080 DI->second.replaceFunction(NF); 1081 1082 // Now that the old function is dead, delete it. 1083 F->eraseFromParent(); 1084 1085 return true; 1086 } 1087 1088 bool DAE::runOnModule(Module &M) { 1089 bool Changed = false; 1090 1091 // Collect debug info descriptors for functions. 1092 CollectFunctionDIs(M); 1093 1094 // First pass: Do a simple check to see if any functions can have their "..." 1095 // removed. We can do this if they never call va_start. This loop cannot be 1096 // fused with the next loop, because deleting a function invalidates 1097 // information computed while surveying other functions. 1098 DEBUG(dbgs() << "DAE - Deleting dead varargs\n"); 1099 for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { 1100 Function &F = *I++; 1101 if (F.getFunctionType()->isVarArg()) 1102 Changed |= DeleteDeadVarargs(F); 1103 } 1104 1105 // Second phase:loop through the module, determining which arguments are live. 1106 // We assume all arguments are dead unless proven otherwise (allowing us to 1107 // determine that dead arguments passed into recursive functions are dead). 1108 // 1109 DEBUG(dbgs() << "DAE - Determining liveness\n"); 1110 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) 1111 SurveyFunction(*I); 1112 1113 // Now, remove all dead arguments and return values from each function in 1114 // turn. 1115 for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { 1116 // Increment now, because the function will probably get removed (ie. 1117 // replaced by a new one). 1118 Function *F = I++; 1119 Changed |= RemoveDeadStuffFromFunction(F); 1120 } 1121 1122 // Finally, look for any unused parameters in functions with non-local 1123 // linkage and replace the passed in parameters with undef. 1124 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) { 1125 Function& F = *I; 1126 1127 Changed |= RemoveDeadArgumentsFromCallers(F); 1128 } 1129 1130 return Changed; 1131 } 1132