DeadArgumentElimination.cpp revision 263508
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> 41using namespace llvm; 42 43STATISTIC(NumArgumentsEliminated, "Number of unread args removed"); 44STATISTIC(NumRetValsEliminated , "Number of unused return values removed"); 45STATISTIC(NumArgumentsReplacedWithUndef, 46 "Number of unread args replaced with undef"); 47namespace { 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 170char DAE::ID = 0; 171INITIALIZE_PASS(DAE, "deadargelim", "Dead Argument Elimination", false, false) 172 173namespace { 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 185char DAH::ID = 0; 186INITIALIZE_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/// 193ModulePass *llvm::createDeadArgEliminationPass() { return new DAE(); } 194ModulePass *llvm::createDeadArgHackingPass() { return new DAH(); } 195 196/// CollectFunctionDIs - Map each function in the module to its debug info 197/// descriptor. 198void 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. 227bool 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. 350bool 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 a function seen at compile time is not necessarily the one linked to 361 // the binary being built, it is illegal to change the actual arguments 362 // passed to it. These functions can be captured by isWeakForLinker(). 363 // *NOTE* that mayBeOverridden() is insufficient for this purpose as it 364 // doesn't include linkage types like AvailableExternallyLinkage and 365 // LinkOnceODRLinkage. Take link_odr* as an example, it indicates a set of 366 // *EQUIVALENT* globals that can be merged at link-time. However, the 367 // semantic of *EQUIVALENT*-functions includes parameters. Changing 368 // parameters breaks this assumption. 369 // 370 if (Fn.isWeakForLinker()) 371 return false; 372 373 if (Fn.use_empty()) 374 return false; 375 376 SmallVector<unsigned, 8> UnusedArgs; 377 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(); 378 I != E; ++I) { 379 Argument *Arg = I; 380 381 if (Arg->use_empty() && !Arg->hasByValAttr()) 382 UnusedArgs.push_back(Arg->getArgNo()); 383 } 384 385 if (UnusedArgs.empty()) 386 return false; 387 388 bool Changed = false; 389 390 for (Function::use_iterator I = Fn.use_begin(), E = Fn.use_end(); 391 I != E; ++I) { 392 CallSite CS(*I); 393 if (!CS || !CS.isCallee(I)) 394 continue; 395 396 // Now go through all unused args and replace them with "undef". 397 for (unsigned I = 0, E = UnusedArgs.size(); I != E; ++I) { 398 unsigned ArgNo = UnusedArgs[I]; 399 400 Value *Arg = CS.getArgument(ArgNo); 401 CS.setArgument(ArgNo, UndefValue::get(Arg->getType())); 402 ++NumArgumentsReplacedWithUndef; 403 Changed = true; 404 } 405 } 406 407 return Changed; 408} 409 410/// Convenience function that returns the number of return values. It returns 0 411/// for void functions and 1 for functions not returning a struct. It returns 412/// the number of struct elements for functions returning a struct. 413static unsigned NumRetVals(const Function *F) { 414 if (F->getReturnType()->isVoidTy()) 415 return 0; 416 else if (StructType *STy = dyn_cast<StructType>(F->getReturnType())) 417 return STy->getNumElements(); 418 else 419 return 1; 420} 421 422/// MarkIfNotLive - This checks Use for liveness in LiveValues. If Use is not 423/// live, it adds Use to the MaybeLiveUses argument. Returns the determined 424/// liveness of Use. 425DAE::Liveness DAE::MarkIfNotLive(RetOrArg Use, UseVector &MaybeLiveUses) { 426 // We're live if our use or its Function is already marked as live. 427 if (LiveFunctions.count(Use.F) || LiveValues.count(Use)) 428 return Live; 429 430 // We're maybe live otherwise, but remember that we must become live if 431 // Use becomes live. 432 MaybeLiveUses.push_back(Use); 433 return MaybeLive; 434} 435 436 437/// SurveyUse - This looks at a single use of an argument or return value 438/// and determines if it should be alive or not. Adds this use to MaybeLiveUses 439/// if it causes the used value to become MaybeLive. 