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 "llvm/Transforms/Utils/BasicBlockUtils.h" 39 #include <map> 40 #include <set> 41 #include <tuple> 42 using namespace llvm; 43 44 #define DEBUG_TYPE "deadargelim" 45 46 STATISTIC(NumArgumentsEliminated, "Number of unread args removed"); 47 STATISTIC(NumRetValsEliminated , "Number of unused return values removed"); 48 STATISTIC(NumArgumentsReplacedWithUndef, 49 "Number of unread args replaced with undef"); 50 namespace { 51 /// DAE - The dead argument elimination pass. 52 /// 53 class DAE : public ModulePass { 54 public: 55 56 /// Struct that represents (part of) either a return value or a function 57 /// argument. Used so that arguments and return values can be used 58 /// interchangeably. 59 struct RetOrArg { 60 RetOrArg(const Function *F, unsigned Idx, bool IsArg) : F(F), Idx(Idx), 61 IsArg(IsArg) {} 62 const Function *F; 63 unsigned Idx; 64 bool IsArg; 65 66 /// Make RetOrArg comparable, so we can put it into a map. 67 bool operator<(const RetOrArg &O) const { 68 return std::tie(F, Idx, IsArg) < std::tie(O.F, O.Idx, O.IsArg); 69 } 70 71 /// Make RetOrArg comparable, so we can easily iterate the multimap. 72 bool operator==(const RetOrArg &O) const { 73 return F == O.F && Idx == O.Idx && IsArg == O.IsArg; 74 } 75 76 std::string getDescription() const { 77 return (Twine(IsArg ? "Argument #" : "Return value #") + utostr(Idx) + 78 " of function " + F->getName()).str(); 79 } 80 }; 81 82 /// Liveness enum - During our initial pass over the program, we determine 83 /// that things are either alive or maybe alive. We don't mark anything 84 /// explicitly dead (even if we know they are), since anything not alive 85 /// with no registered uses (in Uses) will never be marked alive and will 86 /// thus become dead in the end. 87 enum Liveness { Live, MaybeLive }; 88 89 /// Convenience wrapper 90 RetOrArg CreateRet(const Function *F, unsigned Idx) { 91 return RetOrArg(F, Idx, false); 92 } 93 /// Convenience wrapper 94 RetOrArg CreateArg(const Function *F, unsigned Idx) { 95 return RetOrArg(F, Idx, true); 96 } 97 98 typedef std::multimap<RetOrArg, RetOrArg> UseMap; 99 /// This maps a return value or argument to any MaybeLive return values or 100 /// arguments it uses. This allows the MaybeLive values to be marked live 101 /// when any of its users is marked live. 102 /// For example (indices are left out for clarity): 103 /// - Uses[ret F] = ret G 104 /// This means that F calls G, and F returns the value returned by G. 105 /// - Uses[arg F] = ret G 106 /// This means that some function calls G and passes its result as an 107 /// argument to F. 108 /// - Uses[ret F] = arg F 109 /// This means that F returns one of its own arguments. 110 /// - Uses[arg F] = arg G 111 /// This means that G calls F and passes one of its own (G's) arguments 112 /// directly to F. 113 UseMap Uses; 114 115 typedef std::set<RetOrArg> LiveSet; 116 typedef std::set<const Function*> LiveFuncSet; 117 118 /// This set contains all values that have been determined to be live. 119 LiveSet LiveValues; 120 /// This set contains all values that are cannot be changed in any way. 121 LiveFuncSet LiveFunctions; 122 123 typedef SmallVector<RetOrArg, 5> UseVector; 124 125 protected: 126 // DAH uses this to specify a different ID. 127 explicit DAE(char &ID) : ModulePass(ID) {} 128 129 public: 130 static char ID; // Pass identification, replacement for typeid 131 DAE() : ModulePass(ID) { 132 initializeDAEPass(*PassRegistry::getPassRegistry()); 133 } 134 135 bool runOnModule(Module &M) override; 136 137 virtual bool ShouldHackArguments() const { return false; } 138 139 private: 140 Liveness MarkIfNotLive(RetOrArg Use, UseVector &MaybeLiveUses); 141 Liveness SurveyUse(const