DeadArgumentElimination.cpp revision a5affdcf4cdae48ce1e20a42f8dec6fb6f1c3495
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/CallingConv.h" 23#include "llvm/Constant.h" 24#include "llvm/DerivedTypes.h" 25#include "llvm/Instructions.h" 26#include "llvm/IntrinsicInst.h" 27#include "llvm/Module.h" 28#include "llvm/Pass.h" 29#include "llvm/Support/CallSite.h" 30#include "llvm/Support/Debug.h" 31#include "llvm/ADT/SmallVector.h" 32#include "llvm/ADT/Statistic.h" 33#include "llvm/ADT/StringExtras.h" 34#include "llvm/Support/Compiler.h" 35#include <map> 36#include <set> 37using namespace llvm; 38 39STATISTIC(NumArgumentsEliminated, "Number of unread args removed"); 40STATISTIC(NumRetValsEliminated , "Number of unused return values removed"); 41 42namespace { 43 /// DAE - The dead argument elimination pass. 44 /// 45 class VISIBILITY_HIDDEN DAE : public ModulePass { 46 public: 47 48 /// Struct that represents (part of) either a return value or a function 49 /// argument. Used so that arguments and return values can be used 50 /// interchangably. 51 struct RetOrArg { 52 RetOrArg(const Function* F, unsigned Idx, bool IsArg) : F(F), Idx(Idx), 53 IsArg(IsArg) {} 54 const Function *F; 55 unsigned Idx; 56 bool IsArg; 57 58 /// Make RetOrArg comparable, so we can put it into a map. 59 bool operator<(const RetOrArg &O) const { 60 if (F != O.F) 61 return F < O.F; 62 else if (Idx != O.Idx) 63 return Idx < O.Idx; 64 else 65 return IsArg < 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(); 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 public: 123 static char ID; // Pass identification, replacement for typeid 124 DAE() : ModulePass(&ID) {} 125 bool runOnModule(Module &M); 126 127 virtual bool ShouldHackArguments() const { return false; } 128 129 private: 130 Liveness MarkIfNotLive(RetOrArg Use, UseVector &MaybeLiveUses); 131 Liveness SurveyUse(Value::use_iterator U, UseVector &MaybeLiveUses, 132 unsigned RetValNum = 0); 133 Liveness SurveyUses(Value *V, UseVector &MaybeLiveUses); 134 135 void SurveyFunction(Function &F); 136 void MarkValue(const RetOrArg &RA, Liveness L, 137 const UseVector &MaybeLiveUses); 138 void MarkLive(const RetOrArg &RA); 139 void MarkLive(const Function &F); 140 void PropagateLiveness(const RetOrArg &RA); 141 bool RemoveDeadStuffFromFunction(Function *F); 142 bool DeleteDeadVarargs(Function &Fn); 143 }; 144} 145 146 147char DAE::ID = 0; 148static RegisterPass<DAE> 149X("deadargelim", "Dead Argument Elimination"); 150 151namespace { 152 /// DAH - DeadArgumentHacking pass - Same as dead argument elimination, but 153 /// deletes arguments to functions which are external. This is only for use 154 /// by bugpoint. 155 struct DAH : public DAE { 156 static char ID; 157 virtual bool ShouldHackArguments() const { return true; } 158 }; 159} 160 161char DAH::ID = 0; 162static RegisterPass<DAH> 163Y("deadarghaX0r", "Dead Argument Hacking (BUGPOINT USE ONLY; DO NOT USE)"); 164 165/// createDeadArgEliminationPass - This pass removes arguments from functions 166/// which are not used by the body of the function. 167/// 168ModulePass *llvm::createDeadArgEliminationPass() { return new DAE(); } 169ModulePass *llvm::createDeadArgHackingPass() { return new DAH(); } 170 171/// DeleteDeadVarargs - If this is an function that takes a ... list, and if 172/// llvm.vastart is never called, the varargs list is dead for the function. 173bool DAE::DeleteDeadVarargs(Function &Fn) { 174 assert(Fn.getFunctionType()->isVarArg() && "Function isn't varargs!"); 175 if (Fn.isDeclaration() || !Fn.hasLocalLinkage()) return false; 176 177 // Ensure that the function is only directly called. 178 for (Value::use_iterator I = Fn.use_begin(), E = Fn.use_end(); I != E; ++I) { 179 // If this use is anything other than a call site, give up. 180 CallSite CS = CallSite::get(*I); 181 Instruction *TheCall = CS.getInstruction(); 182 if (!TheCall) return false; // Not a direct call site? 183 184 // The addr of this function is passed to the call. 185 if (!CS.isCallee(I)) return false; 186 } 187 188 // Okay, we know we can transform this function if safe. Scan its body 189 // looking for calls to llvm.vastart. 