440/// 441/// RetValNum is the return value number to use when this use is used in a 442/// return instruction. This is used in the recursion, you should always leave 443/// it at 0. 444DAE::Liveness DAE::SurveyUse(Value::const_use_iterator U, 445 UseVector &MaybeLiveUses, unsigned RetValNum) { 446 const User *V = *U; 447 if (const ReturnInst *RI = dyn_cast<ReturnInst>(V)) { 448 // The value is returned from a function. It's only live when the 449 // function's return value is live. We use RetValNum here, for the case 450 // that U is really a use of an insertvalue instruction that uses the 451 // original Use. 452 RetOrArg Use = CreateRet(RI->getParent()->getParent(), RetValNum); 453 // We might be live, depending on the liveness of Use. 454 return MarkIfNotLive(Use, MaybeLiveUses); 455 } 456 if (const InsertValueInst *IV = dyn_cast<InsertValueInst>(V)) { 457 if (U.getOperandNo() != InsertValueInst::getAggregateOperandIndex() 458 && IV->hasIndices()) 459 // The use we are examining is inserted into an aggregate. Our liveness 460 // depends on all uses of that aggregate, but if it is used as a return 461 // value, only index at which we were inserted counts. 462 RetValNum = *IV->idx_begin(); 463 464 // Note that if we are used as the aggregate operand to the insertvalue, 465 // we don't change RetValNum, but do survey all our uses. 466 467 Liveness Result = MaybeLive; 468 for (Value::const_use_iterator I = IV->use_begin(), 469 E = V->use_end(); I != E; ++I) { 470 Result = SurveyUse(I, MaybeLiveUses, RetValNum); 471 if (Result == Live) 472 break; 473 } 474 return Result; 475 } 476 477 if (ImmutableCallSite CS = V) { 478 const Function *F = CS.getCalledFunction(); 479 if (F) { 480 // Used in a direct call. 481 482 // Find the argument number. We know for sure that this use is an 483 // argument, since if it was the function argument this would be an 484 // indirect call and the we know can't be looking at a value of the 485 // label type (for the invoke instruction). 486 unsigned ArgNo = CS.getArgumentNo(U); 487 488 if (ArgNo >= F->getFunctionType()->getNumParams()) 489 // The value is passed in through a vararg! Must be live. 490 return Live; 491 492 assert(CS.getArgument(ArgNo) 493 == CS->getOperand(U.getOperandNo()) 494 && "Argument is not where we expected it"); 495 496 // Value passed to a normal call. It's only live when the corresponding 497 // argument to the called function turns out live. 498 RetOrArg Use = CreateArg(F, ArgNo); 499 return MarkIfNotLive(Use, MaybeLiveUses); 500 } 501 } 502 // Used in any other way? Value must be live. 503 return Live; 504} 505 506/// SurveyUses - This looks at all the uses of the given value 507/// Returns the Liveness deduced from the uses of this value. 508/// 509/// Adds all uses that cause the result to be MaybeLive to MaybeLiveRetUses. If 510/// the result is Live, MaybeLiveUses might be modified but its content should 511/// be ignored (since it might not be complete). 512DAE::Liveness DAE::SurveyUses(const Value *V, UseVector &MaybeLiveUses) { 513 // Assume it's dead (which will only hold if there are no uses at all..). 514 Liveness Result = MaybeLive; 515 // Check each use. 516 for (Value::const_use_iterator I = V->use_begin(), 517 E = V->use_end(); I != E; ++I) { 518 Result = SurveyUse(I, MaybeLiveUses); 519 if (Result == Live) 520 break; 521 } 522 return Result; 523} 524 525// SurveyFunction - This performs the initial survey of the specified function, 526// checking out whether or not it uses any of its incoming arguments or whether 527// any callers use the return value. This fills in the LiveValues set and Uses 528// map. 529// 530// We consider arguments of non-internal functions to be intrinsically alive as 531// well as arguments to functions which have their "address taken". 