Use *U, UseVector &MaybeLiveUses, 142 unsigned RetValNum = -1U); 143 Liveness SurveyUses(const Value *V, UseVector &MaybeLiveUses); 144 145 void SurveyFunction(const Function &F); 146 void MarkValue(const RetOrArg &RA, Liveness L, 147 const UseVector &MaybeLiveUses); 148 void MarkLive(const RetOrArg &RA); 149 void MarkLive(const Function &F); 150 void PropagateLiveness(const RetOrArg &RA); 151 bool RemoveDeadStuffFromFunction(Function *F); 152 bool DeleteDeadVarargs(Function &Fn); 153 bool RemoveDeadArgumentsFromCallers(Function &Fn); 154 }; 155 } 156 157 158 char DAE::ID = 0; 159 INITIALIZE_PASS(DAE, "deadargelim", "Dead Argument Elimination", false, false) 160 161 namespace { 162 /// DAH - DeadArgumentHacking pass - Same as dead argument elimination, but 163 /// deletes arguments to functions which are external. This is only for use 164 /// by bugpoint. 165 struct DAH : public DAE { 166 static char ID; 167 DAH() : DAE(ID) {} 168 169 bool ShouldHackArguments() const override { return true; } 170 }; 171 } 172 173 char DAH::ID = 0; 174 INITIALIZE_PASS(DAH, "deadarghaX0r", 175 "Dead Argument Hacking (BUGPOINT USE ONLY; DO NOT USE)", 176 false, false) 177 178 /// createDeadArgEliminationPass - This pass removes arguments from functions 179 /// which are not used by the body of the function. 180 /// 181 ModulePass *llvm::createDeadArgEliminationPass() { return new DAE(); } 182 ModulePass *llvm::createDeadArgHackingPass() { return new DAH(); } 183 184 /// DeleteDeadVarargs - If this is an function that takes a ... list, and if 185 /// llvm.vastart is never called, the varargs list is dead for the function. 186 bool DAE::DeleteDeadVarargs(Function &Fn) { 187 assert(Fn.getFunctionType()->isVarArg() && "Function isn't varargs!"); 188 if (Fn.isDeclaration() || !Fn.hasLocalLinkage()) return false; 189 190 // Ensure that the function is only directly called. 191 if (Fn.hasAddressTaken()) 192 return false; 193 194 // Don't touch naked functions. The assembly might be using an argument, or 195 // otherwise rely on the frame layout in a way that this analysis will not 196 // see. 197 if (Fn.hasFnAttribute(Attribute::Naked)) { 198 return false; 199 } 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.getIterator(), 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 NF->setSubprogram(Fn.getSubprogram()); 305 306 // Fix up any BlockAddresses that refer to the function. 307 Fn.replaceAllUsesWith(ConstantExpr::getBitCast(NF, Fn.getType())); 308 // Delete the bitcast that we just created, so that NF does not 309 // appear to be address-taken. 310 NF->removeDeadConstantUsers(); 311 // Finally, nuke the old function. 312 Fn.eraseFromParent(); 313 return true; 314 } 315 316 /// RemoveDeadArgumentsFromCallers - Checks if the given function has any 317 /// arguments that are unused, and changes the caller parameters to be undefined 318 /// instead. 319 bool DAE::RemoveDeadArgumentsFromCallers(Function &Fn) 320 { 321 // We cannot change the arguments if this TU does not define the function or 322 // if the linker may choose a function body from another TU, even if the 323 // nominal linkage indicates that other copies of the function have the same 324 // semantics. In the below example, the dead load from %p may not have been 325 // eliminated from the linker-chosen copy of f, so replacing %p with undef 326 // in callers may introduce undefined behavior. 