190 for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) { 191 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 192 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 193 if (II->getIntrinsicID() == Intrinsic::vastart) 194 return false; 195 } 196 } 197 } 198 199 // If we get here, there are no calls to llvm.vastart in the function body, 200 // remove the "..." and adjust all the calls. 201 202 // Start by computing a new prototype for the function, which is the same as 203 // the old function, but doesn't have isVarArg set. 204 const FunctionType *FTy = Fn.getFunctionType(); 205 std::vector<const Type*> Params(FTy->param_begin(), FTy->param_end()); 206 FunctionType *NFTy = FunctionType::get(FTy->getReturnType(), Params, false); 207 unsigned NumArgs = Params.size(); 208 209 // Create the new function body and insert it into the module... 210 Function *NF = Function::Create(NFTy, Fn.getLinkage()); 211 NF->copyAttributesFrom(&Fn); 212 Fn.getParent()->getFunctionList().insert(&Fn, NF); 213 NF->takeName(&Fn); 214 215 // Loop over all of the callers of the function, transforming the call sites 216 // to pass in a smaller number of arguments into the new function. 217 // 218 std::vector<Value*> Args; 219 while (!Fn.use_empty()) { 220 CallSite CS = CallSite::get(Fn.use_back()); 221 Instruction *Call = CS.getInstruction(); 222 223 // Pass all the same arguments. 224 Args.assign(CS.arg_begin(), CS.arg_begin()+NumArgs); 225 226 // Drop any attributes that were on the vararg arguments. 227 AttrListPtr PAL = CS.getAttributes(); 228 if (!PAL.isEmpty() && PAL.getSlot(PAL.getNumSlots() - 1).Index > NumArgs) { 229 SmallVector<AttributeWithIndex, 8> AttributesVec; 230 for (unsigned i = 0; PAL.getSlot(i).Index <= NumArgs; ++i) 231 AttributesVec.push_back(PAL.getSlot(i)); 232 if (Attributes FnAttrs = PAL.getFnAttributes()) 233 AttributesVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 234 PAL = AttrListPtr::get(AttributesVec.begin(), AttributesVec.end()); 235 } 236 237 Instruction *New; 238 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 239 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 240 Args.begin(), Args.end(), "", Call); 241 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 242 cast<InvokeInst>(New)->setAttributes(PAL); 243 } else { 244 New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call); 245 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 246 cast<CallInst>(New)->setAttributes(PAL); 247 if (cast<CallInst>(Call)->isTailCall()) 248 cast<CallInst>(New)->setTailCall(); 249 } 250 Args.clear(); 251 252 if (!Call->use_empty()) 253 Call->replaceAllUsesWith(New); 254 255 New->takeName(Call); 256 257 // Finally, remove the old call from the program, reducing the use-count of 258 // F. 259 Call->eraseFromParent(); 260 } 261 262 // Since we have now created the new function, splice the body of the old 263 // function right into the new function, leaving the old rotting hulk of the 264 // function empty. 265 NF->getBasicBlockList().splice(NF->begin(), Fn.getBasicBlockList()); 266 267 // Loop over the argument list, transfering uses of the old arguments over to 268 // the new arguments, also transfering over the names as well. While we're at 269 // it, remove the dead arguments from the DeadArguments list. 270 // 271 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(), 272 I2 = NF->arg_begin(); I != E; ++I, ++I2) { 273 // Move the name and users over to the new version. 274 I->replaceAllUsesWith(I2); 275 I2->takeName(I); 276 } 277 278 // Finally, nuke the old function. 279 Fn.eraseFromParent(); 280 return true; 281} 282 283/// Convenience function that returns the number of return values. It returns 0 284/// for void functions and 1 for functions not returning a struct. It returns 285/// the number of struct elements for functions returning a struct. 