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 #include "llvm/Transforms/IPO.h" 21 #include "llvm/ADT/DenseMap.h" 22 #include "llvm/ADT/SmallVector.h" 23 #include "llvm/ADT/Statistic.h" 24 #include "llvm/ADT/StringExtras.h" 25 #include "llvm/IR/CallSite.h" 26 #include "llvm/IR/CallingConv.h" 27 #include "llvm/IR/Constant.h" 28 #include "llvm/IR/DIBuilder.h" 29 #include "llvm/IR/DebugInfo.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/Debug.h" 37 #include "llvm/Support/raw_ostream.h" 38 #include <map> 39 #include <set> 40 #include <tuple> 41 using namespace llvm; 42 43 #define DEBUG_TYPE "deadargelim" 44 45 STATISTIC(NumArgumentsEliminated, "Number of unread args removed"); 46 STATISTIC(NumRetValsEliminated , "Number of unused return values removed"); 47 STATISTIC(NumArgumentsReplacedWithUndef, 48 "Number of unread args replaced with undef"); 49 namespace { 50 /// DAE - The dead argument elimination pass. 51 /// 52 class DAE : public ModulePass { 53 public: 54 55 /// Struct that represents (part of) either a return value or a function 56 /// argument. Used so that arguments and return values can be used 57 /// interchangeably. 58 struct RetOrArg { 59 RetOrArg(const Function *F, unsigned Idx, bool IsArg) : F(F), Idx(Idx), 60 IsArg(IsArg) {} 61 const Function *F; 62 unsigned Idx; 63 bool IsArg; 64 65 /// Make RetOrArg comparable, so we can put it into a map. 66 bool operator<(const RetOrArg &O) const { 67 return std::tie(F, Idx, IsArg) < std::tie(O.F, O.Idx, O.IsArg); 68 } 69 70 /// Make RetOrArg comparable, so we can easily iterate the multimap. 71 bool operator==(const RetOrArg &O) const { 72 return F == O.F && Idx == O.Idx && IsArg == O.IsArg; 73 } 74 75 std::string getDescription() const { 76 return (Twine(IsArg ? "Argument #" : "Return value #") + utostr(Idx) + 77 " of function " + F->getName()).str(); 78 } 79 }; 80 81 /// Liveness enum - During our initial pass over the program, we determine 82 /// that things are either alive or maybe alive. We don't mark anything 83 /// explicitly dead (even if we know they are), since anything not alive 84 /// with no registered uses (in Uses) will never be marked alive and will 85 /// thus become dead in the end. 86 enum Liveness { Live, MaybeLive }; 87 88 /// Convenience wrapper 89 RetOrArg CreateRet(const Function *F, unsigned Idx) { 90 return RetOrArg(F, Idx, false); 91 } 92 /// Convenience wrapper 93 RetOrArg CreateArg(const Function *F, unsigned Idx) { 94 return RetOrArg(F, Idx, true); 95 } 96 97 typedef std::multimap<RetOrArg, RetOrArg> UseMap; 98 /// This maps a return value or argument to any MaybeLive return values or 99 /// arguments it uses. This allows the MaybeLive values to be marked live 100 /// when any of its users is marked live. 101 /// For example (indices are left out for clarity): 102 /// - Uses[ret F] = ret G 103 /// This means that F calls G, and F returns the value returned by G. 104 /// - Uses[arg F] = ret G 105 /// This means that some function calls G and passes its result as an 106 /// argument to F. 107 /// - Uses[ret F] = arg F 108 /// This means that F returns one of its own arguments. 109 /// - Uses[arg F] = arg G 110 /// This means that G calls F and passes one of its own (G's) arguments 111 /// directly to F. 112 UseMap Uses; 113 114 typedef std::set<RetOrArg> LiveSet; 115 typedef std::set<const Function*> LiveFuncSet; 116 117 /// This set contains all values that have been determined to be live. 118 LiveSet LiveValues; 119 /// This set contains all values that are cannot be changed in any way. 120 LiveFuncSet LiveFunctions; 121 122 typedef SmallVector<RetOrArg, 5> UseVector; 123 124 // Map each LLVM function to corresponding metadata with debug info. If 125 // the function is replaced with another one, we should patch the pointer 126 // to LLVM function in metadata. 127 // As the code generation for module is finished (and DIBuilder is 128 // finalized) we assume that subprogram descriptors won't be changed, and 129 // they are stored in map for short duration anyway. 130 DenseMap<const Function *, DISubprogram> FunctionDIs; 131 132 protected: 133 // DAH uses this to specify a different ID. 134 explicit DAE(char &ID) : ModulePass(ID) {} 135 136 public: 137 static char ID; // Pass identification, replacement for typeid 138 DAE() : ModulePass(ID) { 139 initializeDAEPass(*PassRegistry::getPassRegistry()); 140 } 141 142 bool runOnModule(Module &M) override; 143 144 virtual bool ShouldHackArguments() const { return false; } 145 146 private: 147 Liveness MarkIfNotLive(RetOrArg Use, UseVector &MaybeLiveUses); 148 Liveness SurveyUse(const Use *U, UseVector &MaybeLiveUses, 149 unsigned RetValNum = -1U); 150 Liveness SurveyUses(const Value *V, UseVector &MaybeLiveUses); 151 152 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 } 163 164 165 char DAE::ID = 0; 166 INITIALIZE_PASS(DAE, "deadargelim", "Dead Argument Elimination", false, false) 167 168 namespace { 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 } 179 180 char DAH::ID = 0; 181 INITIALIZE_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 /// 188 ModulePass *llvm::createDeadArgEliminationPass() { return new DAE(); } 189 ModulePass *llvm::createDeadArgHackingPass() { return new DAH(); } 190 191 /// DeleteDeadVarargs - If this is an function that takes a ... list, and if 192 /// llvm.vastart is never called, the varargs list is dead for the function. 