327 // 328 // define linkonce_odr void @f(i32* %p) { 329 // %v = load i32 %p 330 // ret void 331 // } 332 if (!Fn.isStrongDefinitionForLinker()) 333 return false; 334 335 // Functions with local linkage should already have been handled, except the 336 // fragile (variadic) ones which we can improve here. 337 if (Fn.hasLocalLinkage() && !Fn.getFunctionType()->isVarArg()) 338 return false; 339 340 // Don't touch naked functions. The assembly might be using an argument, or 341 // otherwise rely on the frame layout in a way that this analysis will not 342 // see. 343 if (Fn.hasFnAttribute(Attribute::Naked)) 344 return false; 345 346 if (Fn.use_empty()) 347 return false; 348 349 SmallVector<unsigned, 8> UnusedArgs; 350 for (Argument &Arg : Fn.args()) { 351 if (Arg.use_empty() && !Arg.hasByValOrInAllocaAttr()) 352 UnusedArgs.push_back(Arg.getArgNo()); 353 } 354 355 if (UnusedArgs.empty()) 356 return false; 357 358 bool Changed = false; 359 360 for (Use &U : Fn.uses()) { 361 CallSite CS(U.getUser()); 362 if (!CS || !CS.isCallee(&U)) 363 continue; 364 365 // Now go through all unused args and replace them with "undef". 366 for (unsigned I = 0, E = UnusedArgs.size(); I != E; ++I) { 367 unsigned ArgNo = UnusedArgs[I]; 368 369 Value *Arg = CS.getArgument(ArgNo); 370 CS.setArgument(ArgNo, UndefValue::get(Arg->getType())); 371 ++NumArgumentsReplacedWithUndef; 372 Changed = true; 373 } 374 } 375 376 return Changed; 377 } 378 379 /// Convenience function that returns the number of return values. It returns 0 380 /// for void functions and 1 for functions not returning a struct. It returns 381 /// the number of struct elements for functions returning a struct. 382 static unsigned NumRetVals(const Function *F) { 383 Type *RetTy = F->getReturnType(); 384 if (RetTy->isVoidTy()) 385 return 0; 386 else if (StructType *STy = dyn_cast<StructType>(RetTy)) 387 return STy->getNumElements(); 388 else if (ArrayType *ATy = dyn_cast<ArrayType>(RetTy)) 389 return ATy->getNumElements(); 390 else 391 return 1; 392 } 393 394 /// Returns the sub-type a function will return at a given Idx. Should 395 /// correspond to the result type of an ExtractValue instruction executed with 396 /// just that one Idx (i.e. only top-level structure is considered). 397 static Type *getRetComponentType(const Function *F, unsigned Idx) { 398 Type *RetTy = F->getReturnType(); 399 assert(!RetTy->isVoidTy() && "void type has no subtype"); 400 401 if (StructType *STy = dyn_cast<StructType>(RetTy)) 402 return STy->getElementType(Idx); 403 else if (ArrayType *ATy = dyn_cast<ArrayType>(RetTy)) 404 return ATy->getElementType(); 405 else 406 return RetTy; 407 } 408 409 /// MarkIfNotLive - This checks Use for liveness in LiveValues. If Use is not 410 /// live, it adds Use to the MaybeLiveUses argument. Returns the determined 411 /// liveness of Use. 412 DAE::Liveness DAE::MarkIfNotLive(RetOrArg Use, UseVector &MaybeLiveUses) { 413 // We're live if our use or its Function is already marked as live. 414 if (LiveFunctions.count(Use.F) || LiveValues.count(Use)) 415 return Live; 416 417 // We're maybe live otherwise, but remember that we must become live if 418 // Use becomes live. 419 MaybeLiveUses.push_back(Use); 420 return MaybeLive; 421 } 422 423 424 /// SurveyUse - This looks at a single use of an argument or return value 425 /// and determines if it should be alive or not. Adds this use to MaybeLiveUses 426 /// if it causes the used value to become MaybeLive. 427 /// 428 /// RetValNum is the return value number to use when this use is used in a 429 /// return instruction. This is used in the recursion, you should always leave 430 /// it at 0. 