286static unsigned NumRetVals(const Function *F) { 287 if (F->getReturnType() == Type::VoidTy) 288 return 0; 289 else if (const StructType *STy = dyn_cast<StructType>(F->getReturnType())) 290 return STy->getNumElements(); 291 else 292 return 1; 293} 294 295/// MarkIfNotLive - This checks Use for liveness in LiveValues. If Use is not 296/// live, it adds Use to the MaybeLiveUses argument. Returns the determined 297/// liveness of Use. 298DAE::Liveness DAE::MarkIfNotLive(RetOrArg Use, UseVector &MaybeLiveUses) { 299 // We're live if our use or its Function is already marked as live. 300 if (LiveFunctions.count(Use.F) || LiveValues.count(Use)) 301 return Live; 302 303 // We're maybe live otherwise, but remember that we must become live if 304 // Use becomes live. 305 MaybeLiveUses.push_back(Use); 306 return MaybeLive; 307} 308 309 310/// SurveyUse - This looks at a single use of an argument or return value 311/// and determines if it should be alive or not. Adds this use to MaybeLiveUses 312/// if it causes the used value to become MaybeAlive. 313/// 314/// RetValNum is the return value number to use when this use is used in a 315/// return instruction. This is used in the recursion, you should always leave 316/// it at 0. 317DAE::Liveness DAE::SurveyUse(Value::use_iterator U, UseVector &MaybeLiveUses, 318 unsigned RetValNum) { 319 Value *V = *U; 320 if (ReturnInst *RI = dyn_cast<ReturnInst>(V)) { 321 // The value is returned from a function. It's only live when the 322 // function's return value is live. We use RetValNum here, for the case 323 // that U is really a use of an insertvalue instruction that uses the 324 // orginal Use. 325 RetOrArg Use = CreateRet(RI->getParent()->getParent(), RetValNum); 326 // We might be live, depending on the liveness of Use. 327 return MarkIfNotLive(Use, MaybeLiveUses); 328 } 329 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(V)) { 330 if (U.getOperandNo() != InsertValueInst::getAggregateOperandIndex() 331 && IV->hasIndices()) 332 // The use we are examining is inserted into an aggregate. Our liveness 333 // depends on all uses of that aggregate, but if it is used as a return 334 // value, only index at which we were inserted counts. 335 RetValNum = *IV->idx_begin(); 336 337 // Note that if we are used as the aggregate operand to the insertvalue, 338 // we don't change RetValNum, but do survey all our uses. 339 340 Liveness Result = MaybeLive; 341 for (Value::use_iterator I = IV->use_begin(), 342 E = V->use_end(); I != E; ++I) { 343 Result = SurveyUse(I, MaybeLiveUses, RetValNum); 344 if (Result == Live) 345 break; 346 } 347 return Result; 348 } 349 CallSite CS = CallSite::get(V); 350 if (CS.getInstruction()) { 351 Function *F = CS.getCalledFunction(); 352 if (F) { 353 // Used in a direct call. 354 355 // Find the argument number. We know for sure that this use is an 356 // argument, since if it was the function argument this would be an 357 // indirect call and the we know can't be looking at a value of the 358 // label type (for the invoke instruction). 359 unsigned ArgNo = CS.getArgumentNo(U.getOperandNo()); 360 361 if (ArgNo >= F->getFunctionType()->getNumParams()) 362 // The value is passed in through a vararg! Must be live. 363 return Live; 364 365 assert(CS.getArgument(ArgNo) 366 == CS.getInstruction()->getOperand(U.getOperandNo()) 367 && "Argument is not where we expected it"); 368 369 // Value passed to a normal call. It's only live when the corresponding 370 // argument to the called function turns out live. 371 RetOrArg Use = CreateArg(F, ArgNo); 372 return MarkIfNotLive(Use, MaybeLiveUses); 373 } 374 } 375 // Used in any other way? Value must be live. 376 return Live; 377} 378 379/// SurveyUses - This looks at all the uses of the given value 380/// Returns the Liveness deduced from the uses of this value. 381/// 382/// Adds all uses that cause the result to be MaybeLive to MaybeLiveRetUses. If 383/// the result is Live, MaybeLiveUses might be modified but its content should 384/// be ignored (since it might not be complete). 385DAE::Liveness DAE::SurveyUses(Value *V, UseVector &MaybeLiveUses) { 386 // Assume it's dead (which will only hold if there are no uses at all..). 387 Liveness Result = MaybeLive; 388 // Check each use. 389 for (Value::use_iterator I = V->use_begin(), 390 E = V->use_end(); I != E; ++I) { 391 Result = SurveyUse(I, MaybeLiveUses); 392 if (Result == Live) 393 break; 394 } 395 return Result; 396} 397 398// SurveyFunction - This performs the initial survey of the specified function, 399// checking out whether or not it uses any of its incoming arguments or whether 400// any callers use the return value. This fills in the LiveValues set and Uses 401// map. 402// 403// We consider arguments of non-internal functions to be intrinsically alive as 404// well as arguments to functions which have their "address taken". 