532// 533void DAE::SurveyFunction(const Function &F) { 534 unsigned RetCount = NumRetVals(&F); 535 // Assume all return values are dead 536 typedef SmallVector<Liveness, 5> RetVals; 537 RetVals RetValLiveness(RetCount, MaybeLive); 538 539 typedef SmallVector<UseVector, 5> RetUses; 540 // These vectors map each return value to the uses that make it MaybeLive, so 541 // we can add those to the Uses map if the return value really turns out to be 542 // MaybeLive. Initialized to a list of RetCount empty lists. 543 RetUses MaybeLiveRetUses(RetCount); 544 545 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 546 if (const ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) 547 if (RI->getNumOperands() != 0 && RI->getOperand(0)->getType() 548 != F.getFunctionType()->getReturnType()) { 549 // We don't support old style multiple return values. 550 MarkLive(F); 551 return; 552 } 553 554 if (!F.hasLocalLinkage() && (!ShouldHackArguments() || F.isIntrinsic())) { 555 MarkLive(F); 556 return; 557 } 558 559 DEBUG(dbgs() << "DAE - Inspecting callers for fn: " << F.getName() << "\n"); 560 // Keep track of the number of live retvals, so we can skip checks once all 561 // of them turn out to be live. 562 unsigned NumLiveRetVals = 0; 563 Type *STy = dyn_cast<StructType>(F.getReturnType()); 564 // Loop all uses of the function. 565 for (Value::const_use_iterator I = F.use_begin(), E = F.use_end(); 566 I != E; ++I) { 567 // If the function is PASSED IN as an argument, its address has been 568 // taken. 569 ImmutableCallSite CS(*I); 570 if (!CS || !CS.isCallee(I)) { 571 MarkLive(F); 572 return; 573 } 574 575 // If this use is anything other than a call site, the function is alive. 576 const Instruction *TheCall = CS.getInstruction(); 577 if (!TheCall) { // Not a direct call site? 578 MarkLive(F); 579 return; 580 } 581 582 // If we end up here, we are looking at a direct call to our function. 583 584 // Now, check how our return value(s) is/are used in this caller. Don't 585 // bother checking return values if all of them are live already. 586 if (NumLiveRetVals != RetCount) { 587 if (STy) { 588 // Check all uses of the return value. 589 for (Value::const_use_iterator I = TheCall->use_begin(), 590 E = TheCall->use_end(); I != E; ++I) { 591 const ExtractValueInst *Ext = dyn_cast<ExtractValueInst>(*I); 592 if (Ext && Ext->hasIndices()) { 593 // This use uses a part of our return value, survey the uses of 594 // that part and store the results for this index only. 595 unsigned Idx = *Ext->idx_begin(); 596 if (RetValLiveness[Idx] != Live) { 597 RetValLiveness[Idx] = SurveyUses(Ext, MaybeLiveRetUses[Idx]); 598 if (RetValLiveness[Idx] == Live) 599 NumLiveRetVals++; 600 } 601 } else { 602 // Used by something else than extractvalue. Mark all return 603 // values as live. 604 for (unsigned i = 0; i != RetCount; ++i ) 605 RetValLiveness[i] = Live; 606 NumLiveRetVals = RetCount; 607 break; 608 } 609 } 610 } else { 611 // Single return value 612 RetValLiveness[0] = SurveyUses(TheCall, MaybeLiveRetUses[0]); 613 if (RetValLiveness[0] == Live) 614 NumLiveRetVals = RetCount; 615 } 616 } 617 } 618 619 // Now we've inspected all callers, record the liveness of our return values. 620 for (unsigned i = 0; i != RetCount; ++i) 621 MarkValue(CreateRet(&F, i), RetValLiveness[i], MaybeLiveRetUses[i]); 622 623 DEBUG(dbgs() << "DAE - Inspecting args for fn: " << F.getName() << "\n"); 624 625 // Now, check all of our arguments. 626 unsigned i = 0; 627 UseVector MaybeLiveArgUses; 628 for (Function::const_arg_iterator AI = F.arg_begin(), 629 E = F.arg_end(); AI != E; ++AI, ++i) { 630 Liveness Result; 631 if (F.getFunctionType()->isVarArg()) { 632 // Variadic functions will already have a va_arg function expanded inside 633 // them, making them potentially very sensitive to ABI changes resulting 634 // from removing arguments entirely, so don't. For example AArch64 handles 635 // register and stack HFAs very differently, and this is reflected in the 636 // IR which has already been generated. 637 Result = Live; 638 } else { 639 // See what the effect of this use is (recording any uses that cause 640 // MaybeLive in MaybeLiveArgUses). 641 Result = SurveyUses(AI, MaybeLiveArgUses); 642 } 643 644 // Mark the result. 645 MarkValue(CreateArg(&F, i), Result, MaybeLiveArgUses); 646 // Clear the vector again for the next iteration. 647 MaybeLiveArgUses.clear(); 648 } 649} 650 651/// MarkValue - This function marks the liveness of RA depending on L. If L is 652/// MaybeLive, it also takes all uses in MaybeLiveUses and records them in Uses, 653/// such that RA will be marked live if any use in MaybeLiveUses gets marked 654/// live later on. 655void DAE::MarkValue(const RetOrArg &RA, Liveness L, 656 const UseVector &MaybeLiveUses) { 657 switch (L) { 658 case Live: MarkLive(RA); break; 659 case MaybeLive: 660 { 661 // Note any uses of this value, so this return value can be 662 // marked live whenever one of the uses becomes live. 663 for (UseVector::const_iterator UI = MaybeLiveUses.begin(), 664 UE = MaybeLiveUses.end(); UI != UE; ++UI) 665 Uses.insert(std::make_pair(*UI, RA)); 666 break; 667 } 668 } 669} 670 671/// MarkLive - Mark the given Function as alive, meaning that it cannot be 672/// changed in any way. Additionally, 673/// mark any values that are used as this function's parameters or by its return 674/// values (according to Uses) live as well. 675void DAE::MarkLive(const Function &F) { 676 DEBUG(dbgs() << "DAE - Intrinsically live fn: " << F.getName() << "\n"); 677 // Mark the function as live. 678 LiveFunctions.insert(&F); 679 // Mark all arguments as live. 680 for (unsigned i = 0, e = F.arg_size(); i != e; ++i) 681 PropagateLiveness(CreateArg(&F, i)); 682 // Mark all return values as live. 683 for (unsigned i = 0, e = NumRetVals(&F); i != e; ++i) 684 PropagateLiveness(CreateRet(&F, i)); 685} 686 687/// MarkLive - Mark the given return value or argument as live. Additionally, 688/// mark any values that are used by this value (according to Uses) live as 689/// well. 690void DAE::MarkLive(const RetOrArg &RA) { 691 if (LiveFunctions.count(RA.F)) 692 return; // Function was already marked Live. 693 694 if (!LiveValues.insert(RA).second) 695 return; // We were already marked Live. 696 697 DEBUG(dbgs() << "DAE - Marking " << RA.getDescription() << " live\n"); 698 PropagateLiveness(RA); 699} 700 701/// PropagateLiveness - Given that RA is a live value, propagate it's liveness 702/// to any other values it uses (according to Uses). 703void DAE::PropagateLiveness(const RetOrArg &RA) { 704 // We don't use upper_bound (or equal_range) here, because our recursive call 705 // to ourselves is likely to cause the upper_bound (which is the first value 706 // not belonging to RA) to become erased and the iterator invalidated. 707 UseMap::iterator Begin = Uses.lower_bound(RA); 708 UseMap::iterator E = Uses.end(); 709 UseMap::iterator I; 710 for (I = Begin; I != E && I->first == RA; ++I) 711 MarkLive(I->second); 712 713 // Erase RA from the Uses map (from the lower bound to wherever we ended up 714 // after the loop). 715 Uses.erase(Begin, I); 716} 717 718// RemoveDeadStuffFromFunction - Remove any arguments and return values from F 719// that are not in LiveValues. Transform the function and all of the callees of 720// the function to not have these arguments and return values. 721// 722bool DAE::RemoveDeadStuffFromFunction(Function *F) { 723 // Don't modify fully live functions 724 if (LiveFunctions.count(F)) 725 return false; 726 727 // Start by computing a new prototype for the function, which is the same as 728 // the old function, but has fewer arguments and a different return type. 729 FunctionType *FTy = F->getFunctionType(); 730 std::vector<Type*> Params; 731 732 // Keep track of if we have a live 'returned' argument 733 bool HasLiveReturnedArg = false; 734 735 // Set up to build a new list of parameter attributes. 736 SmallVector<AttributeSet, 8> AttributesVec; 737 const AttributeSet &PAL = F->getAttributes(); 738 739 // Remember which arguments are still alive. 740 SmallVector<bool, 10> ArgAlive(FTy->getNumParams(), false); 741 // Construct the new parameter list from non-dead arguments. Also construct 742 // a new set of parameter attributes to correspond. Skip the first parameter 743 // attribute, since that belongs to the return value. 744 unsigned i = 0; 745 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 746 I != E; ++I, ++i) { 747 RetOrArg Arg = CreateArg(F, i); 748 if (LiveValues.erase(Arg)) { 749 Params.push_back(I->getType()); 750 ArgAlive[i] = true; 751 752 // Get the original parameter attributes (skipping the first one, that is 753 // for the return value. 754 if (PAL.hasAttributes(i + 1)) { 755 AttrBuilder B(PAL, i + 1); 756 if (B.contains(Attribute::Returned)) 757 HasLiveReturnedArg = true; 758 AttributesVec. 759 push_back(AttributeSet::get(F->getContext(), Params.size(), B)); 760 } 761 } else { 762 ++NumArgumentsEliminated; 763 DEBUG(dbgs() << "DAE - Removing argument " << i << " (" << I->getName() 764 << ") from " << F->getName() << "\n"); 765 } 766 } 767 768 // Find out the new return value. 769 Type *RetTy = FTy->getReturnType(); 770 Type *NRetTy = NULL; 771 unsigned RetCount = NumRetVals(F); 772 773 // -1 means unused, other numbers are the new index 774 SmallVector<int, 5> NewRetIdxs(RetCount, -1); 775 std::vector<Type*> RetTypes; 776 777 // If there is a function with a live 'returned' argument but a dead return 778 // value, then there are two possible actions: 779 // 1) Eliminate the return value and take off the 'returned' attribute on the 780 // argument. 781 // 2) Retain the 'returned' attribute and treat the return value (but not the 782 // entire function) as live so that it is not eliminated. 783 // 784 // It's not clear in the general case which option is more profitable because, 785 // even in the absence of explicit uses of the return value, code generation 786 // is free to use the 'returned' attribute to do things like eliding 787 // save/restores of registers across calls. Whether or not this happens is 788 // target and ABI-specific as well as depending on the amount of register 789 // pressure, so there's no good way for an IR-level pass to figure this out. 790 // 791 // Fortunately, the only places where 'returned' is currently generated by 792 // the FE are places where 'returned' is basically free and almost always a 793 // performance win, so the second option can just be used always for now. 794 // 795 // This should be revisited if 'returned' is ever applied more liberally. 796 if (RetTy->isVoidTy() || HasLiveReturnedArg) { 797 NRetTy = RetTy; 798 } else { 799 StructType *STy = dyn_cast<StructType>(RetTy); 800 if (STy) 801 // Look at each of the original return values individually. 802 for (unsigned i = 0; i != RetCount; ++i) { 803 RetOrArg Ret = CreateRet(F, i); 804 if (LiveValues.erase(Ret)) { 805 RetTypes.push_back(STy->getElementType(i)); 806 NewRetIdxs[i] = RetTypes.size() - 1; 807 } else { 808 ++NumRetValsEliminated; 809 DEBUG(dbgs() << "DAE - Removing return value " << i << " from " 810 << F->getName() << "\n"); 811 } 812 } 813 else 814 // We used to return a single value. 815 if (LiveValues.erase(CreateRet(F, 0))) { 816 RetTypes.push_back(RetTy); 817 NewRetIdxs[0] = 0; 818 } else { 819 DEBUG(dbgs() << "DAE - Removing return value from " << F->getName() 820 << "\n"); 821 ++NumRetValsEliminated; 822 } 823 if (RetTypes.size() > 1) 824 // More than one return type? Return a struct with them. Also, if we used 825 // to return a struct and didn't change the number of return values, 826 // return a struct again. This prevents changing {something} into 827 // something and {} into void. 828 // Make the new struct packed if we used to return a packed struct 829 // already. 830 NRetTy = StructType::get(STy->getContext(), RetTypes, STy->isPacked()); 831 else if (RetTypes.size() == 1) 832 // One return type? Just a simple value then, but only if we didn't use to 833 // return a struct with that simple value before. 834 NRetTy = RetTypes.front(); 835 else if (RetTypes.size() == 0) 836 // No return types? Make it void, but only if we didn't use to return {}. 837 NRetTy = Type::getVoidTy(F->getContext()); 838 } 839 840 assert(NRetTy && "No new return type found?"); 841 842 // The existing function return attributes. 843 AttributeSet RAttrs = PAL.getRetAttributes(); 844 845 // Remove any incompatible attributes, but only if we removed all return 846 // values. Otherwise, ensure that we don't have any conflicting attributes 847 // here. Currently, this should not be possible, but special handling might be 848 // required when new return value attributes are added. 849 if (NRetTy->isVoidTy()) 850 RAttrs = 851 AttributeSet::get(NRetTy->getContext(), AttributeSet::ReturnIndex, 852 AttrBuilder(RAttrs, AttributeSet::ReturnIndex). 853 removeAttributes(AttributeFuncs:: 854 typeIncompatible(NRetTy, AttributeSet::ReturnIndex), 855 AttributeSet::ReturnIndex)); 856 else 857 assert(!AttrBuilder(RAttrs, AttributeSet::ReturnIndex). 858 hasAttributes(AttributeFuncs:: 859 typeIncompatible(NRetTy, AttributeSet::ReturnIndex), 860 AttributeSet::ReturnIndex) && 861 "Return attributes no longer compatible?"); 862 863 if (RAttrs.hasAttributes(AttributeSet::ReturnIndex)) 864 AttributesVec.push_back(AttributeSet::get(NRetTy->getContext(), RAttrs)); 865 866 if (PAL.hasAttributes(AttributeSet::FunctionIndex)) 867 AttributesVec.push_back(AttributeSet::get(F->getContext(), 868 PAL.getFnAttributes())); 869 870 // Reconstruct the AttributesList based on the vector we constructed. 871 AttributeSet NewPAL = AttributeSet::get(F->getContext(), AttributesVec); 872 873 // Create the new function type based on the recomputed parameters. 874 FunctionType *NFTy = FunctionType::get(NRetTy, Params, FTy->isVarArg()); 875 876 // No change? 877 if (NFTy == FTy) 878 return false; 879 880 // Create the new function body and insert it into the module... 881 Function *NF = Function::Create(NFTy, F->getLinkage()); 882 NF->copyAttributesFrom(F); 883 NF->setAttributes(NewPAL); 884 // Insert the new function before the old function, so we won't be processing 885 // it again. 886 F->getParent()->getFunctionList().insert(F, NF); 887 NF->takeName(F); 888 889 // Loop over all of the callers of the function, transforming the call sites 890 // to pass in a smaller number of arguments into the new function. 891 // 892 std::vector<Value*> Args; 893 while (!F->use_empty()) { 894 CallSite CS(F->use_back()); 895 Instruction *Call = CS.getInstruction(); 896 897 AttributesVec.clear(); 898 const AttributeSet &CallPAL = CS.getAttributes(); 899 900 // The call return attributes. 901 AttributeSet RAttrs = CallPAL.getRetAttributes(); 902 903 // Adjust in case the function was changed to return void. 904 RAttrs = 905 AttributeSet::get(NF->getContext(), AttributeSet::ReturnIndex, 906 AttrBuilder(RAttrs, AttributeSet::ReturnIndex). 907 removeAttributes(AttributeFuncs:: 908 typeIncompatible(NF->getReturnType(), 909 AttributeSet::ReturnIndex), 910 AttributeSet::ReturnIndex)); 911 if (RAttrs.hasAttributes(AttributeSet::ReturnIndex)) 912 AttributesVec.push_back(AttributeSet::get(NF->getContext(), RAttrs)); 913 914 // Declare these outside of the loops, so we can reuse them for the second 915 // loop, which loops the varargs. 916 CallSite::arg_iterator I = CS.arg_begin(); 917 unsigned i = 0; 918 // Loop over those operands, corresponding to the normal arguments to the 919 // original function, and add those that are still alive. 920 for (unsigned e = FTy->getNumParams(); i != e; ++I, ++i) 921 if (ArgAlive[i]) { 922 Args.push_back(*I); 923 // Get original parameter attributes, but skip return attributes. 924 if (CallPAL.hasAttributes(i + 1)) { 925 AttrBuilder B(CallPAL, i + 1); 926 // If the return type has changed, then get rid of 'returned' on the 927 // call site. The alternative is to make all 'returned' attributes on 928 // call sites keep the return value alive just like 'returned' 929 // attributes on function declaration but it's less clearly a win 930 // and this is not an expected case anyway 931 if (NRetTy != RetTy && B.contains(Attribute::Returned)) 932 B.removeAttribute(Attribute::Returned); 933 AttributesVec. 934 push_back(AttributeSet::get(F->getContext(), Args.size(), B)); 935 } 936 } 937 938 // Push any varargs arguments on the list. Don't forget their attributes. 939 for (CallSite::arg_iterator E = CS.arg_end(); I != E; ++I, ++i) { 940 Args.push_back(*I); 941 if (CallPAL.hasAttributes(i + 1)) { 942 AttrBuilder B(CallPAL, i + 1); 943 AttributesVec. 944 push_back(AttributeSet::get(F->getContext(), Args.size(), B)); 945 } 946 } 947 948 if (CallPAL.hasAttributes(AttributeSet::FunctionIndex)) 949 AttributesVec.push_back(AttributeSet::get(Call->getContext(), 950 CallPAL.getFnAttributes())); 951 952 // Reconstruct the AttributesList based on the vector we constructed. 953 AttributeSet NewCallPAL = AttributeSet::get(F->getContext(), AttributesVec); 954 955 Instruction *New; 956 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 957 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 958 Args, "", Call); 959 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 960 cast<InvokeInst>(New)->setAttributes(NewCallPAL); 961 } else { 962 New = CallInst::Create(NF, Args, "", Call); 963 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 964 cast<CallInst>(New)->setAttributes(NewCallPAL); 965 if (cast<CallInst>(Call)->isTailCall()) 966 cast<CallInst>(New)->setTailCall(); 967 } 968 New->setDebugLoc(Call->getDebugLoc()); 969 970 Args.clear(); 971 972 if (!Call->use_empty()) { 973 if (New->getType() == Call->getType()) { 974 // Return type not changed? Just replace users then. 975 Call->replaceAllUsesWith(New); 976 New->takeName(Call); 977 } else if (New->getType()->isVoidTy()) { 978 // Our return value has uses, but they will get removed later on. 979 // Replace by null for now. 980 if (!Call->getType()->isX86_MMXTy()) 981 Call->replaceAllUsesWith(Constant::getNullValue(Call->getType())); 982 } else { 983 assert(RetTy->isStructTy() && 984 "Return type changed, but not into a void. The old return type" 985 " must have been a struct!"); 986 Instruction *InsertPt = Call; 987 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 988 BasicBlock::iterator IP = II->getNormalDest()->begin(); 989 while (isa<PHINode>(IP)) ++IP; 990 InsertPt = IP; 991 } 992 993 // We used to return a struct. Instead of doing smart stuff with all the 994 // uses of this struct, we will just rebuild it using 995 // extract/insertvalue chaining and let instcombine clean that up. 996 // 997 // Start out building up our return value from undef 998 Value *RetVal = UndefValue::get(RetTy); 999 for (unsigned i = 0; i != RetCount; ++i) 1000 if (NewRetIdxs[i] != -1) { 1001 Value *V; 1002 if (RetTypes.size() > 1) 1003 // We are still returning a struct, so extract the value from our 1004 // return value 1005 V = ExtractValueInst::Create(New, NewRetIdxs[i], "newret", 1006 InsertPt); 1007 else 1008 // We are now returning a single element, so just insert that 1009 V = New; 1010 // Insert the value at the old position 1011 RetVal = InsertValueInst::Create(RetVal, V, i, "oldret", InsertPt); 1012 } 1013 // Now, replace all uses of the old call instruction with the return 1014 // struct we built 1015 Call->replaceAllUsesWith(RetVal); 1016 New->takeName(Call); 1017 } 1018 } 1019 1020 // Finally, remove the old call from the program, reducing the use-count of 1021 // F. 1022 Call->eraseFromParent(); 1023 } 1024 1025 // Since we have now created the new function, splice the body of the old 1026 // function right into the new function, leaving the old rotting hulk of the 1027 // function empty. 1028 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 1029 1030 // Loop over the argument list, transferring uses of the old arguments over to 1031 // the new arguments, also transferring over the names as well. 1032 i = 0; 1033 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), 1034 I2 = NF->arg_begin(); I != E; ++I, ++i) 1035 if (ArgAlive[i]) { 1036 // If this is a live argument, move the name and users over to the new 1037 // version. 1038 I->replaceAllUsesWith(I2); 1039 I2->takeName(I); 1040 ++I2; 1041 } else { 1042 // If this argument is dead, replace any uses of it with null constants 1043 // (these are guaranteed to become unused later on). 1044 if (!I->getType()->isX86_MMXTy()) 1045 I->replaceAllUsesWith(Constant::getNullValue(I->getType())); 1046 } 1047 1048 // If we change the return value of the function we must rewrite any return 1049 // instructions. Check this now. 1050 if (F->getReturnType() != NF->getReturnType()) 1051 for (Function::iterator BB = NF->begin(), E = NF->end(); BB != E; ++BB) 1052 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 1053 Value *RetVal; 1054 1055 if (NFTy->getReturnType()->isVoidTy()) { 1056 RetVal = 0; 1057 } else { 1058 assert (RetTy->isStructTy()); 1059 // The original return value was a struct, insert 1060 // extractvalue/insertvalue chains to extract only the values we need 1061 // to return and insert them into our new result. 1062 // This does generate messy code, but we'll let it to instcombine to 1063 // clean that up. 1064 Value *OldRet = RI->getOperand(0); 1065 // Start out building up our return value from undef 1066 RetVal = UndefValue::get(NRetTy); 1067 for (unsigned i = 0; i != RetCount; ++i) 1068 if (NewRetIdxs[i] != -1) { 1069 ExtractValueInst *EV = ExtractValueInst::Create(OldRet, i, 1070 "oldret", RI); 1071 if (RetTypes.size() > 1) { 1072 // We're still returning a struct, so reinsert the value into 1073 // our new return value at the new index 1074 1075 RetVal = InsertValueInst::Create(RetVal, EV, NewRetIdxs[i], 1076 "newret", RI); 1077 } else { 1078 // We are now only returning a simple value, so just return the 1079 // extracted value. 1080 RetVal = EV; 1081 } 1082 } 1083 } 1084 // Replace the return instruction with one returning the new return 1085 // value (possibly 0 if we became void). 1086 ReturnInst::Create(F->getContext(), RetVal, RI); 1087 BB->getInstList().erase(RI); 1088 } 1089 1090 // Patch the pointer to LLVM function in debug info descriptor. 1091 FunctionDIMap::iterator DI = FunctionDIs.find(F); 1092 if (DI != FunctionDIs.end()) 1093 DI->second.replaceFunction(NF); 1094 1095 // Now that the old function is dead, delete it. 1096 F->eraseFromParent(); 1097 1098 return true; 1099} 1100 1101bool DAE::runOnModule(Module &M) { 1102 bool Changed = false; 1103 1104 // Collect debug info descriptors for functions. 1105 CollectFunctionDIs(M); 1106 1107 // First pass: Do a simple check to see if any functions can have their "..." 1108 // removed. We can do this if they never call va_start. This loop cannot be 1109 // fused with the next loop, because deleting a function invalidates 1110 // information computed while surveying other functions. 1111 DEBUG(dbgs() << "DAE - Deleting dead varargs\n"); 1112 for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { 1113 Function &F = *I++; 1114 if (F.getFunctionType()->isVarArg()) 1115 Changed |= DeleteDeadVarargs(F); 1116 } 1117 1118 // Second phase:loop through the module, determining which arguments are live. 1119 // We assume all arguments are dead unless proven otherwise (allowing us to 1120 // determine that dead arguments passed into recursive functions are dead). 1121 // 1122 DEBUG(dbgs() << "DAE - Determining liveness\n"); 1123 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) 1124 SurveyFunction(*I); 1125 1126 // Now, remove all dead arguments and return values from each function in 1127 // turn. 1128 for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { 1129 // Increment now, because the function will probably get removed (ie. 1130 // replaced by a new one). 1131 Function *F = I++; 1132 Changed |= RemoveDeadStuffFromFunction(F); 1133 } 1134 1135 // Finally, look for any unused parameters in functions with non-local 1136 // linkage and replace the passed in parameters with undef. 1137 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) { 1138 Function& F = *I; 1139 1140 Changed |= RemoveDeadArgumentsFromCallers(F); 1141 } 1142 1143 return Changed; 1144} 1145