431 DAE::Liveness DAE::SurveyUse(const Use *U, 432 UseVector &MaybeLiveUses, unsigned RetValNum) { 433 const User *V = U->getUser(); 434 if (const ReturnInst *RI = dyn_cast<ReturnInst>(V)) { 435 // The value is returned from a function. It's only live when the 436 // function's return value is live. We use RetValNum here, for the case 437 // that U is really a use of an insertvalue instruction that uses the 438 // original Use. 439 const Function *F = RI->getParent()->getParent(); 440 if (RetValNum != -1U) { 441 RetOrArg Use = CreateRet(F, RetValNum); 442 // We might be live, depending on the liveness of Use. 443 return MarkIfNotLive(Use, MaybeLiveUses); 444 } else { 445 DAE::Liveness Result = MaybeLive; 446 for (unsigned i = 0; i < NumRetVals(F); ++i) { 447 RetOrArg Use = CreateRet(F, i); 448 // We might be live, depending on the liveness of Use. If any 449 // sub-value is live, then the entire value is considered live. This 450 // is a conservative choice, and better tracking is possible. 451 DAE::Liveness SubResult = MarkIfNotLive(Use, MaybeLiveUses); 452 if (Result != Live) 453 Result = SubResult; 454 } 455 return Result; 456 } 457 } 458 if (const InsertValueInst *IV = dyn_cast<InsertValueInst>(V)) { 459 if (U->getOperandNo() != InsertValueInst::getAggregateOperandIndex() 460 && IV->hasIndices()) 461 // The use we are examining is inserted into an aggregate. Our liveness 462 // depends on all uses of that aggregate, but if it is used as a return 463 // value, only index at which we were inserted counts. 464 RetValNum = *IV->idx_begin(); 465 466 // Note that if we are used as the aggregate operand to the insertvalue, 467 // we don't change RetValNum, but do survey all our uses. 468 469 Liveness Result = MaybeLive; 470 for (const Use &UU : IV->uses()) { 471 Result = SurveyUse(&UU, MaybeLiveUses, RetValNum); 472 if (Result == Live) 473 break; 474 } 475 return Result; 476 } 477 478 if (auto CS = ImmutableCallSite(V)) { 479 const Function *F = CS.getCalledFunction(); 480 if (F) { 481 // Used in a direct call. 482 483 // Find the argument number. We know for sure that this use is an 484 // argument, since if it was the function argument this would be an 485 // indirect call and the we know can't be looking at a value of the 486 // label type (for the invoke instruction). 487 unsigned ArgNo = CS.getArgumentNo(U); 488 489 if (ArgNo >= F->getFunctionType()->getNumParams()) 490 // The value is passed in through a vararg! Must be live. 491 return Live; 492 493 assert(CS.getArgument(ArgNo) 494 == CS->getOperand(U->getOperandNo()) 495 && "Argument is not where we expected it"); 496 497 // Value passed to a normal call. It's only live when the corresponding 498 // argument to the called function turns out live. 499 RetOrArg Use = CreateArg(F, ArgNo); 500 return MarkIfNotLive(Use, MaybeLiveUses); 501 } 502 } 503 // Used in any other way? Value must be live. 504 return Live; 505 } 506 507 /// SurveyUses - This looks at all the uses of the given value 508 /// Returns the Liveness deduced from the uses of this value. 509 /// 510 /// Adds all uses that cause the result to be MaybeLive to MaybeLiveRetUses. If 511 /// the result is Live, MaybeLiveUses might be modified but its content should 512 /// be ignored (since it might not be complete). 513 DAE::Liveness DAE::SurveyUses(const Value *V, UseVector &MaybeLiveUses) { 514 // Assume it's dead (which will only hold if there are no uses at all..). 515 Liveness Result = MaybeLive; 516 // Check each use. 517 for (const Use &U : V->uses()) { 518 Result = SurveyUse(&U, MaybeLiveUses); 519 if (Result == Live) 520 break; 521 } 522 return Result; 523 } 524 525 // SurveyFunction - This performs the initial survey of the specified function, 526 // checking out whether or not it uses any of its incoming arguments or whether 527 // any callers use the return value. This fills in the LiveValues set and Uses 528 // map. 529 // 530 // We consider arguments of non-internal functions to be intrinsically alive as 531 // well as arguments to functions which have their "address taken". 532 // 533 void DAE::SurveyFunction(const Function &F) { 534 // Functions with inalloca parameters are expecting args in a particular 535 // register and memory layout. 