405// 406void DAE::SurveyFunction(Function &F) { 407 unsigned RetCount = NumRetVals(&F); 408 // Assume all return values are dead 409 typedef SmallVector<Liveness, 5> RetVals; 410 RetVals RetValLiveness(RetCount, MaybeLive); 411 412 typedef SmallVector<UseVector, 5> RetUses; 413 // These vectors map each return value to the uses that make it MaybeLive, so 414 // we can add those to the Uses map if the return value really turns out to be 415 // MaybeLive. Initialized to a list of RetCount empty lists. 416 RetUses MaybeLiveRetUses(RetCount); 417 418 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 419 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) 420 if (RI->getNumOperands() != 0 && RI->getOperand(0)->getType() 421 != F.getFunctionType()->getReturnType()) { 422 // We don't support old style multiple return values. 423 MarkLive(F); 424 return; 425 } 426 427 if (!F.hasLocalLinkage() && (!ShouldHackArguments() || F.isIntrinsic())) { 428 MarkLive(F); 429 return; 430 } 431 432 DOUT << "DAE - Inspecting callers for fn: " << F.getName() << "\n"; 433 // Keep track of the number of live retvals, so we can skip checks once all 434 // of them turn out to be live. 435 unsigned NumLiveRetVals = 0; 436 const Type *STy = dyn_cast<StructType>(F.getReturnType()); 437 // Loop all uses of the function. 438 for (Value::use_iterator I = F.use_begin(), E = F.use_end(); I != E; ++I) { 439 // If the function is PASSED IN as an argument, its address has been 440 // taken. 441 CallSite CS = CallSite::get(*I); 442 if (!CS.getInstruction() || !CS.isCallee(I)) { 443 MarkLive(F); 444 return; 445 } 446 447 // If this use is anything other than a call site, the function is alive. 448 Instruction *TheCall = CS.getInstruction(); 449 if (!TheCall) { // Not a direct call site? 450 MarkLive(F); 451 return; 452 } 453 454 // If we end up here, we are looking at a direct call to our function. 455 456 // Now, check how our return value(s) is/are used in this caller. Don't 457 // bother checking return values if all of them are live already. 458 if (NumLiveRetVals != RetCount) { 459 if (STy) { 460 // Check all uses of the return value. 461 for (Value::use_iterator I = TheCall->use_begin(), 462 E = TheCall->use_end(); I != E; ++I) { 463 ExtractValueInst *Ext = dyn_cast<ExtractValueInst>(*I); 464 if (Ext && Ext->hasIndices()) { 465 // This use uses a part of our return value, survey the uses of 466 // that part and store the results for this index only. 467 unsigned Idx = *Ext->idx_begin(); 468 if (RetValLiveness[Idx] != Live) { 469 RetValLiveness[Idx] = SurveyUses(Ext, MaybeLiveRetUses[Idx]); 470 if (RetValLiveness[Idx] == Live) 471 NumLiveRetVals++; 472 } 473 } else { 474 // Used by something else than extractvalue. Mark all return 475 // values as live. 476 for (unsigned i = 0; i != RetCount; ++i ) 477 RetValLiveness[i] = Live; 478 NumLiveRetVals = RetCount; 479 break; 480 } 481 } 482 } else { 483 // Single return value 484 RetValLiveness[0] = SurveyUses(TheCall, MaybeLiveRetUses[0]); 485 if (RetValLiveness[0] == Live) 486 NumLiveRetVals = RetCount; 487 } 488 } 489 } 490 491 // Now we've inspected all callers, record the liveness of our return values. 492 for (unsigned i = 0; i != RetCount; ++i) 493 MarkValue(CreateRet(&F, i), RetValLiveness[i], MaybeLiveRetUses[i]); 494 495 DOUT << "DAE - Inspecting args for fn: " << F.getName() << "\n"; 496 497 // Now, check all of our arguments. 498 unsigned i = 0; 499 UseVector MaybeLiveArgUses; 500 for (Function::arg_iterator AI = F.arg_begin(), 501 E = F.arg_end(); AI != E; ++AI, ++i) { 502 // See what the effect of this use is (recording any uses that cause 503 // MaybeLive in MaybeLiveArgUses). 504 Liveness Result = SurveyUses(AI, MaybeLiveArgUses); 505 // Mark the result. 506 MarkValue(CreateArg(&F, i), Result, MaybeLiveArgUses); 507 // Clear the vector again for the next iteration. 508 MaybeLiveArgUses.clear(); 509 } 510} 511 512/// MarkValue - This function marks the liveness of RA depending on L. If L is 513/// MaybeLive, it also takes all uses in MaybeLiveUses and records them in Uses, 514/// such that RA will be marked live if any use in MaybeLiveUses gets marked 515/// live later on. 516void DAE::MarkValue(const RetOrArg &RA, Liveness L, 517 const UseVector &MaybeLiveUses) { 518 switch (L) { 519 case Live: MarkLive(RA); break; 520 case MaybeLive: 521 { 522 // Note any uses of this value, so this return value can be 523 // marked live whenever one of the uses becomes live. 524 for (UseVector::const_iterator UI = MaybeLiveUses.begin(), 525 UE = MaybeLiveUses.end(); UI != UE; ++UI) 526 Uses.insert(std::make_pair(*UI, RA)); 527 break; 528 } 529 } 530} 531 532/// MarkLive - Mark the given Function as alive, meaning that it cannot be 533/// changed in any way. Additionally, 534/// mark any values that are used as this function's parameters or by its return 535/// values (according to Uses) live as well. 536void DAE::MarkLive(const Function &F) { 537 DOUT << "DAE - Intrinsically live fn: " << F.getName() << "\n"; 538 // Mark the function as live. 539 LiveFunctions.insert(&F); 540 // Mark all arguments as live. 541 for (unsigned i = 0, e = F.arg_size(); i != e; ++i) 542 PropagateLiveness(CreateArg(&F, i)); 543 // Mark all return values as live. 544 for (unsigned i = 0, e = NumRetVals(&F); i != e; ++i) 545 PropagateLiveness(CreateRet(&F, i)); 546} 547 548/// MarkLive - Mark the given return value or argument as live. Additionally, 549/// mark any values that are used by this value (according to Uses) live as 550/// well. 551void DAE::MarkLive(const RetOrArg &RA) { 552 if (LiveFunctions.count(RA.F)) 553 return; // Function was already marked Live. 554 555 if (!LiveValues.insert(RA).second) 556 return; // We were already marked Live. 557 558 DOUT << "DAE - Marking " << RA.getDescription() << " live\n"; 559 PropagateLiveness(RA); 560} 561 562/// PropagateLiveness - Given that RA is a live value, propagate it's liveness 563/// to any other values it uses (according to Uses). 564void DAE::PropagateLiveness(const RetOrArg &RA) { 565 // We don't use upper_bound (or equal_range) here, because our recursive call 566 // to ourselves is likely to cause the upper_bound (which is the first value 567 // not belonging to RA) to become erased and the iterator invalidated. 568 UseMap::iterator Begin = Uses.lower_bound(RA); 569 UseMap::iterator E = Uses.end(); 570 UseMap::iterator I; 571 for (I = Begin; I != E && I->first == RA; ++I) 572 MarkLive(I->second); 573 574 // Erase RA from the Uses map (from the lower bound to wherever we ended up 575 // after the loop). 576 Uses.erase(Begin, I); 577} 578 579// RemoveDeadStuffFromFunction - Remove any arguments and return values from F 580// that are not in LiveValues. Transform the function and all of the callees of 581// the function to not have these arguments and return values. 582// 583bool DAE::RemoveDeadStuffFromFunction(Function *F) { 584 // Don't modify fully live functions 585 if (LiveFunctions.count(F)) 586 return false; 587 588 // Start by computing a new prototype for the function, which is the same as 589 // the old function, but has fewer arguments and a different return type. 590 const FunctionType *FTy = F->getFunctionType(); 591 std::vector<const Type*> Params; 592 593 // Set up to build a new list of parameter attributes. 594 SmallVector<AttributeWithIndex, 8> AttributesVec; 595 const AttrListPtr &PAL = F->getAttributes(); 596 597 // The existing function return attributes. 598 Attributes RAttrs = PAL.getRetAttributes(); 599 Attributes FnAttrs = PAL.getFnAttributes(); 600 601 // Find out the new return value. 602 603 const Type *RetTy = FTy->getReturnType(); 604 const Type *NRetTy = NULL; 605 unsigned RetCount = NumRetVals(F); 606 // -1 means unused, other numbers are the new index 607 SmallVector<int, 5> NewRetIdxs(RetCount, -1); 608 std::vector<const Type*> RetTypes; 609 if (RetTy == Type::VoidTy) { 610 NRetTy = Type::VoidTy; 611 } else { 612 const StructType *STy = dyn_cast<StructType>(RetTy); 613 if (STy) 614 // Look at each of the original return values individually. 615 for (unsigned i = 0; i != RetCount; ++i) { 616 RetOrArg Ret = CreateRet(F, i); 617 if (LiveValues.erase(Ret)) { 618 RetTypes.push_back(STy->getElementType(i)); 619 NewRetIdxs[i] = RetTypes.size() - 1; 620 } else { 621 ++NumRetValsEliminated; 622 DOUT << "DAE - Removing return value " << i << " from " 623 << F->getNameStart() << "\n"; 624 } 625 } 626 else 627 // We used to return a single value. 628 if (LiveValues.erase(CreateRet(F, 0))) { 629 RetTypes.push_back(RetTy); 630 NewRetIdxs[0] = 0; 631 } else { 632 DOUT << "DAE - Removing return value from " << F->getNameStart() 633 << "\n"; 634 ++NumRetValsEliminated; 635 } 636 if (RetTypes.size() > 1) 637 // More than one return type? Return a struct with them. Also, if we used 638 // to return a struct and didn't change the number of return values, 639 // return a struct again. This prevents changing {something} into 640 // something and {} into void. 641 // Make the new struct packed if we used to return a packed struct 642 // already. 643 NRetTy = StructType::get(RetTypes, STy->isPacked()); 644 else if (RetTypes.size() == 1) 645 // One return type? Just a simple value then, but only if we didn't use to 646 // return a struct with that simple value before. 647 NRetTy = RetTypes.front(); 648 else if (RetTypes.size() == 0) 649 // No return types? Make it void, but only if we didn't use to return {}. 650 NRetTy = Type::VoidTy; 651 } 652 653 assert(NRetTy && "No new return type found?"); 654 655 // Remove any incompatible attributes, but only if we removed all return 656 // values. Otherwise, ensure that we don't have any conflicting attributes 657 // here. Currently, this should not be possible, but special handling might be 658 // required when new return value attributes are added. 659 if (NRetTy == Type::VoidTy) 660 RAttrs &= ~Attribute::typeIncompatible(NRetTy); 661 else 662 assert((RAttrs & Attribute::typeIncompatible(NRetTy)) == 0 663 && "Return attributes no longer compatible?"); 664 665 if (RAttrs) 666 AttributesVec.push_back(AttributeWithIndex::get(0, RAttrs)); 667 668 // Remember which arguments are still alive. 669 SmallVector<bool, 10> ArgAlive(FTy->getNumParams(), false); 670 // Construct the new parameter list from non-dead arguments. Also construct 671 // a new set of parameter attributes to correspond. Skip the first parameter 672 // attribute, since that belongs to the return value. 673 unsigned i = 0; 674 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 675 I != E; ++I, ++i) { 676 RetOrArg Arg = CreateArg(F, i); 677 if (LiveValues.erase(Arg)) { 678 Params.push_back(I->getType()); 679 ArgAlive[i] = true; 680 681 // Get the original parameter attributes (skipping the first one, that is 682 // for the return value. 683 if (Attributes Attrs = PAL.getParamAttributes(i + 1)) 684 AttributesVec.push_back(AttributeWithIndex::get(Params.size(), Attrs)); 685 } else { 686 ++NumArgumentsEliminated; 687 DOUT << "DAE - Removing argument " << i << " (" << I->getNameStart() 688 << ") from " << F->getNameStart() << "\n"; 689 } 690 } 691 692 if (FnAttrs != Attribute::None) 693 AttributesVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 694 695 // Reconstruct the AttributesList based on the vector we constructed. 696 AttrListPtr NewPAL = AttrListPtr::get(AttributesVec.begin(), AttributesVec.end()); 697 698 // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which 699 // have zero fixed arguments. 700 // 701 // Note that we apply this hack for a vararg fuction that does not have any 702 // arguments anymore, but did have them before (so don't bother fixing 703 // functions that were already broken wrt CWriter). 704 bool ExtraArgHack = false; 705 if (Params.empty() && FTy->isVarArg() && FTy->getNumParams() != 0) { 706 ExtraArgHack = true; 707 Params.push_back(Type::Int32Ty); 708 } 709 710 // Create the new function type based on the recomputed parameters. 711 FunctionType *NFTy = FunctionType::get(NRetTy, Params, FTy->isVarArg()); 712 713 // No change? 714 if (NFTy == FTy) 715 return false; 716 717 // Create the new function body and insert it into the module... 718 Function *NF = Function::Create(NFTy, F->getLinkage()); 719 NF->copyAttributesFrom(F); 720 NF->setAttributes(NewPAL); 721 // Insert the new function before the old function, so we won't be processing 722 // it again. 723 F->getParent()->getFunctionList().insert(F, NF); 724 NF->takeName(F); 725 726 // Loop over all of the callers of the function, transforming the call sites 727 // to pass in a smaller number of arguments into the new function. 728 // 729 std::vector<Value*> Args; 730 while (!F->use_empty()) { 731 CallSite CS = CallSite::get(F->use_back()); 732 Instruction *Call = CS.getInstruction(); 733 734 AttributesVec.clear(); 735 const AttrListPtr &CallPAL = CS.getAttributes(); 736 737 // The call return attributes. 738 Attributes RAttrs = CallPAL.getRetAttributes(); 739 Attributes FnAttrs = CallPAL.getFnAttributes(); 740 // Adjust in case the function was changed to return void. 741 RAttrs &= ~Attribute::typeIncompatible(NF->getReturnType()); 742 if (RAttrs) 743 AttributesVec.push_back(AttributeWithIndex::get(0, RAttrs)); 744 745 // Declare these outside of the loops, so we can reuse them for the second 746 // loop, which loops the varargs. 747 CallSite::arg_iterator I = CS.arg_begin(); 748 unsigned i = 0; 749 // Loop over those operands, corresponding to the normal arguments to the 750 // original function, and add those that are still alive. 751 for (unsigned e = FTy->getNumParams(); i != e; ++I, ++i) 752 if (ArgAlive[i]) { 753 Args.push_back(*I); 754 // Get original parameter attributes, but skip return attributes. 755 if (Attributes Attrs = CallPAL.getParamAttributes(i + 1)) 756 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 757 } 758 759 if (ExtraArgHack) 760 Args.push_back(UndefValue::get(Type::Int32Ty)); 761 762 // Push any varargs arguments on the list. Don't forget their attributes. 763 for (CallSite::arg_iterator E = CS.arg_end(); I != E; ++I, ++i) { 764 Args.push_back(*I); 765 if (Attributes Attrs = CallPAL.getParamAttributes(i + 1)) 766 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 767 } 768 769 if (FnAttrs != Attribute::None) 770 AttributesVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 771 772 // Reconstruct the AttributesList based on the vector we constructed. 773 AttrListPtr NewCallPAL = AttrListPtr::get(AttributesVec.begin(), 774 AttributesVec.end()); 775 776 Instruction *New; 777 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 778 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 779 Args.begin(), Args.end(), "", Call); 780 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 781 cast<InvokeInst>(New)->setAttributes(NewCallPAL); 782 } else { 783 New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call); 784 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 785 cast<CallInst>(New)->setAttributes(NewCallPAL); 786 if (cast<CallInst>(Call)->isTailCall()) 787 cast<CallInst>(New)->setTailCall(); 788 } 789 Args.clear(); 790 791 if (!Call->use_empty()) { 792 if (New->getType() == Call->getType()) { 793 // Return type not changed? Just replace users then. 794 Call->replaceAllUsesWith(New); 795 New->takeName(Call); 796 } else if (New->getType() == Type::VoidTy) { 797 // Our return value has uses, but they will get removed later on. 798 // Replace by null for now. 799 Call->replaceAllUsesWith(Constant::getNullValue(Call->getType())); 800 } else { 801 assert(isa<StructType>(RetTy) && 802 "Return type changed, but not into a void. The old return type" 803 " must have been a struct!"); 804 Instruction *InsertPt = Call; 805 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 806 BasicBlock::iterator IP = II->getNormalDest()->begin(); 807 while (isa<PHINode>(IP)) ++IP; 808 InsertPt = IP; 809 } 810 811 // We used to return a struct. Instead of doing smart stuff with all the 812 // uses of this struct, we will just rebuild it using 813 // extract/insertvalue chaining and let instcombine clean that up. 814 // 815 // Start out building up our return value from undef 816 Value *RetVal = llvm::UndefValue::get(RetTy); 817 for (unsigned i = 0; i != RetCount; ++i) 818 if (NewRetIdxs[i] != -1) { 819 Value *V; 820 if (RetTypes.size() > 1) 821 // We are still returning a struct, so extract the value from our 822 // return value 823 V = ExtractValueInst::Create(New, NewRetIdxs[i], "newret", 824 InsertPt); 825 else 826 // We are now returning a single element, so just insert that 827 V = New; 828 // Insert the value at the old position 829 RetVal = InsertValueInst::Create(RetVal, V, i, "oldret", InsertPt); 830 } 831 // Now, replace all uses of the old call instruction with the return 832 // struct we built 833 Call->replaceAllUsesWith(RetVal); 834 New->takeName(Call); 835 } 836 } 837 838 // Finally, remove the old call from the program, reducing the use-count of 839 // F. 840 Call->eraseFromParent(); 841 } 842 843 // Since we have now created the new function, splice the body of the old 844 // function right into the new function, leaving the old rotting hulk of the 845 // function empty. 846 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 847 848 // Loop over the argument list, transfering uses of the old arguments over to 849 // the new arguments, also transfering over the names as well. 850 i = 0; 851 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), 852 I2 = NF->arg_begin(); I != E; ++I, ++i) 853 if (ArgAlive[i]) { 854 // If this is a live argument, move the name and users over to the new 855 // version. 856 I->replaceAllUsesWith(I2); 857 I2->takeName(I); 858 ++I2; 859 } else { 860 // If this argument is dead, replace any uses of it with null constants 861 // (these are guaranteed to become unused later on). 862 I->replaceAllUsesWith(Constant::getNullValue(I->getType())); 863 } 864 865 // If we change the return value of the function we must rewrite any return 866 // instructions. Check this now. 867 if (F->getReturnType() != NF->getReturnType()) 868 for (Function::iterator BB = NF->begin(), E = NF->end(); BB != E; ++BB) 869 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 870 Value *RetVal; 871 872 if (NFTy->getReturnType() == Type::VoidTy) { 873 RetVal = 0; 874 } else { 875 assert (isa<StructType>(RetTy)); 876 // The original return value was a struct, insert 877 // extractvalue/insertvalue chains to extract only the values we need 878 // to return and insert them into our new result. 879 // This does generate messy code, but we'll let it to instcombine to 880 // clean that up. 881 Value *OldRet = RI->getOperand(0); 882 // Start out building up our return value from undef 883 RetVal = llvm::UndefValue::get(NRetTy); 884 for (unsigned i = 0; i != RetCount; ++i) 885 if (NewRetIdxs[i] != -1) { 886 ExtractValueInst *EV = ExtractValueInst::Create(OldRet, i, 887 "oldret", RI); 888 if (RetTypes.size() > 1) { 889 // We're still returning a struct, so reinsert the value into 890 // our new return value at the new index 891 892 RetVal = InsertValueInst::Create(RetVal, EV, NewRetIdxs[i], 893 "newret", RI); 894 } else { 895 // We are now only returning a simple value, so just return the 896 // extracted value. 897 RetVal = EV; 898 } 899 } 900 } 901 // Replace the return instruction with one returning the new return 902 // value (possibly 0 if we became void). 903 ReturnInst::Create(RetVal, RI); 904 BB->getInstList().erase(RI); 905 } 906 907 // Now that the old function is dead, delete it. 908 F->eraseFromParent(); 909 910 return true; 911} 912 913bool DAE::runOnModule(Module &M) { 914 bool Changed = false; 915 916 // First pass: Do a simple check to see if any functions can have their "..." 917 // removed. We can do this if they never call va_start. This loop cannot be 918 // fused with the next loop, because deleting a function invalidates 919 // information computed while surveying other functions. 920 DOUT << "DAE - Deleting dead varargs\n"; 921 for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { 922 Function &F = *I++; 923 if (F.getFunctionType()->isVarArg()) 924 Changed |= DeleteDeadVarargs(F); 925 } 926 927 // Second phase:loop through the module, determining which arguments are live. 928 // We assume all arguments are dead unless proven otherwise (allowing us to 929 // determine that dead arguments passed into recursive functions are dead). 930 // 931 DOUT << "DAE - Determining liveness\n"; 932 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) 933 SurveyFunction(*I); 934 935 // Now, remove all dead arguments and return values from each function in 936 // turn 937 for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { 938 // Increment now, because the function will probably get removed (ie 939 // replaced by a new one). 940 Function *F = I++; 941 Changed |= RemoveDeadStuffFromFunction(F); 942 } 943 return Changed; 944} 945