193 bool DAE::DeleteDeadVarargs(Function &Fn) { 194 assert(Fn.getFunctionType()->isVarArg() && "Function isn't varargs!"); 195 if (Fn.isDeclaration() || !Fn.hasLocalLinkage()) return false; 196 197 // Ensure that the function is only directly called. 198 if (Fn.hasAddressTaken()) 199 return false; 200 201 // Okay, we know we can transform this function if safe. Scan its body 202 // looking for calls marked musttail or calls to llvm.vastart. 203 for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) { 204 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 205 CallInst *CI = dyn_cast<CallInst>(I); 206 if (!CI) 207 continue; 208 if (CI->isMustTailCall()) 209 return false; 210 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) { 211 if (II->getIntrinsicID() == Intrinsic::vastart) 212 return false; 213 } 214 } 215 } 216 217 // If we get here, there are no calls to llvm.vastart in the function body, 218 // remove the "..." and adjust all the calls. 219 220 // Start by computing a new prototype for the function, which is the same as 221 // the old function, but doesn't have isVarArg set. 222 FunctionType *FTy = Fn.getFunctionType(); 223 224 std::vector<Type*> Params(FTy->param_begin(), FTy->param_end()); 225 FunctionType *NFTy = FunctionType::get(FTy->getReturnType(), 226 Params, false); 227 unsigned NumArgs = Params.size(); 228 229 // Create the new function body and insert it into the module... 230 Function *NF = Function::Create(NFTy, Fn.getLinkage()); 231 NF->copyAttributesFrom(&Fn); 232 Fn.getParent()->getFunctionList().insert(&Fn, NF); 233 NF->takeName(&Fn); 234 235 // Loop over all of the callers of the function, transforming the call sites 236 // to pass in a smaller number of arguments into the new function. 237 // 238 std::vector<Value*> Args; 239 for (Value::user_iterator I = Fn.user_begin(), E = Fn.user_end(); I != E; ) { 240 CallSite CS(*I++); 241 if (!CS) 242 continue; 243 Instruction *Call = CS.getInstruction(); 244 245 // Pass all the same arguments. 246 Args.assign(CS.arg_begin(), CS.arg_begin() + NumArgs); 247 248 // Drop any attributes that were on the vararg arguments. 249 AttributeSet PAL = CS.getAttributes(); 250 if (!PAL.isEmpty() && PAL.getSlotIndex(PAL.getNumSlots() - 1) > NumArgs) { 251 SmallVector<AttributeSet, 8> AttributesVec; 252 for (unsigned i = 0; PAL.getSlotIndex(i) <= NumArgs; ++i) 253 AttributesVec.push_back(PAL.getSlotAttributes(i)); 254 if (PAL.hasAttributes(AttributeSet::FunctionIndex)) 255 AttributesVec.push_back(AttributeSet::get(Fn.getContext(), 256 PAL.getFnAttributes())); 257 PAL = AttributeSet::get(Fn.getContext(), AttributesVec); 258 } 259 260 Instruction *New; 261 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 262 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 263 Args, "", Call); 264 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 265 cast<InvokeInst>(New)->setAttributes(PAL); 266 } else { 267 New = CallInst::Create(NF, Args, "", Call); 268 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 269 cast<CallInst>(New)->setAttributes(PAL); 270 if (cast<CallInst>(Call)->isTailCall()) 271 cast<CallInst>(New)->setTailCall(); 272 } 273 New->setDebugLoc(Call->getDebugLoc()); 274 275 Args.clear(); 276 277 if (!Call->use_empty()) 278 Call->replaceAllUsesWith(New); 279 280 New->takeName(Call); 281 282 // Finally, remove the old call from the program, reducing the use-count of 283 // F. 284 Call->eraseFromParent(); 285 } 286 287 // Since we have now created the new function, splice the body of the old 288 // function right into the new function, leaving the old rotting hulk of the 289 // function empty. 290 NF->getBasicBlockList().splice(NF->begin(), Fn.getBasicBlockList()); 291 292 // Loop over the argument list, transferring uses of the old arguments over to 293 // the new arguments, also transferring over the names as well. While we're at 294 // it, remove the dead arguments from the DeadArguments list. 295 // 296 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(), 297 I2 = NF->arg_begin(); I != E; ++I, ++I2) { 298 // Move the name and users over to the new version. 299 I->replaceAllUsesWith(I2); 300 I2->takeName(I); 301 } 302 303 // Patch the pointer to LLVM function in debug info descriptor. 304 auto DI = FunctionDIs.find(&Fn); 305 if (DI != FunctionDIs.end()) { 306 DISubprogram SP = DI->second; 307 SP->replaceFunction(NF); 308 // Ensure the map is updated so it can be reused on non-varargs argument 309 // eliminations of the same function. 310 FunctionDIs.erase(DI); 311 FunctionDIs[NF] = SP; 312 } 313 314 // Fix up any BlockAddresses that refer to the function. 315 Fn.replaceAllUsesWith(ConstantExpr::getBitCast(NF, Fn.getType())); 316 // Delete the bitcast that we just created, so that NF does not 317 // appear to be address-taken. 318 NF->removeDeadConstantUsers(); 319 // Finally, nuke the old function. 320 Fn.eraseFromParent(); 321 return true; 322 } 323 324 /// RemoveDeadArgumentsFromCallers - Checks if the given function has any 325 /// arguments that are unused, and changes the caller parameters to be undefined 326 /// instead. 327 bool DAE::RemoveDeadArgumentsFromCallers(Function &Fn) 328 { 329 if (Fn.isDeclaration() || Fn.mayBeOverridden()) 330 return false; 331 332 // Functions with local linkage should already have been handled, except the 333 // fragile (variadic) ones which we can improve here. 334 if (Fn.hasLocalLinkage() && !Fn.getFunctionType()->isVarArg()) 335 return false; 336 337 // If a function seen at compile time is not necessarily the one linked to 338 // the binary being built, it is illegal to change the actual arguments 339 // passed to it. These functions can be captured by isWeakForLinker(). 340 // *NOTE* that mayBeOverridden() is insufficient for this purpose as it 341 // doesn't include linkage types like AvailableExternallyLinkage and 342 // LinkOnceODRLinkage. Take link_odr* as an example, it indicates a set of 343 // *EQUIVALENT* globals that can be merged at link-time. However, the 344 // semantic of *EQUIVALENT*-functions includes parameters. Changing 345 // parameters breaks this assumption. 346 // 347 if (Fn.isWeakForLinker()) 348 return false; 349 350 if (Fn.use_empty()) 351 return false; 352 353 SmallVector<unsigned, 8> UnusedArgs; 354 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(); 355 I != E; ++I) { 356 Argument *Arg = I; 357 358 if (Arg->use_empty() && !Arg->hasByValOrInAllocaAttr()) 359 UnusedArgs.push_back(Arg->getArgNo()); 360 } 361 362 if (UnusedArgs.empty()) 363 return false; 364 365 bool Changed = false; 366 367 for (Use &U : Fn.uses()) { 368 CallSite CS(U.getUser()); 369 if (!CS || !CS.isCallee(&U)) 370 continue; 371 372 // Now go through all unused args and replace them with "undef". 373 for (unsigned I = 0, E = UnusedArgs.size(); I != E; ++I) { 374 unsigned ArgNo = UnusedArgs[I]; 375 376 Value *Arg = CS.getArgument(ArgNo); 377 CS.setArgument(ArgNo, UndefValue::get(Arg->getType())); 378 ++NumArgumentsReplacedWithUndef; 379 Changed = true; 380 } 381 } 382 383 return Changed; 384 } 385 386 /// Convenience function that returns the number of return values. It returns 0 387 /// for void functions and 1 for functions not returning a struct. It returns 388 /// the number of struct elements for functions returning a struct. 389 static unsigned NumRetVals(const Function *F) { 390 Type *RetTy = F->getReturnType(); 391 if (RetTy->isVoidTy()) 392 return 0; 393 else if (StructType *STy = dyn_cast<StructType>(RetTy)) 394 return STy->getNumElements(); 395 else if (ArrayType *ATy = dyn_cast<ArrayType>(RetTy)) 396 return ATy->getNumElements(); 397 else 398 return 1; 399 } 400 401 /// Returns the sub-type a function will return at a given Idx. Should 402 /// correspond to the result type of an ExtractValue instruction executed with 403 /// just that one Idx (i.e. only top-level structure is considered). 404 static Type *getRetComponentType(const Function *F, unsigned Idx) { 405 Type *RetTy = F->getReturnType(); 406 assert(!RetTy->isVoidTy() && "void type has no subtype"); 407 408 if (StructType *STy = dyn_cast<StructType>(RetTy)) 409 return STy->getElementType(Idx); 410 else if (ArrayType *ATy = dyn_cast<ArrayType>(RetTy)) 411 return ATy->getElementType(); 412 else 413 return RetTy; 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. 419 DAE::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. 438 DAE::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 const Function *F = RI->getParent()->getParent(); 447 if (RetValNum != -1U) { 448 RetOrArg Use = CreateRet(F, RetValNum); 449 // We might be live, depending on the liveness of Use. 450 return MarkIfNotLive(Use, MaybeLiveUses); 451 } else { 452 DAE::Liveness Result = MaybeLive; 453 for (unsigned i = 0; i < NumRetVals(F); ++i) { 454 RetOrArg Use = CreateRet(F, i); 455 // We might be live, depending on the liveness of Use. If any 456 // sub-value is live, then the entire value is considered live. This 457 // is a conservative choice, and better tracking is possible. 458 DAE::Liveness SubResult = MarkIfNotLive(Use, MaybeLiveUses); 459 if (Result != Live) 460 Result = SubResult; 461 } 462 return Result; 463 } 464 } 465 if (const InsertValueInst *IV = dyn_cast<InsertValueInst>(V)) { 466 if (U->getOperandNo() != InsertValueInst::getAggregateOperandIndex() 467 && IV->hasIndices()) 468 // The use we are examining is inserted into an aggregate. Our liveness 469 // depends on all uses of that aggregate, but if it is used as a return 470 // value, only index at which we were inserted counts. 471 RetValNum = *IV->idx_begin(); 472 473 // Note that if we are used as the aggregate operand to the insertvalue, 474 // we don't change RetValNum, but do survey all our uses. 475 476 Liveness Result = MaybeLive; 477 for (const Use &UU : IV->uses()) { 478 Result = SurveyUse(&UU, MaybeLiveUses, RetValNum); 479 if (Result == Live) 480 break; 481 } 482 return Result; 483 } 484 485 if (auto CS = ImmutableCallSite(V)) { 486 const Function *F = CS.getCalledFunction(); 487 if (F) { 488 // Used in a direct call. 489 490 // Find the argument number. We know for sure that this use is an 491 // argument, since if it was the function argument this would be an 492 // indirect call and the we know can't be looking at a value of the 493 // label type (for the invoke instruction). 494 unsigned ArgNo = CS.getArgumentNo(U); 495 496 if (ArgNo >= F->getFunctionType()->getNumParams()) 497 // The value is passed in through a vararg! Must be live. 498 return Live; 499 500 assert(CS.getArgument(ArgNo) 501 == CS->getOperand(U->getOperandNo()) 502 && "Argument is not where we expected it"); 503 504 // Value passed to a normal call. It's only live when the corresponding 505 // argument to the called function turns out live. 506 RetOrArg Use = CreateArg(F, ArgNo); 507 return MarkIfNotLive(Use, MaybeLiveUses); 508 } 509 } 510 // Used in any other way? Value must be live. 511 return Live; 512 } 513 514 /// SurveyUses - This looks at all the uses of the given value 515 /// Returns the Liveness deduced from the uses of this value. 516 /// 517 /// Adds all uses that cause the result to be MaybeLive to MaybeLiveRetUses. If 518 /// the result is Live, MaybeLiveUses might be modified but its content should 519 /// be ignored (since it might not be complete). 520 DAE::Liveness DAE::SurveyUses(const Value *V, UseVector &MaybeLiveUses) { 521 // Assume it's dead (which will only hold if there are no uses at all..). 522 Liveness Result = MaybeLive; 523 // Check each use. 524 for (const Use &U : V->uses()) { 525 Result = SurveyUse(&U, MaybeLiveUses); 526 if (Result == Live) 527 break; 528 } 529 return Result; 530 } 531 532 // SurveyFunction - This performs the initial survey of the specified function, 533 // checking out whether or not it uses any of its incoming arguments or whether 534 // any callers use the return value. This fills in the LiveValues set and Uses 535 // map. 536 // 537 // We consider arguments of non-internal functions to be intrinsically alive as 538 // well as arguments to functions which have their "address taken". 539 // 540 void DAE::SurveyFunction(const Function &F) { 541 // Functions with inalloca parameters are expecting args in a particular 542 // register and memory layout. 543 if (F.getAttributes().hasAttrSomewhere(Attribute::InAlloca)) { 544 MarkLive(F); 545 return; 546 } 547 548 unsigned RetCount = NumRetVals(&F); 549 // Assume all return values are dead 550 typedef SmallVector<Liveness, 5> RetVals; 551 RetVals RetValLiveness(RetCount, MaybeLive); 552 553 typedef SmallVector<UseVector, 5> RetUses; 554 // These vectors map each return value to the uses that make it MaybeLive, so 555 // we can add those to the Uses map if the return value really turns out to be 556 // MaybeLive. Initialized to a list of RetCount empty lists. 557 RetUses MaybeLiveRetUses(RetCount); 558 559 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 560 if (const ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) 561 if (RI->getNumOperands() != 0 && RI->getOperand(0)->getType() 562 != F.getFunctionType()->getReturnType()) { 563 // We don't support old style multiple return values. 564 MarkLive(F); 565 return; 566 } 567 568 if (!F.hasLocalLinkage() && (!ShouldHackArguments() || F.isIntrinsic())) { 569 MarkLive(F); 570 return; 571 } 572 573 DEBUG(dbgs() << "DAE - Inspecting callers for fn: " << F.getName() << "\n"); 574 // Keep track of the number of live retvals, so we can skip checks once all 575 // of them turn out to be live. 576 unsigned NumLiveRetVals = 0; 577 // Loop all uses of the function. 578 for (const Use &U : F.uses()) { 579 // If the function is PASSED IN as an argument, its address has been 580 // taken. 581 ImmutableCallSite CS(U.getUser()); 582 if (!CS || !CS.isCallee(&U)) { 583 MarkLive(F); 584 return; 585 } 586 587 // If this use is anything other than a call site, the function is alive. 588 const Instruction *TheCall = CS.getInstruction(); 589 if (!TheCall) { // Not a direct call site? 590 MarkLive(F); 591 return; 592 } 593 594 // If we end up here, we are looking at a direct call to our function. 595 596 // Now, check how our return value(s) is/are used in this caller. Don't 597 // bother checking return values if all of them are live already. 598 if (NumLiveRetVals == RetCount) 599 continue; 600 601 // Check all uses of the return value. 602 for (const Use &U : TheCall->uses()) { 603 if (ExtractValueInst *Ext = dyn_cast<ExtractValueInst>(U.getUser())) { 604 // This use uses a part of our return value, survey the uses of 605 // that part and store the results for this index only. 606 unsigned Idx = *Ext->idx_begin(); 607 if (RetValLiveness[Idx] != Live) { 608 RetValLiveness[Idx] = SurveyUses(Ext, MaybeLiveRetUses[Idx]); 609 if (RetValLiveness[Idx] == Live) 610 NumLiveRetVals++; 611 } 612 } else { 613 // Used by something else than extractvalue. Survey, but assume that the 614 // result applies to all sub-values. 615 UseVector MaybeLiveAggregateUses; 616 if (SurveyUse(&U, MaybeLiveAggregateUses) == Live) { 617 NumLiveRetVals = RetCount; 618 RetValLiveness.assign(RetCount, Live); 619 break; 620 } else { 621 for (unsigned i = 0; i != RetCount; ++i) { 622 if (RetValLiveness[i] != Live) 623 MaybeLiveRetUses[i].append(MaybeLiveAggregateUses.begin(), 624 MaybeLiveAggregateUses.end()); 625 } 626 } 627 } 628 } 629 } 630 631 // Now we've inspected all callers, record the liveness of our return values. 632 for (unsigned i = 0; i != RetCount; ++i) 633 MarkValue(CreateRet(&F, i), RetValLiveness[i], MaybeLiveRetUses[i]); 634 635 DEBUG(dbgs() << "DAE - Inspecting args for fn: " << F.getName() << "\n"); 636 637 // Now, check all of our arguments. 638 unsigned i = 0; 639 UseVector MaybeLiveArgUses; 640 for (Function::const_arg_iterator AI = F.arg_begin(), 641 E = F.arg_end(); AI != E; ++AI, ++i) { 642 Liveness Result; 643 if (F.getFunctionType()->isVarArg()) { 644 // Variadic functions will already have a va_arg function expanded inside 645 // them, making them potentially very sensitive to ABI changes resulting 646 // from removing arguments entirely, so don't. For example AArch64 handles 647 // register and stack HFAs very differently, and this is reflected in the 648 // IR which has already been generated. 649 Result = Live; 650 } else { 651 // See what the effect of this use is (recording any uses that cause 652 // MaybeLive in MaybeLiveArgUses). 653 Result = SurveyUses(AI, MaybeLiveArgUses); 654 } 655 656 // Mark the result. 657 MarkValue(CreateArg(&F, i), Result, MaybeLiveArgUses); 658 // Clear the vector again for the next iteration. 659 MaybeLiveArgUses.clear(); 660 } 661 } 662 663 /// MarkValue - This function marks the liveness of RA depending on L. If L is 664 /// MaybeLive, it also takes all uses in MaybeLiveUses and records them in Uses, 665 /// such that RA will be marked live if any use in MaybeLiveUses gets marked 666 /// live later on. 667 void DAE::MarkValue(const RetOrArg &RA, Liveness L, 668 const UseVector &MaybeLiveUses) { 669 switch (L) { 670 case Live: MarkLive(RA); break; 671 case MaybeLive: 672 { 673 // Note any uses of this value, so this return value can be 674 // marked live whenever one of the uses becomes live. 675 for (UseVector::const_iterator UI = MaybeLiveUses.begin(), 676 UE = MaybeLiveUses.end(); UI != UE; ++UI) 677 Uses.insert(std::make_pair(*UI, RA)); 678 break; 679 } 680 } 681 } 682 683 /// MarkLive - Mark the given Function as alive, meaning that it cannot be 684 /// changed in any way. Additionally, 685 /// mark any values that are used as this function's parameters or by its return 686 /// values (according to Uses) live as well. 687 void DAE::MarkLive(const Function &F) { 688 DEBUG(dbgs() << "DAE - Intrinsically live fn: " << F.getName() << "\n"); 689 // Mark the function as live. 690 LiveFunctions.insert(&F); 691 // Mark all arguments as live. 692 for (unsigned i = 0, e = F.arg_size(); i != e; ++i) 693 PropagateLiveness(CreateArg(&F, i)); 694 // Mark all return values as live. 695 for (unsigned i = 0, e = NumRetVals(&F); i != e; ++i) 696 PropagateLiveness(CreateRet(&F, i)); 697 } 698 699 /// MarkLive - Mark the given return value or argument as live. Additionally, 700 /// mark any values that are used by this value (according to Uses) live as 701 /// well. 702 void DAE::MarkLive(const RetOrArg &RA) { 703 if (LiveFunctions.count(RA.F)) 704 return; // Function was already marked Live. 705 706 if (!LiveValues.insert(RA).second) 707 return; // We were already marked Live. 708 709 DEBUG(dbgs() << "DAE - Marking " << RA.getDescription() << " live\n"); 710 PropagateLiveness(RA); 711 } 712 713 /// PropagateLiveness - Given that RA is a live value, propagate it's liveness 714 /// to any other values it uses (according to Uses). 715 void DAE::PropagateLiveness(const RetOrArg &RA) { 716 // We don't use upper_bound (or equal_range) here, because our recursive call 717 // to ourselves is likely to cause the upper_bound (which is the first value 718 // not belonging to RA) to become erased and the iterator invalidated. 719 UseMap::iterator Begin = Uses.lower_bound(RA); 720 UseMap::iterator E = Uses.end(); 721 UseMap::iterator I; 722 for (I = Begin; I != E && I->first == RA; ++I) 723 MarkLive(I->second); 724 725 // Erase RA from the Uses map (from the lower bound to wherever we ended up 726 // after the loop). 727 Uses.erase(Begin, I); 728 } 729 730 // RemoveDeadStuffFromFunction - Remove any arguments and return values from F 731 // that are not in LiveValues. Transform the function and all of the callees of 732 // the function to not have these arguments and return values. 733 // 734 bool DAE::RemoveDeadStuffFromFunction(Function *F) { 735 // Don't modify fully live functions 736 if (LiveFunctions.count(F)) 737 return false; 738 739 // Start by computing a new prototype for the function, which is the same as 740 // the old function, but has fewer arguments and a different return type. 741 FunctionType *FTy = F->getFunctionType(); 742 std::vector<Type*> Params; 743 744 // Keep track of if we have a live 'returned' argument 745 bool HasLiveReturnedArg = false; 746 747 // Set up to build a new list of parameter attributes. 748 SmallVector<AttributeSet, 8> AttributesVec; 749 const AttributeSet &PAL = F->getAttributes(); 750 751 // Remember which arguments are still alive. 752 SmallVector<bool, 10> ArgAlive(FTy->getNumParams(), false); 753 // Construct the new parameter list from non-dead arguments. Also construct 754 // a new set of parameter attributes to correspond. Skip the first parameter 755 // attribute, since that belongs to the return value. 756 unsigned i = 0; 757 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 758 I != E; ++I, ++i) { 759 RetOrArg Arg = CreateArg(F, i); 760 if (LiveValues.erase(Arg)) { 761 Params.push_back(I->getType()); 762 ArgAlive[i] = true; 763 764 // Get the original parameter attributes (skipping the first one, that is 765 // for the return value. 766 if (PAL.hasAttributes(i + 1)) { 767 AttrBuilder B(PAL, i + 1); 768 if (B.contains(Attribute::Returned)) 769 HasLiveReturnedArg = true; 770 AttributesVec. 771 push_back(AttributeSet::get(F->getContext(), Params.size(), B)); 772 } 773 } else { 774 ++NumArgumentsEliminated; 775 DEBUG(dbgs() << "DAE - Removing argument " << i << " (" << I->getName() 776 << ") from " << F->getName() << "\n"); 777 } 778 } 779 780 // Find out the new return value. 781 Type *RetTy = FTy->getReturnType(); 782 Type *NRetTy = nullptr; 783 unsigned RetCount = NumRetVals(F); 784 785 // -1 means unused, other numbers are the new index 786 SmallVector<int, 5> NewRetIdxs(RetCount, -1); 787 std::vector<Type*> RetTypes; 788 789 // If there is a function with a live 'returned' argument but a dead return 790 // value, then there are two possible actions: 791 // 1) Eliminate the return value and take off the 'returned' attribute on the 792 // argument. 793 // 2) Retain the 'returned' attribute and treat the return value (but not the 794 // entire function) as live so that it is not eliminated. 795 // 796 // It's not clear in the general case which option is more profitable because, 797 // even in the absence of explicit uses of the return value, code generation 798 // is free to use the 'returned' attribute to do things like eliding 799 // save/restores of registers across calls. Whether or not this happens is 800 // target and ABI-specific as well as depending on the amount of register 801 // pressure, so there's no good way for an IR-level pass to figure this out. 802 // 803 // Fortunately, the only places where 'returned' is currently generated by 804 // the FE are places where 'returned' is basically free and almost always a 805 // performance win, so the second option can just be used always for now. 806 // 807 // This should be revisited if 'returned' is ever applied more liberally. 808 if (RetTy->isVoidTy() || HasLiveReturnedArg) { 809 NRetTy = RetTy; 810 } else { 811 // Look at each of the original return values individually. 812 for (unsigned i = 0; i != RetCount; ++i) { 813 RetOrArg Ret = CreateRet(F, i); 814 if (LiveValues.erase(Ret)) { 815 RetTypes.push_back(getRetComponentType(F, i)); 816 NewRetIdxs[i] = RetTypes.size() - 1; 817 } else { 818 ++NumRetValsEliminated; 819 DEBUG(dbgs() << "DAE - Removing return value " << i << " from " 820 << F->getName() << "\n"); 821 } 822 } 823 if (RetTypes.size() > 1) { 824 // More than one return type? Reduce it down to size. 825 if (StructType *STy = dyn_cast<StructType>(RetTy)) { 826 // Make the new struct packed if we used to return a packed struct 827 // already. 828 NRetTy = StructType::get(STy->getContext(), RetTypes, STy->isPacked()); 829 } else { 830 assert(isa<ArrayType>(RetTy) && "unexpected multi-value return"); 831 NRetTy = ArrayType::get(RetTypes[0], RetTypes.size()); 832 } 833 } else if (RetTypes.size() == 1) 834 // One return type? Just a simple value then, but only if we didn't use to 835 // return a struct with that simple value before. 836 NRetTy = RetTypes.front(); 837 else if (RetTypes.size() == 0) 838 // No return types? Make it void, but only if we didn't use to return {}. 839 NRetTy = Type::getVoidTy(F->getContext()); 840 } 841 842 assert(NRetTy && "No new return type found?"); 843 844 // The existing function return attributes. 845 AttributeSet RAttrs = PAL.getRetAttributes(); 846 847 // Remove any incompatible attributes, but only if we removed all return 848 // values. Otherwise, ensure that we don't have any conflicting attributes 849 // here. Currently, this should not be possible, but special handling might be 850 // required when new return value attributes are added. 851 if (NRetTy->isVoidTy()) 852 RAttrs = 853 AttributeSet::get(NRetTy->getContext(), AttributeSet::ReturnIndex, 854 AttrBuilder(RAttrs, AttributeSet::ReturnIndex). 855 removeAttributes(AttributeFuncs:: 856 typeIncompatible(NRetTy, AttributeSet::ReturnIndex), 857 AttributeSet::ReturnIndex)); 858 else 859 assert(!AttrBuilder(RAttrs, AttributeSet::ReturnIndex). 860 hasAttributes(AttributeFuncs:: 861 typeIncompatible(NRetTy, AttributeSet::ReturnIndex), 862 AttributeSet::ReturnIndex) && 863 "Return attributes no longer compatible?"); 864 865 if (RAttrs.hasAttributes(AttributeSet::ReturnIndex)) 866 AttributesVec.push_back(AttributeSet::get(NRetTy->getContext(), RAttrs)); 867 868 if (PAL.hasAttributes(AttributeSet::FunctionIndex)) 869 AttributesVec.push_back(AttributeSet::get(F->getContext(), 870 PAL.getFnAttributes())); 871 872 // Reconstruct the AttributesList based on the vector we constructed. 873 AttributeSet NewPAL = AttributeSet::get(F->getContext(), AttributesVec); 874 875 // Create the new function type based on the recomputed parameters. 876 FunctionType *NFTy = FunctionType::get(NRetTy, Params, FTy->isVarArg()); 877 878 // No change? 879 if (NFTy == FTy) 880 return false; 881 882 // Create the new function body and insert it into the module... 883 Function *NF = Function::Create(NFTy, F->getLinkage()); 884 NF->copyAttributesFrom(F); 885 NF->setAttributes(NewPAL); 886 // Insert the new function before the old function, so we won't be processing 887 // it again. 888 F->getParent()->getFunctionList().insert(F, NF); 889 NF->takeName(F); 890 891 // Loop over all of the callers of the function, transforming the call sites 892 // to pass in a smaller number of arguments into the new function. 893 // 894 std::vector<Value*> Args; 895 while (!F->use_empty()) { 896 CallSite CS(F->user_back()); 897 Instruction *Call = CS.getInstruction(); 898 899 AttributesVec.clear(); 900 const AttributeSet &CallPAL = CS.getAttributes(); 901 902 // The call return attributes. 903 AttributeSet RAttrs = CallPAL.getRetAttributes(); 904 905 // Adjust in case the function was changed to return void. 906 RAttrs = 907 AttributeSet::get(NF->getContext(), AttributeSet::ReturnIndex, 908 AttrBuilder(RAttrs, AttributeSet::ReturnIndex). 909 removeAttributes(AttributeFuncs:: 910 typeIncompatible(NF->getReturnType(), 911 AttributeSet::ReturnIndex), 912 AttributeSet::ReturnIndex)); 913 if (RAttrs.hasAttributes(AttributeSet::ReturnIndex)) 914 AttributesVec.push_back(AttributeSet::get(NF->getContext(), RAttrs)); 915 916 // Declare these outside of the loops, so we can reuse them for the second 917 // loop, which loops the varargs. 918 CallSite::arg_iterator I = CS.arg_begin(); 919 unsigned i = 0; 920 // Loop over those operands, corresponding to the normal arguments to the 921 // original function, and add those that are still alive. 922 for (unsigned e = FTy->getNumParams(); i != e; ++I, ++i) 923 if (ArgAlive[i]) { 924 Args.push_back(*I); 925 // Get original parameter attributes, but skip return attributes. 926 if (CallPAL.hasAttributes(i + 1)) { 927 AttrBuilder B(CallPAL, i + 1); 928 // If the return type has changed, then get rid of 'returned' on the 929 // call site. The alternative is to make all 'returned' attributes on 930 // call sites keep the return value alive just like 'returned' 931 // attributes on function declaration but it's less clearly a win 932 // and this is not an expected case anyway 933 if (NRetTy != RetTy && B.contains(Attribute::Returned)) 934 B.removeAttribute(Attribute::Returned); 935 AttributesVec. 936 push_back(AttributeSet::get(F->getContext(), Args.size(), B)); 937 } 938 } 939 940 // Push any varargs arguments on the list. Don't forget their attributes. 941 for (CallSite::arg_iterator E = CS.arg_end(); I != E; ++I, ++i) { 942 Args.push_back(*I); 943 if (CallPAL.hasAttributes(i + 1)) { 944 AttrBuilder B(CallPAL, i + 1); 945 AttributesVec. 946 push_back(AttributeSet::get(F->getContext(), Args.size(), B)); 947 } 948 } 949 950 if (CallPAL.hasAttributes(AttributeSet::FunctionIndex)) 951 AttributesVec.push_back(AttributeSet::get(Call->getContext(), 952 CallPAL.getFnAttributes())); 953 954 // Reconstruct the AttributesList based on the vector we constructed. 955 AttributeSet NewCallPAL = AttributeSet::get(F->getContext(), AttributesVec); 956 957 Instruction *New; 958 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 959 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 960 Args, "", Call); 961 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 962 cast<InvokeInst>(New)->setAttributes(NewCallPAL); 963 } else { 964 New = CallInst::Create(NF, Args, "", Call); 965 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 966 cast<CallInst>(New)->setAttributes(NewCallPAL); 967 if (cast<CallInst>(Call)->isTailCall()) 968 cast<CallInst>(New)->setTailCall(); 969 } 970 New->setDebugLoc(Call->getDebugLoc()); 971 972 Args.clear(); 973 974 if (!Call->use_empty()) { 975 if (New->getType() == Call->getType()) { 976 // Return type not changed? Just replace users then. 977 Call->replaceAllUsesWith(New); 978 New->takeName(Call); 979 } else if (New->getType()->isVoidTy()) { 980 // Our return value has uses, but they will get removed later on. 981 // Replace by null for now. 982 if (!Call->getType()->isX86_MMXTy()) 983 Call->replaceAllUsesWith(Constant::getNullValue(Call->getType())); 984 } else { 985 assert((RetTy->isStructTy() || RetTy->isArrayTy()) && 986 "Return type changed, but not into a void. The old return type" 987 " must have been a struct or an array!"); 988 Instruction *InsertPt = Call; 989 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 990 BasicBlock::iterator IP = II->getNormalDest()->begin(); 991 while (isa<PHINode>(IP)) ++IP; 992 InsertPt = IP; 993 } 994 995 // We used to return a struct or array. Instead of doing smart stuff 996 // with all the uses, we will just rebuild it using extract/insertvalue 997 // chaining and let instcombine clean that up. 998 // 999 // Start out building up our return value from undef 1000 Value *RetVal = UndefValue::get(RetTy); 1001 for (unsigned i = 0; i != RetCount; ++i) 1002 if (NewRetIdxs[i] != -1) { 1003 Value *V; 1004 if (RetTypes.size() > 1) 1005 // We are still returning a struct, so extract the value from our 1006 // return value 1007 V = ExtractValueInst::Create(New, NewRetIdxs[i], "newret", 1008 InsertPt); 1009 else 1010 // We are now returning a single element, so just insert that 1011 V = New; 1012 // Insert the value at the old position 1013 RetVal = InsertValueInst::Create(RetVal, V, i, "oldret", InsertPt); 1014 } 1015 // Now, replace all uses of the old call instruction with the return 1016 // struct we built 1017 Call->replaceAllUsesWith(RetVal); 1018 New->takeName(Call); 1019 } 1020 } 1021 1022 // Finally, remove the old call from the program, reducing the use-count of 1023 // F. 1024 Call->eraseFromParent(); 1025 } 1026 1027 // Since we have now created the new function, splice the body of the old 1028 // function right into the new function, leaving the old rotting hulk of the 1029 // function empty. 1030 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 1031 1032 // Loop over the argument list, transferring uses of the old arguments over to 1033 // the new arguments, also transferring over the names as well. 1034 i = 0; 1035 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), 1036 I2 = NF->arg_begin(); I != E; ++I, ++i) 1037 if (ArgAlive[i]) { 1038 // If this is a live argument, move the name and users over to the new 1039 // version. 1040 I->replaceAllUsesWith(I2); 1041 I2->takeName(I); 1042 ++I2; 1043 } else { 1044 // If this argument is dead, replace any uses of it with null constants 1045 // (these are guaranteed to become unused later on). 1046 if (!I->getType()->isX86_MMXTy()) 1047 I->replaceAllUsesWith(Constant::getNullValue(I->getType())); 1048 } 1049 1050 // If we change the return value of the function we must rewrite any return 1051 // instructions. Check this now. 1052 if (F->getReturnType() != NF->getReturnType()) 1053 for (Function::iterator BB = NF->begin(), E = NF->end(); BB != E; ++BB) 1054 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 1055 Value *RetVal; 1056 1057 if (NFTy->getReturnType()->isVoidTy()) { 1058 RetVal = nullptr; 1059 } else { 1060 assert(RetTy->isStructTy() || RetTy->isArrayTy()); 1061 // The original return value was a struct or array, insert 1062 // extractvalue/insertvalue chains to extract only the values we need 1063 // to return and insert them into our new result. 1064 // This does generate messy code, but we'll let it to instcombine to 1065 // clean that up. 1066 Value *OldRet = RI->getOperand(0); 1067 // Start out building up our return value from undef 1068 RetVal = UndefValue::get(NRetTy); 1069 for (unsigned i = 0; i != RetCount; ++i) 1070 if (NewRetIdxs[i] != -1) { 1071 ExtractValueInst *EV = ExtractValueInst::Create(OldRet, i, 1072 "oldret", RI); 1073 if (RetTypes.size() > 1) { 1074 // We're still returning a struct, so reinsert the value into 1075 // our new return value at the new index 1076 1077 RetVal = InsertValueInst::Create(RetVal, EV, NewRetIdxs[i], 1078 "newret", RI); 1079 } else { 1080 // We are now only returning a simple value, so just return the 1081 // extracted value. 1082 RetVal = EV; 1083 } 1084 } 1085 } 1086 // Replace the return instruction with one returning the new return 1087 // value (possibly 0 if we became void). 1088 ReturnInst::Create(F->getContext(), RetVal, RI); 1089 BB->getInstList().erase(RI); 1090 } 1091 1092 // Patch the pointer to LLVM function in debug info descriptor. 1093 auto DI = FunctionDIs.find(F); 1094 if (DI != FunctionDIs.end()) 1095 DI->second->replaceFunction(NF); 1096 1097 // Now that the old function is dead, delete it. 1098 F->eraseFromParent(); 1099 1100 return true; 1101 } 1102 1103 bool DAE::runOnModule(Module &M) { 1104 bool Changed = false; 1105 1106 // Collect debug info descriptors for functions. 1107 FunctionDIs = makeSubprogramMap(M); 1108 1109 // First pass: Do a simple check to see if any functions can have their "..." 1110 // removed. We can do this if they never call va_start. This loop cannot be 1111 // fused with the next loop, because deleting a function invalidates 1112 // information computed while surveying other functions. 1113 DEBUG(dbgs() << "DAE - Deleting dead varargs\n"); 1114 for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { 1115 Function &F = *I++; 1116 if (F.getFunctionType()->isVarArg()) 1117 Changed |= DeleteDeadVarargs(F); 1118 } 1119 1120 // Second phase:loop through the module, determining which arguments are live. 1121 // We assume all arguments are dead unless proven otherwise (allowing us to 1122 // determine that dead arguments passed into recursive functions are dead). 1123 // 1124 DEBUG(dbgs() << "DAE - Determining liveness\n"); 1125 for (auto &F : M) 1126 SurveyFunction(F); 1127 1128 // Now, remove all dead arguments and return values from each function in 1129 // turn. 1130 for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { 1131 // Increment now, because the function will probably get removed (ie. 1132 // replaced by a new one). 1133 Function *F = I++; 1134 Changed |= RemoveDeadStuffFromFunction(F); 1135 } 1136 1137 // Finally, look for any unused parameters in functions with non-local 1138 // linkage and replace the passed in parameters with undef. 1139 for (auto &F : M) 1140 Changed |= RemoveDeadArgumentsFromCallers(F); 1141 1142 return Changed; 1143 } 1144