536 if (F.getAttributes().hasAttrSomewhere(Attribute::InAlloca)) { 537 MarkLive(F); 538 return; 539 } 540 541 // Don't touch naked functions. The assembly might be using an argument, or 542 // otherwise rely on the frame layout in a way that this analysis will not 543 // see. 544 if (F.hasFnAttribute(Attribute::Naked)) { 545 MarkLive(F); 546 return; 547 } 548 549 unsigned RetCount = NumRetVals(&F); 550 // Assume all return values are dead 551 typedef SmallVector<Liveness, 5> RetVals; 552 RetVals RetValLiveness(RetCount, MaybeLive); 553 554 typedef SmallVector<UseVector, 5> RetUses; 555 // These vectors map each return value to the uses that make it MaybeLive, so 556 // we can add those to the Uses map if the return value really turns out to be 557 // MaybeLive. Initialized to a list of RetCount empty lists. 558 RetUses MaybeLiveRetUses(RetCount); 559 560 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 561 if (const ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) 562 if (RI->getNumOperands() != 0 && RI->getOperand(0)->getType() 563 != F.getFunctionType()->getReturnType()) { 564 // We don't support old style multiple return values. 565 MarkLive(F); 566 return; 567 } 568 569 if (!F.hasLocalLinkage() && (!ShouldHackArguments() || F.isIntrinsic())) { 570 MarkLive(F); 571 return; 572 } 573 574 DEBUG(dbgs() << "DAE - Inspecting callers for fn: " << F.getName() << "\n"); 575 // Keep track of the number of live retvals, so we can skip checks once all 576 // of them turn out to be live. 577 unsigned NumLiveRetVals = 0; 578 // Loop all uses of the function. 579 for (const Use &U : F.uses()) { 580 // If the function is PASSED IN as an argument, its address has been 581 // taken. 582 ImmutableCallSite CS(U.getUser()); 583 if (!CS || !CS.isCallee(&U)) { 584 MarkLive(F); 585 return; 586 } 587 588 // If this use is anything other than a call site, the function is alive. 589 const Instruction *TheCall = CS.getInstruction(); 590 if (!TheCall) { // Not a direct call site? 591 MarkLive(F); 592 return; 593 } 594 595 // If we end up here, we are looking at a direct call to our function. 596 597 // Now, check how our return value(s) is/are used in this caller. Don't 598 // bother checking return values if all of them are live already. 599 if (NumLiveRetVals == RetCount) 600 continue; 601 602 // Check all uses of the return value. 603 for (const Use &U : TheCall->uses()) { 604 if (ExtractValueInst *Ext = dyn_cast<ExtractValueInst>(U.getUser())) { 605 // This use uses a part of our return value, survey the uses of 606 // that part and store the results for this index only. 607 unsigned Idx = *Ext->idx_begin(); 608 if (RetValLiveness[Idx] != Live) { 609 RetValLiveness[Idx] = SurveyUses(Ext, MaybeLiveRetUses[Idx]); 610 if (RetValLiveness[Idx] == Live) 611 NumLiveRetVals++; 612 } 613 } else { 614 // Used by something else than extractvalue. Survey, but assume that the 615 // result applies to all sub-values. 616 UseVector MaybeLiveAggregateUses; 617 if (SurveyUse(&U, MaybeLiveAggregateUses) == Live) { 618 NumLiveRetVals = RetCount; 619 RetValLiveness.assign(RetCount, Live); 620 break; 621 } else { 622 for (unsigned i = 0; i != RetCount; ++i) { 623 if (RetValLiveness[i] != Live) 624 MaybeLiveRetUses[i].append(MaybeLiveAggregateUses.begin(), 625 MaybeLiveAggregateUses.end()); 626 } 627 } 628 } 629 } 630 } 631 632 // Now we've inspected all callers, record the liveness of our return values. 633 for (unsigned i = 0; i != RetCount; ++i) 634 MarkValue(CreateRet(&F, i), RetValLiveness[i], MaybeLiveRetUses[i]); 635 636 DEBUG(dbgs() << "DAE - Inspecting args for fn: " << F.getName() << "\n"); 637 638 // Now, check all of our arguments. 639 unsigned i = 0; 640 UseVector MaybeLiveArgUses; 641 for (Function::const_arg_iterator AI = F.arg_begin(), 642 E = F.arg_end(); AI != E; ++AI, ++i) { 643 Liveness Result; 644 if (F.getFunctionType()->isVarArg()) { 645 // Variadic functions will already have a va_arg function expanded inside 646 // them, making them potentially very sensitive to ABI changes resulting 647 // from removing arguments entirely, so don't. For example AArch64 handles 648 // register and stack HFAs very differently, and this is reflected in the 649 // IR which has already been generated. 650 Result = Live; 651 } else { 652 // See what the effect of this use is (recording any uses that cause 653 // MaybeLive in MaybeLiveArgUses). 654 Result = SurveyUses(&*AI, MaybeLiveArgUses); 655 } 656 657 // Mark the result. 658 MarkValue(CreateArg(&F, i), Result, MaybeLiveArgUses); 659 // Clear the vector again for the next iteration. 660 MaybeLiveArgUses.clear(); 661 } 662 } 663 664 /// MarkValue - This function marks the liveness of RA depending on L. If L is 665 /// MaybeLive, it also takes all uses in MaybeLiveUses and records them in Uses, 666 /// such that RA will be marked live if any use in MaybeLiveUses gets marked 667 /// live later on. 668 void DAE::MarkValue(const RetOrArg &RA, Liveness L, 669 const UseVector &MaybeLiveUses) { 670 switch (L) { 671 case Live: MarkLive(RA); break; 672 case MaybeLive: 673 { 674 // Note any uses of this value, so this return value can be 675 // marked live whenever one of the uses becomes live. 676 for (UseVector::const_iterator UI = MaybeLiveUses.begin(), 677 UE = MaybeLiveUses.end(); UI != UE; ++UI) 678 Uses.insert(std::make_pair(*UI, RA)); 679 break; 680 } 681 } 682 } 683 684 /// MarkLive - Mark the given Function as alive, meaning that it cannot be 685 /// changed in any way. Additionally, 686 /// mark any values that are used as this function's parameters or by its return 687 /// values (according to Uses) live as well. 688 void DAE::MarkLive(const Function &F) { 689 DEBUG(dbgs() << "DAE - Intrinsically live fn: " << F.getName() << "\n"); 690 // Mark the function as live. 691 LiveFunctions.insert(&F); 692 // Mark all arguments as live. 693 for (unsigned i = 0, e = F.arg_size(); i != e; ++i) 694 PropagateLiveness(CreateArg(&F, i)); 695 // Mark all return values as live. 696 for (unsigned i = 0, e = NumRetVals(&F); i != e; ++i) 697 PropagateLiveness(CreateRet(&F, i)); 698 } 699 700 /// MarkLive - Mark the given return value or argument as live. Additionally, 701 /// mark any values that are used by this value (according to Uses) live as 702 /// well. 703 void DAE::MarkLive(const RetOrArg &RA) { 704 if (LiveFunctions.count(RA.F)) 705 return; // Function was already marked Live. 706 707 if (!LiveValues.insert(RA).second) 708 return; // We were already marked Live. 709 710 DEBUG(dbgs() << "DAE - Marking " << RA.getDescription() << " live\n"); 711 PropagateLiveness(RA); 712 } 713 714 /// PropagateLiveness - Given that RA is a live value, propagate it's liveness 715 /// to any other values it uses (according to Uses). 716 void DAE::PropagateLiveness(const RetOrArg &RA) { 717 // We don't use upper_bound (or equal_range) here, because our recursive call 718 // to ourselves is likely to cause the upper_bound (which is the first value 719 // not belonging to RA) to become erased and the iterator invalidated. 720 UseMap::iterator Begin = Uses.lower_bound(RA); 721 UseMap::iterator E = Uses.end(); 722 UseMap::iterator I; 723 for (I = Begin; I != E && I->first == RA; ++I) 724 MarkLive(I->second); 725 726 // Erase RA from the Uses map (from the lower bound to wherever we ended up 727 // after the loop). 728 Uses.erase(Begin, I); 729 } 730 731 // RemoveDeadStuffFromFunction - