DeadArgumentElimination.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
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/IR/CallSite.h" 27#include "llvm/IR/CallingConv.h" 28#include "llvm/IR/Constant.h" 29#include "llvm/IR/DIBuilder.h" 30#include "llvm/IR/DebugInfo.h" 31#include "llvm/IR/DerivedTypes.h" 32#include "llvm/IR/Instructions.h" 33#include "llvm/IR/IntrinsicInst.h" 34#include "llvm/IR/LLVMContext.h" 35#include "llvm/IR/Module.h" 36#include "llvm/Pass.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 return std::tie(F, Idx, IsArg) < std::tie(O.F, O.Idx, O.IsArg); 66 } 67 68 /// Make RetOrArg comparable, so we can easily iterate the multimap. 69 bool operator==(const RetOrArg &O) const { 70 return F == O.F && Idx == O.Idx && IsArg == O.IsArg; 71 } 72 73 std::string getDescription() const { 74 return std::string((IsArg ? "Argument #" : "Return value #")) 75 + utostr(Idx) + " of function " + F->getName().str(); 76 } 77 }; 78 79 /// Liveness enum - During our initial pass over the program, we determine 80 /// that things are either alive or maybe alive. We don't mark anything 81 /// explicitly dead (even if we know they are), since anything not alive 82 /// with no registered uses (in Uses) will never be marked alive and will 83 /// thus become dead in the end. 84 enum Liveness { Live, MaybeLive }; 85 86 /// Convenience wrapper 87 RetOrArg CreateRet(const Function *F, unsigned Idx) { 88 return RetOrArg(F, Idx, false); 89 } 90 /// Convenience wrapper 91 RetOrArg CreateArg(const Function *F, unsigned Idx) { 92 return RetOrArg(F, Idx, true); 93 } 94 95 typedef std::multimap<RetOrArg, RetOrArg> UseMap; 96 /// This maps a return value or argument to any MaybeLive return values or 97 /// arguments it uses. This allows the MaybeLive values to be marked live 98 /// when any of its users is marked live. 99 /// For example (indices are left out for clarity): 100 /// - Uses[ret F] = ret G 101 /// This means that F calls G, and F returns the value returned by G. 102 /// - Uses[arg F] = ret G 103 /// This means that some function calls G and passes its result as an 104 /// argument to F. 105 /// - Uses[ret F] = arg F 106 /// This means that F returns one of its own arguments. 107 /// - Uses[arg F] = arg G 108 /// This means that G calls F and passes one of its own (G's) arguments 109 /// directly to F. 110 UseMap Uses; 111 112 typedef std::set<RetOrArg> LiveSet; 113 typedef std::set<const Function*> LiveFuncSet; 114 115 /// This set contains all values that have been determined to be live. 116 LiveSet LiveValues; 117 /// This set contains all values that are cannot be changed in any way. 118 LiveFuncSet LiveFunctions; 119 120 typedef SmallVector<RetOrArg, 5> UseVector; 121 122 // Map each LLVM function to corresponding metadata with debug info. If 123 // the function is replaced with another one, we should patch the pointer 124 // to LLVM function in metadata. 125 // As the code generation for module is finished (and DIBuilder is 126 // finalized) we assume that subprogram descriptors won't be changed, and 127 // they are stored in map for short duration anyway. 128 typedef DenseMap<Function*, DISubprogram> FunctionDIMap; 129 FunctionDIMap FunctionDIs; 130 131 protected: 132 // DAH uses this to specify a different ID. 133 explicit DAE(char &ID) : ModulePass(ID) {} 134 135 public: 136 static char ID; // Pass identification, replacement for typeid 137 DAE() : ModulePass(ID) { 138 initializeDAEPass(*PassRegistry::getPassRegistry()); 139 } 140 141 bool runOnModule(Module &M) override; 142 143 virtual bool ShouldHackArguments() const { return false; } 144 145 private: 146 Liveness MarkIfNotLive(RetOrArg Use, UseVector &MaybeLiveUses); 147 Liveness SurveyUse(const Use *U, UseVector &MaybeLiveUses, 148 unsigned RetValNum = 0); 149 Liveness SurveyUses(const Value *V, UseVector &MaybeLiveUses); 150 151 void CollectFunctionDIs(Module &M); 152 void SurveyFunction(const Function &F); 153 void MarkValue(const RetOrArg &RA, Liveness L, 154 const UseVector &MaybeLiveUses); 155 void MarkLive(const RetOrArg &RA); 156 void MarkLive(const Function &F); 157 void PropagateLiveness(const RetOrArg &RA); 158 bool RemoveDeadStuffFromFunction(Function *F); 159 bool DeleteDeadVarargs(Function &Fn); 160 bool RemoveDeadArgumentsFromCallers(Function &Fn); 161 }; 162} 163 164 165char DAE::ID = 0; 166INITIALIZE_PASS(DAE, "deadargelim", "Dead Argument Elimination", false, false) 167 168namespace { 169 /// DAH - DeadArgumentHacking pass - Same as dead argument elimination, but 170 /// deletes arguments to functions which are external. This is only for use 171 /// by bugpoint. 172 struct DAH : public DAE { 173 static char ID; 174 DAH() : DAE(ID) {} 175 176 bool ShouldHackArguments() const override { return true; } 177 }; 178} 179 180char DAH::ID = 0; 181INITIALIZE_PASS(DAH, "deadarghaX0r", 182 "Dead Argument Hacking (BUGPOINT USE ONLY; DO NOT USE)", 183 false, false) 184 185/// createDeadArgEliminationPass - This pass removes arguments from functions 186/// which are not used by the body of the function. 187/// 188ModulePass *llvm::createDeadArgEliminationPass() { return new DAE(); } 189ModulePass *llvm::createDeadArgHackingPass() { return new DAH(); } 190 191/// CollectFunctionDIs - Map each function in the module to its debug info 192/// descriptor. 193void DAE::CollectFunctionDIs(Module &M) { 194 FunctionDIs.clear(); 195 196 for (Module::named_metadata_iterator I = M.named_metadata_begin(), 197 E = M.named_metadata_end(); I != E; ++I) { 198 NamedMDNode &NMD = *I; 199 for (unsigned MDIndex = 0, MDNum = NMD.getNumOperands(); 200 MDIndex < MDNum; ++MDIndex) { 201 MDNode *Node = NMD.getOperand(MDIndex); 202 if (!DIDescriptor(Node).isCompileUnit()) 203 continue; 204 DICompileUnit CU(Node); 205 const DIArray &SPs = CU.getSubprograms(); 206 for (unsigned SPIndex = 0, SPNum = SPs.getNumElements(); 207 SPIndex < SPNum; ++SPIndex) { 208 DISubprogram SP(SPs.getElement(SPIndex)); 209 assert((!SP || SP.isSubprogram()) && 210 "A MDNode in subprograms of a CU should be null or a DISubprogram."); 211 if (!SP) 212 continue; 213 if (Function *F = SP.getFunction()) 214 FunctionDIs[F] = SP; 215 } 216 } 217 } 218} 219 220/// DeleteDeadVarargs - If this is an function that takes a ... list, and if 221/// llvm.vastart is never called, the varargs list is dead for the function. 222bool DAE::DeleteDeadVarargs(Function &Fn) { 223 assert(Fn.getFunctionType()->isVarArg() && "Function isn't varargs!"); 224 if (Fn.isDeclaration() || !Fn.hasLocalLinkage()) return false; 225 226 // Ensure that the function is only directly called. 227 if (Fn.hasAddressTaken()) 228 return false; 229 230 // Okay, we know we can transform this function if safe. Scan its body 231 // looking for calls to llvm.vastart. 232 for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) { 233 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 234 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 235 if (II->getIntrinsicID() == Intrinsic::vastart) 236 return false; 237 } 238 } 239 } 240 241 // If we get here, there are no calls to llvm.vastart in the function body, 242 // remove the "..." and adjust all the calls. 243 244 // Start by computing a new prototype for the function, which is the same as 245 // the old function, but doesn't have isVarArg set. 246 FunctionType *FTy = Fn.getFunctionType(); 247 248 std::vector<Type*> Params(FTy->param_begin(), FTy->param_end()); 249 FunctionType *NFTy = FunctionType::get(FTy->getReturnType(), 250 Params, false); 251 unsigned NumArgs = Params.size(); 252 253 // Create the new function body and insert it into the module... 254 Function *NF = Function::Create(NFTy, Fn.getLinkage()); 255 NF->copyAttributesFrom(&Fn); 256 Fn.getParent()->getFunctionList().insert(&Fn, NF); 257 NF->takeName(&Fn); 258 259 // Loop over all of the callers of the function, transforming the call sites 260 // to pass in a smaller number of arguments into the new function. 261 // 262 std::vector<Value*> Args; 263 for (Value::user_iterator I = Fn.user_begin(), E = Fn.user_end(); I != E; ) { 264 CallSite CS(*I++); 265 if (!CS) 266 continue; 267 Instruction *Call = CS.getInstruction(); 268 269 // Pass all the same arguments. 270 Args.assign(CS.arg_begin(), CS.arg_begin() + NumArgs); 271 272 // Drop any attributes that were on the vararg arguments. 273 AttributeSet PAL = CS.getAttributes(); 274 if (!PAL.isEmpty() && PAL.getSlotIndex(PAL.getNumSlots() - 1) > NumArgs) { 275 SmallVector<AttributeSet, 8> AttributesVec; 276 for (unsigned i = 0; PAL.getSlotIndex(i) <= NumArgs; ++i) 277 AttributesVec.push_back(PAL.getSlotAttributes(i)); 278 if (PAL.hasAttributes(AttributeSet::FunctionIndex)) 279 AttributesVec.push_back(AttributeSet::get(Fn.getContext(), 280 PAL.getFnAttributes())); 281 PAL = AttributeSet::get(Fn.getContext(), AttributesVec); 282 } 283 284 Instruction *New; 285 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 286 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 287 Args, "", Call); 288 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 289 cast<InvokeInst>(New)->setAttributes(PAL); 290 } else { 291 New = CallInst::Create(NF, Args, "", Call); 292 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 293 cast<CallInst>(New)->setAttributes(PAL); 294 if (cast<CallInst>(Call)->isTailCall()) 295 cast<CallInst>(New)->setTailCall(); 296 } 297 New->setDebugLoc(Call->getDebugLoc()); 298 299 Args.clear(); 300 301 if (!Call->use_empty()) 302 Call->replaceAllUsesWith(New); 303 304 New->takeName(Call); 305 306 // Finally, remove the old call from the program, reducing the use-count of 307 // F. 308 Call->eraseFromParent(); 309 } 310 311 // Since we have now created the new function, splice the body of the old 312 // function right into the new function, leaving the old rotting hulk of the 313 // function empty. 314 NF->getBasicBlockList().splice(NF->begin(), Fn.getBasicBlockList()); 315 316 // Loop over the argument list, transferring uses of the old arguments over to 317 // the new arguments, also transferring over the names as well. While we're at 318 // it, remove the dead arguments from the DeadArguments list. 319 // 320 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(), 321 I2 = NF->arg_begin(); I != E; ++I, ++I2) { 322 // Move the name and users over to the new version. 323 I->replaceAllUsesWith(I2); 324 I2->takeName(I); 325 } 326 327 // Patch the pointer to LLVM function in debug info descriptor. 328 FunctionDIMap::iterator DI = FunctionDIs.find(&Fn); 329 if (DI != FunctionDIs.end()) 330 DI->second.replaceFunction(NF); 331 332 // Fix up any BlockAddresses that refer to the function. 333 Fn.replaceAllUsesWith(ConstantExpr::getBitCast(NF, Fn.getType())); 334 // Delete the bitcast that we just created, so that NF does not 335 // appear to be address-taken. 336 NF->removeDeadConstantUsers(); 337 // Finally, nuke the old function. 338 Fn.eraseFromParent(); 339 return true; 340} 341 342/// RemoveDeadArgumentsFromCallers - Checks if the given function has any 343/// arguments that are unused, and changes the caller parameters to be undefined 344/// instead. 345bool DAE::RemoveDeadArgumentsFromCallers(Function &Fn) 346{ 347 if (Fn.isDeclaration() || Fn.mayBeOverridden()) 348 return false; 349 350 // Functions with local linkage should already have been handled, except the 351 // fragile (variadic) ones which we can improve here. 352 if (Fn.hasLocalLinkage() && !Fn.getFunctionType()->isVarArg()) 353 return false; 354 355 // If a function seen at compile time is not necessarily the one linked to 356 // the binary being built, it is illegal to change the actual arguments 357 // passed to it. These functions can be captured by isWeakForLinker(). 358 // *NOTE* that mayBeOverridden() is insufficient for this purpose as it 359 // doesn't include linkage types like AvailableExternallyLinkage and 360 // LinkOnceODRLinkage. Take link_odr* as an example, it indicates a set of 361 // *EQUIVALENT* globals that can be merged at link-time. However, the 362 // semantic of *EQUIVALENT*-functions includes parameters. Changing 363 // parameters breaks this assumption. 364 // 365 if (Fn.isWeakForLinker()) 366 return false; 367 368 if (Fn.use_empty()) 369 return false; 370 371 SmallVector<unsigned, 8> UnusedArgs; 372 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(); 373 I != E; ++I) { 374 Argument *Arg = I; 375 376 if (Arg->use_empty() && !Arg->hasByValOrInAllocaAttr()) 377 UnusedArgs.push_back(Arg->getArgNo()); 378 } 379 380 if (UnusedArgs.empty()) 381 return false; 382 383 bool Changed = false; 384 385 for (Use &U : Fn.uses()) { 386 CallSite CS(U.getUser()); 387 if (!CS || !CS.isCallee(&U)) 388 continue; 389 390 // Now go through all unused args and replace them with "undef". 391 for (unsigned I = 0, E = UnusedArgs.size(); I != E; ++I) { 392 unsigned ArgNo = UnusedArgs[I]; 393 394 Value *Arg = CS.getArgument(ArgNo); 395 CS.setArgument(ArgNo, UndefValue::get(Arg->getType())); 396 ++NumArgumentsReplacedWithUndef; 397 Changed = true; 398 } 399 } 400 401 return Changed; 402} 403 404/// Convenience function that returns the number of return values. It returns 0 405/// for void functions and 1 for functions not returning a struct. It returns 406/// the number of struct elements for functions returning a struct. 407static unsigned NumRetVals(const Function *F) { 408 if (F->getReturnType()->isVoidTy()) 409 return 0; 410 else if (StructType *STy = dyn_cast<StructType>(F->getReturnType())) 411 return STy->getNumElements(); 412 else 413 return 1; 414} 415 416/// MarkIfNotLive - This checks Use for liveness in LiveValues. If Use is not 417/// live, it adds Use to the MaybeLiveUses argument. Returns the determined 418/// liveness of Use. 419DAE::Liveness DAE::MarkIfNotLive(RetOrArg Use, UseVector &MaybeLiveUses) { 420 // We're live if our use or its Function is already marked as live. 421 if (LiveFunctions.count(Use.F) || LiveValues.count(Use)) 422 return Live; 423 424 // We're maybe live otherwise, but remember that we must become live if 425 // Use becomes live. 426 MaybeLiveUses.push_back(Use); 427 return MaybeLive; 428} 429 430 431/// SurveyUse - This looks at a single use of an argument or return value 432/// and determines if it should be alive or not. Adds this use to MaybeLiveUses 433/// if it causes the used value to become MaybeLive. 434/// 435/// RetValNum is the return value number to use when this use is used in a 436/// return instruction. This is used in the recursion, you should always leave 437/// it at 0. 438DAE::Liveness DAE::SurveyUse(const Use *U, 439 UseVector &MaybeLiveUses, unsigned RetValNum) { 440 const User *V = U->getUser(); 441 if (const ReturnInst *RI = dyn_cast<ReturnInst>(V)) { 442 // The value is returned from a function. It's only live when the 443 // function's return value is live. We use RetValNum here, for the case 444 // that U is really a use of an insertvalue instruction that uses the 445 // original Use. 446 RetOrArg Use = CreateRet(RI->getParent()->getParent(), RetValNum); 447 // We might be live, depending on the liveness of Use. 448 return MarkIfNotLive(Use, MaybeLiveUses); 449 } 450 if (const InsertValueInst *IV = dyn_cast<InsertValueInst>(V)) { 451 if (U->getOperandNo() != InsertValueInst::getAggregateOperandIndex() 452 && IV->hasIndices()) 453 // The use we are examining is inserted into an aggregate. Our liveness 454 // depends on all uses of that aggregate, but if it is used as a return 455 // value, only index at which we were inserted counts. 456 RetValNum = *IV->idx_begin(); 457 458 // Note that if we are used as the aggregate operand to the insertvalue, 459 // we don't change RetValNum, but do survey all our uses. 460 461 Liveness Result = MaybeLive; 462 for (const Use &UU : IV->uses()) { 463 Result = SurveyUse(&UU, MaybeLiveUses, RetValNum); 464 if (Result == Live) 465 break; 466 } 467 return Result; 468 } 469 470 if (ImmutableCallSite CS = V) { 471 const Function *F = CS.getCalledFunction(); 472 if (F) { 473 // Used in a direct call. 474 475 // Find the argument number. We know for sure that this use is an 476 // argument, since if it was the function argument this would be an 477 // indirect call and the we know can't be looking at a value of the 478 // label type (for the invoke instruction). 479 unsigned ArgNo = CS.getArgumentNo(U); 480 481 if (ArgNo >= F->getFunctionType()->getNumParams()) 482 // The value is passed in through a vararg! Must be live. 483 return Live; 484 485 assert(CS.getArgument(ArgNo) 486 == CS->getOperand(U->getOperandNo()) 487 && "Argument is not where we expected it"); 488 489 // Value passed to a normal call. It's only live when the corresponding 490 // argument to the called function turns out live. 491 RetOrArg Use = CreateArg(F, ArgNo); 492 return MarkIfNotLive(Use, MaybeLiveUses); 493 } 494 } 495 // Used in any other way? Value must be live. 496 return Live; 497} 498 499/// SurveyUses - This looks at all the uses of the given value 500/// Returns the Liveness deduced from the uses of this value. 501/// 502/// Adds all uses that cause the result to be MaybeLive to MaybeLiveRetUses. If 503/// the result is Live, MaybeLiveUses might be modified but its content should 504/// be ignored (since it might not be complete). 505DAE::Liveness DAE::SurveyUses(const Value *V, UseVector &MaybeLiveUses) { 506 // Assume it's dead (which will only hold if there are no uses at all..). 507 Liveness Result = MaybeLive; 508 // Check each use. 509 for (const Use &U : V->uses()) { 510 Result = SurveyUse(&U, MaybeLiveUses); 511 if (Result == Live) 512 break; 513 } 514 return Result; 515} 516 517// SurveyFunction - This performs the initial survey of the specified function, 518// checking out whether or not it uses any of its incoming arguments or whether 519// any callers use the return value. This fills in the LiveValues set and Uses 520// map. 521// 522// We consider arguments of non-internal functions to be intrinsically alive as 523// well as arguments to functions which have their "address taken". 524// 525void DAE::SurveyFunction(const Function &F) { 526 // Functions with inalloca parameters are expecting args in a particular 527 // register and memory layout. 528 if (F.getAttributes().hasAttrSomewhere(Attribute::InAlloca)) { 529 MarkLive(F); 530 return; 531 } 532 533 unsigned RetCount = NumRetVals(&F); 534 // Assume all return values are dead 535 typedef SmallVector<Liveness, 5> RetVals; 536 RetVals RetValLiveness(RetCount, MaybeLive); 537 538 typedef SmallVector<UseVector, 5> RetUses; 539 // These vectors map each return value to the uses that make it MaybeLive, so 540 // we can add those to the Uses map if the return value really turns out to be 541 // MaybeLive. Initialized to a list of RetCount empty lists. 542 RetUses MaybeLiveRetUses(RetCount); 543 544 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 545 if (const ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) 546 if (RI->getNumOperands() != 0 && RI->getOperand(0)->getType() 547 != F.getFunctionType()->getReturnType()) { 548 // We don't support old style multiple return values. 549 MarkLive(F); 550 return; 551 } 552 553 if (!F.hasLocalLinkage() && (!ShouldHackArguments() || F.isIntrinsic())) { 554 MarkLive(F); 555 return; 556 } 557 558 DEBUG(dbgs() << "DAE - Inspecting callers for fn: " << F.getName() << "\n"); 559 // Keep track of the number of live retvals, so we can skip checks once all 560 // of them turn out to be live. 561 unsigned NumLiveRetVals = 0; 562 Type *STy = dyn_cast<StructType>(F.getReturnType()); 563 // Loop all uses of the function. 564 for (const Use &U : F.uses()) { 565 // If the function is PASSED IN as an argument, its address has been 566 // taken. 567 ImmutableCallSite CS(U.getUser()); 568 if (!CS || !CS.isCallee(&U)) { 569 MarkLive(F); 570 return; 571 } 572 573 // If this use is anything other than a call site, the function is alive. 574 const Instruction *TheCall = CS.getInstruction(); 575 if (!TheCall) { // Not a direct call site? 576 MarkLive(F); 577 return; 578 } 579 580 // If we end up here, we are looking at a direct call to our function. 581 582 // Now, check how our return value(s) is/are used in this caller. Don't 583 // bother checking return values if all of them are live already. 584 if (NumLiveRetVals != RetCount) { 585 if (STy) { 586 // Check all uses of the return value. 587 for (const User *U : TheCall->users()) { 588 const ExtractValueInst *Ext = dyn_cast<ExtractValueInst>(U); 589 if (Ext && Ext->hasIndices()) { 590 // This use uses a part of our return value, survey the uses of 591 // that part and store the results for this index only. 592 unsigned Idx = *Ext->idx_begin(); 593 if (RetValLiveness[Idx] != Live) { 594 RetValLiveness[Idx] = SurveyUses(Ext, MaybeLiveRetUses[Idx]); 595 if (RetValLiveness[Idx] == Live) 596 NumLiveRetVals++; 597 } 598 } else { 599 // Used by something else than extractvalue. Mark all return 600 // values as live. 601 for (unsigned i = 0; i != RetCount; ++i ) 602 RetValLiveness[i] = Live; 603 NumLiveRetVals = RetCount; 604 break; 605 } 606 } 607 } else { 608 // Single return value 609 RetValLiveness[0] = SurveyUses(TheCall, MaybeLiveRetUses[0]); 610 if (RetValLiveness[0] == Live) 611 NumLiveRetVals = RetCount; 612 } 613 } 614 } 615 616 // Now we've inspected all callers, record the liveness of our return values. 617 for (unsigned i = 0; i != RetCount; ++i) 618 MarkValue(CreateRet(&F, i), RetValLiveness[i], MaybeLiveRetUses[i]); 619 620 DEBUG(dbgs() << "DAE - Inspecting args for fn: " << F.getName() << "\n"); 621 622 // Now, check all of our arguments. 623 unsigned i = 0; 624 UseVector MaybeLiveArgUses; 625 for (Function::const_arg_iterator AI = F.arg_begin(), 626 E = F.arg_end(); AI != E; ++AI, ++i) { 627 Liveness Result; 628 if (F.getFunctionType()->isVarArg()) { 629 // Variadic functions will already have a va_arg function expanded inside 630 // them, making them potentially very sensitive to ABI changes resulting 631 // from removing arguments entirely, so don't. For example AArch64 handles 632 // register and stack HFAs very differently, and this is reflected in the 633 // IR which has already been generated. 634 Result = Live; 635 } else { 636 // See what the effect of this use is (recording any uses that cause 637 // MaybeLive in MaybeLiveArgUses). 638 Result = SurveyUses(AI, MaybeLiveArgUses); 639 } 640 641 // Mark the result. 642 MarkValue(CreateArg(&F, i), Result, MaybeLiveArgUses); 643 // Clear the vector again for the next iteration. 644 MaybeLiveArgUses.clear(); 645 } 646} 647 648/// MarkValue - This function marks the liveness of RA depending on L. If L is 649/// MaybeLive, it also takes all uses in MaybeLiveUses and records them in Uses, 650/// such that RA will be marked live if any use in MaybeLiveUses gets marked 651/// live later on. 652void DAE::MarkValue(const RetOrArg &RA, Liveness L, 653 const UseVector &MaybeLiveUses) { 654 switch (L) { 655 case Live: MarkLive(RA); break; 656 case MaybeLive: 657 { 658 // Note any uses of this value, so this return value can be 659 // marked live whenever one of the uses becomes live. 660 for (UseVector::const_iterator UI = MaybeLiveUses.begin(), 661 UE = MaybeLiveUses.end(); UI != UE; ++UI) 662 Uses.insert(std::make_pair(*UI, RA)); 663 break; 664 } 665 } 666} 667 668/// MarkLive - Mark the given Function as alive, meaning that it cannot be 669/// changed in any way. Additionally, 670/// mark any values that are used as this function's parameters or by its return 671/// values (according to Uses) live as well. 672void DAE::MarkLive(const Function &F) { 673 DEBUG(dbgs() << "DAE - Intrinsically live fn: " << F.getName() << "\n"); 674 // Mark the function as live. 675 LiveFunctions.insert(&F); 676 // Mark all arguments as live. 677 for (unsigned i = 0, e = F.arg_size(); i != e; ++i) 678 PropagateLiveness(CreateArg(&F, i)); 679 // Mark all return values as live. 680 for (unsigned i = 0, e = NumRetVals(&F); i != e; ++i) 681 PropagateLiveness(CreateRet(&F, i)); 682} 683 684/// MarkLive - Mark the given return value or argument as live. Additionally, 685/// mark any values that are used by this value (according to Uses) live as 686/// well. 687void DAE::MarkLive(const RetOrArg &RA) { 688 if (LiveFunctions.count(RA.F)) 689 return; // Function was already marked Live. 690 691 if (!LiveValues.insert(RA).second) 692 return; // We were already marked Live. 693 694 DEBUG(dbgs() << "DAE - Marking " << RA.getDescription() << " live\n"); 695 PropagateLiveness(RA); 696} 697 698/// PropagateLiveness - Given that RA is a live value, propagate it's liveness 699/// to any other values it uses (according to Uses). 700void DAE::PropagateLiveness(const RetOrArg &RA) { 701 // We don't use upper_bound (or equal_range) here, because our recursive call 702 // to ourselves is likely to cause the upper_bound (which is the first value 703 // not belonging to RA) to become erased and the iterator invalidated. 704 UseMap::iterator Begin = Uses.lower_bound(RA); 705 UseMap::iterator E = Uses.end(); 706 UseMap::iterator I; 707 for (I = Begin; I != E && I->first == RA; ++I) 708 MarkLive(I->second); 709 710 // Erase RA from the Uses map (from the lower bound to wherever we ended up 711 // after the loop). 712 Uses.erase(Begin, I); 713} 714 715// RemoveDeadStuffFromFunction - Remove any arguments and return values from F 716// that are not in LiveValues. Transform the function and all of the callees of 717// the function to not have these arguments and return values. 718// 719bool DAE::RemoveDeadStuffFromFunction(Function *F) { 720 // Don't modify fully live functions 721 if (LiveFunctions.count(F)) 722 return false; 723 724 // Start by computing a new prototype for the function, which is the same as 725 // the old function, but has fewer arguments and a different return type. 726 FunctionType *FTy = F->getFunctionType(); 727 std::vector<Type*> Params; 728 729 // Keep track of if we have a live 'returned' argument 730 bool HasLiveReturnedArg = false; 731 732 // Set up to build a new list of parameter attributes. 733 SmallVector<AttributeSet, 8> AttributesVec; 734 const AttributeSet &PAL = F->getAttributes(); 735 736 // Remember which arguments are still alive. 737 SmallVector<bool, 10> ArgAlive(FTy->getNumParams(), false); 738 // Construct the new parameter list from non-dead arguments. Also construct 739 // a new set of parameter attributes to correspond. Skip the first parameter 740 // attribute, since that belongs to the return value. 741 unsigned i = 0; 742 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 743 I != E; ++I, ++i) { 744 RetOrArg Arg = CreateArg(F, i); 745 if (LiveValues.erase(Arg)) { 746 Params.push_back(I->getType()); 747 ArgAlive[i] = true; 748 749 // Get the original parameter attributes (skipping the first one, that is 750 // for the return value. 751 if (PAL.hasAttributes(i + 1)) { 752 AttrBuilder B(PAL, i + 1); 753 if (B.contains(Attribute::Returned)) 754 HasLiveReturnedArg = true; 755 AttributesVec. 756 push_back(AttributeSet::get(F->getContext(), Params.size(), B)); 757 } 758 } else { 759 ++NumArgumentsEliminated; 760 DEBUG(dbgs() << "DAE - Removing argument " << i << " (" << I->getName() 761 << ") from " << F->getName() << "\n"); 762 } 763 } 764 765 // Find out the new return value. 766 Type *RetTy = FTy->getReturnType(); 767 Type *NRetTy = NULL; 768 unsigned RetCount = NumRetVals(F); 769 770 // -1 means unused, other numbers are the new index 771 SmallVector<int, 5> NewRetIdxs(RetCount, -1); 772 std::vector<Type*> RetTypes; 773 774 // If there is a function with a live 'returned' argument but a dead return 775 // value, then there are two possible actions: 776 // 1) Eliminate the return value and take off the 'returned' attribute on the 777 // argument. 778 // 2) Retain the 'returned' attribute and treat the return value (but not the 779 // entire function) as live so that it is not eliminated. 780 // 781 // It's not clear in the general case which option is more profitable because, 782 // even in the absence of explicit uses of the return value, code generation 783 // is free to use the 'returned' attribute to do things like eliding 784 // save/restores of registers across calls. Whether or not this happens is 785 // target and ABI-specific as well as depending on the amount of register 786 // pressure, so there's no good way for an IR-level pass to figure this out. 787 // 788 // Fortunately, the only places where 'returned' is currently generated by 789 // the FE are places where 'returned' is basically free and almost always a 790 // performance win, so the second option can just be used always for now. 791 // 792 // This should be revisited if 'returned' is ever applied more liberally. 793 if (RetTy->isVoidTy() || HasLiveReturnedArg) { 794 NRetTy = RetTy; 795 } else { 796 StructType *STy = dyn_cast<StructType>(RetTy); 797 if (STy) 798 // Look at each of the original return values individually. 799 for (unsigned i = 0; i != RetCount; ++i) { 800 RetOrArg Ret = CreateRet(F, i); 801 if (LiveValues.erase(Ret)) { 802 RetTypes.push_back(STy->getElementType(i)); 803 NewRetIdxs[i] = RetTypes.size() - 1; 804 } else { 805 ++NumRetValsEliminated; 806 DEBUG(dbgs() << "DAE - Removing return value " << i << " from " 807 << F->getName() << "\n"); 808 } 809 } 810 else 811 // We used to return a single value. 812 if (LiveValues.erase(CreateRet(F, 0))) { 813 RetTypes.push_back(RetTy); 814 NewRetIdxs[0] = 0; 815 } else { 816 DEBUG(dbgs() << "DAE - Removing return value from " << F->getName() 817 << "\n"); 818 ++NumRetValsEliminated; 819 } 820 if (RetTypes.size() > 1) 821 // More than one return type? Return a struct with them. Also, if we used 822 // to return a struct and didn't change the number of return values, 823 // return a struct again. This prevents changing {something} into 824 // something and {} into void. 825 // Make the new struct packed if we used to return a packed struct 826 // already. 827 NRetTy = StructType::get(STy->getContext(), RetTypes, STy->isPacked()); 828 else if (RetTypes.size() == 1) 829 // One return type? Just a simple value then, but only if we didn't use to 830 // return a struct with that simple value before. 831 NRetTy = RetTypes.front(); 832 else if (RetTypes.size() == 0) 833 // No return types? Make it void, but only if we didn't use to return {}. 834 NRetTy = Type::getVoidTy(F->getContext()); 835 } 836 837 assert(NRetTy && "No new return type found?"); 838 839 // The existing function return attributes. 840 AttributeSet RAttrs = PAL.getRetAttributes(); 841 842 // Remove any incompatible attributes, but only if we removed all return 843 // values. Otherwise, ensure that we don't have any conflicting attributes 844 // here. Currently, this should not be possible, but special handling might be 845 // required when new return value attributes are added. 846 if (NRetTy->isVoidTy()) 847 RAttrs = 848 AttributeSet::get(NRetTy->getContext(), AttributeSet::ReturnIndex, 849 AttrBuilder(RAttrs, AttributeSet::ReturnIndex). 850 removeAttributes(AttributeFuncs:: 851 typeIncompatible(NRetTy, AttributeSet::ReturnIndex), 852 AttributeSet::ReturnIndex)); 853 else 854 assert(!AttrBuilder(RAttrs, AttributeSet::ReturnIndex). 855 hasAttributes(AttributeFuncs:: 856 typeIncompatible(NRetTy, AttributeSet::ReturnIndex), 857 AttributeSet::ReturnIndex) && 858 "Return attributes no longer compatible?"); 859 860 if (RAttrs.hasAttributes(AttributeSet::ReturnIndex)) 861 AttributesVec.push_back(AttributeSet::get(NRetTy->getContext(), RAttrs)); 862 863 if (PAL.hasAttributes(AttributeSet::FunctionIndex)) 864 AttributesVec.push_back(AttributeSet::get(F->getContext(), 865 PAL.getFnAttributes())); 866 867 // Reconstruct the AttributesList based on the vector we constructed. 868 AttributeSet NewPAL = AttributeSet::get(F->getContext(), AttributesVec); 869 870 // Create the new function type based on the recomputed parameters. 871 FunctionType *NFTy = FunctionType::get(NRetTy, Params, FTy->isVarArg()); 872 873 // No change? 874 if (NFTy == FTy) 875 return false; 876 877 // Create the new function body and insert it into the module... 878 Function *NF = Function::Create(NFTy, F->getLinkage()); 879 NF->copyAttributesFrom(F); 880 NF->setAttributes(NewPAL); 881 // Insert the new function before the old function, so we won't be processing 882 // it again. 883 F->getParent()->getFunctionList().insert(F, NF); 884 NF->takeName(F); 885 886 // Loop over all of the callers of the function, transforming the call sites 887 // to pass in a smaller number of arguments into the new function. 888 // 889 std::vector<Value*> Args; 890 while (!F->use_empty()) { 891 CallSite CS(F->user_back()); 892 Instruction *Call = CS.getInstruction(); 893 894 AttributesVec.clear(); 895 const AttributeSet &CallPAL = CS.getAttributes(); 896 897 // The call return attributes. 898 AttributeSet RAttrs = CallPAL.getRetAttributes(); 899 900 // Adjust in case the function was changed to return void. 901 RAttrs = 902 AttributeSet::get(NF->getContext(), AttributeSet::ReturnIndex, 903 AttrBuilder(RAttrs, AttributeSet::ReturnIndex). 904 removeAttributes(AttributeFuncs:: 905 typeIncompatible(NF->getReturnType(), 906 AttributeSet::ReturnIndex), 907 AttributeSet::ReturnIndex)); 908 if (RAttrs.hasAttributes(AttributeSet::ReturnIndex)) 909 AttributesVec.push_back(AttributeSet::get(NF->getContext(), RAttrs)); 910 911 // Declare these outside of the loops, so we can reuse them for the second 912 // loop, which loops the varargs. 913 CallSite::arg_iterator I = CS.arg_begin(); 914 unsigned i = 0; 915 // Loop over those operands, corresponding to the normal arguments to the 916 // original function, and add those that are still alive. 917 for (unsigned e = FTy->getNumParams(); i != e; ++I, ++i) 918 if (ArgAlive[i]) { 919 Args.push_back(*I); 920 // Get original parameter attributes, but skip return attributes. 921 if (CallPAL.hasAttributes(i + 1)) { 922 AttrBuilder B(CallPAL, i + 1); 923 // If the return type has changed, then get rid of 'returned' on the 924 // call site. The alternative is to make all 'returned' attributes on 925 // call sites keep the return value alive just like 'returned' 926 // attributes on function declaration but it's less clearly a win 927 // and this is not an expected case anyway 928 if (NRetTy != RetTy && B.contains(Attribute::Returned)) 929 B.removeAttribute(Attribute::Returned); 930 AttributesVec. 931 push_back(AttributeSet::get(F->getContext(), Args.size(), B)); 932 } 933 } 934 935 // Push any varargs arguments on the list. Don't forget their attributes. 936 for (CallSite::arg_iterator E = CS.arg_end(); I != E; ++I, ++i) { 937 Args.push_back(*I); 938 if (CallPAL.hasAttributes(i + 1)) { 939 AttrBuilder B(CallPAL, i + 1); 940 AttributesVec. 941 push_back(AttributeSet::get(F->getContext(), Args.size(), B)); 942 } 943 } 944 945 if (CallPAL.hasAttributes(AttributeSet::FunctionIndex)) 946 AttributesVec.push_back(AttributeSet::get(Call->getContext(), 947 CallPAL.getFnAttributes())); 948 949 // Reconstruct the AttributesList based on the vector we constructed. 950 AttributeSet NewCallPAL = AttributeSet::get(F->getContext(), AttributesVec); 951 952 Instruction *New; 953 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 954 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 955 Args, "", Call); 956 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 957 cast<InvokeInst>(New)->setAttributes(NewCallPAL); 958 } else { 959 New = CallInst::Create(NF, Args, "", Call); 960 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 961 cast<CallInst>(New)->setAttributes(NewCallPAL); 962 if (cast<CallInst>(Call)->isTailCall()) 963 cast<CallInst>(New)->setTailCall(); 964 } 965 New->setDebugLoc(Call->getDebugLoc()); 966 967 Args.clear(); 968 969 if (!Call->use_empty()) { 970 if (New->getType() == Call->getType()) { 971 // Return type not changed? Just replace users then. 972 Call->replaceAllUsesWith(New); 973 New->takeName(Call); 974 } else if (New->getType()->isVoidTy()) { 975 // Our return value has uses, but they will get removed later on. 976 // Replace by null for now. 977 if (!Call->getType()->isX86_MMXTy()) 978 Call->replaceAllUsesWith(Constant::getNullValue(Call->getType())); 979 } else { 980 assert(RetTy->isStructTy() && 981 "Return type changed, but not into a void. The old return type" 982 " must have been a struct!"); 983 Instruction *InsertPt = Call; 984 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 985 BasicBlock::iterator IP = II->getNormalDest()->begin(); 986 while (isa<PHINode>(IP)) ++IP; 987 InsertPt = IP; 988 } 989 990 // We used to return a struct. Instead of doing smart stuff with all the 991 // uses of this struct, we will just rebuild it using 992 // extract/insertvalue chaining and let instcombine clean that up. 993 // 994 // Start out building up our return value from undef 995 Value *RetVal = UndefValue::get(RetTy); 996 for (unsigned i = 0; i != RetCount; ++i) 997 if (NewRetIdxs[i] != -1) { 998 Value *V; 999 if (RetTypes.size() > 1) 1000 // We are still returning a struct, so extract the value from our 1001 // return value 1002 V = ExtractValueInst::Create(New, NewRetIdxs[i], "newret", 1003 InsertPt); 1004 else 1005 // We are now returning a single element, so just insert that 1006 V = New; 1007 // Insert the value at the old position 1008 RetVal = InsertValueInst::Create(RetVal, V, i, "oldret", InsertPt); 1009 } 1010 // Now, replace all uses of the old call instruction with the return 1011 // struct we built 1012 Call->replaceAllUsesWith(RetVal); 1013 New->takeName(Call); 1014 } 1015 } 1016 1017 // Finally, remove the old call from the program, reducing the use-count of 1018 // F. 1019 Call->eraseFromParent(); 1020 } 1021 1022 // Since we have now created the new function, splice the body of the old 1023 // function right into the new function, leaving the old rotting hulk of the 1024 // function empty. 1025 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 1026 1027 // Loop over the argument list, transferring uses of the old arguments over to 1028 // the new arguments, also transferring over the names as well. 1029 i = 0; 1030 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), 1031 I2 = NF->arg_begin(); I != E; ++I, ++i) 1032 if (ArgAlive[i]) { 1033 // If this is a live argument, move the name and users over to the new 1034 // version. 1035 I->replaceAllUsesWith(I2); 1036 I2->takeName(I); 1037 ++I2; 1038 } else { 1039 // If this argument is dead, replace any uses of it with null constants 1040 // (these are guaranteed to become unused later on). 1041 if (!I->getType()->isX86_MMXTy()) 1042 I->replaceAllUsesWith(Constant::getNullValue(I->getType())); 1043 } 1044 1045 // If we change the return value of the function we must rewrite any return 1046 // instructions. Check this now. 1047 if (F->getReturnType() != NF->getReturnType()) 1048 for (Function::iterator BB = NF->begin(), E = NF->end(); BB != E; ++BB) 1049 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 1050 Value *RetVal; 1051 1052 if (NFTy->getReturnType()->isVoidTy()) { 1053 RetVal = 0; 1054 } else { 1055 assert (RetTy->isStructTy()); 1056 // The original return value was a struct, insert 1057 // extractvalue/insertvalue chains to extract only the values we need 1058 // to return and insert them into our new result. 1059 // This does generate messy code, but we'll let it to instcombine to 1060 // clean that up. 1061 Value *OldRet = RI->getOperand(0); 1062 // Start out building up our return value from undef 1063 RetVal = UndefValue::get(NRetTy); 1064 for (unsigned i = 0; i != RetCount; ++i) 1065 if (NewRetIdxs[i] != -1) { 1066 ExtractValueInst *EV = ExtractValueInst::Create(OldRet, i, 1067 "oldret", RI); 1068 if (RetTypes.size() > 1) { 1069 // We're still returning a struct, so reinsert the value into 1070 // our new return value at the new index 1071 1072 RetVal = InsertValueInst::Create(RetVal, EV, NewRetIdxs[i], 1073 "newret", RI); 1074 } else { 1075 // We are now only returning a simple value, so just return the 1076 // extracted value. 1077 RetVal = EV; 1078 } 1079 } 1080 } 1081 // Replace the return instruction with one returning the new return 1082 // value (possibly 0 if we became void). 1083 ReturnInst::Create(F->getContext(), RetVal, RI); 1084 BB->getInstList().erase(RI); 1085 } 1086 1087 // Patch the pointer to LLVM function in debug info descriptor. 1088 FunctionDIMap::iterator DI = FunctionDIs.find(F); 1089 if (DI != FunctionDIs.end()) 1090 DI->second.replaceFunction(NF); 1091 1092 // Now that the old function is dead, delete it. 1093 F->eraseFromParent(); 1094 1095 return true; 1096} 1097 1098bool DAE::runOnModule(Module &M) { 1099 bool Changed = false; 1100 1101 // Collect debug info descriptors for functions. 1102 CollectFunctionDIs(M); 1103 1104 // First pass: Do a simple check to see if any functions can have their "..." 1105 // removed. We can do this if they never call va_start. This loop cannot be 1106 // fused with the next loop, because deleting a function invalidates 1107 // information computed while surveying other functions. 1108 DEBUG(dbgs() << "DAE - Deleting dead varargs\n"); 1109 for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { 1110 Function &F = *I++; 1111 if (F.getFunctionType()->isVarArg()) 1112 Changed |= DeleteDeadVarargs(F); 1113 } 1114 1115 // Second phase:loop through the module, determining which arguments are live. 1116 // We assume all arguments are dead unless proven otherwise (allowing us to 1117 // determine that dead arguments passed into recursive functions are dead). 1118 // 1119 DEBUG(dbgs() << "DAE - Determining liveness\n"); 1120 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) 1121 SurveyFunction(*I); 1122 1123 // Now, remove all dead arguments and return values from each function in 1124 // turn. 1125 for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { 1126 // Increment now, because the function will probably get removed (ie. 1127 // replaced by a new one). 1128 Function *F = I++; 1129 Changed |= RemoveDeadStuffFromFunction(F); 1130 } 1131 1132 // Finally, look for any unused parameters in functions with non-local 1133 // linkage and replace the passed in parameters with undef. 1134 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) { 1135 Function& F = *I; 1136 1137 Changed |= RemoveDeadArgumentsFromCallers(F); 1138 } 1139 1140 return Changed; 1141} 1142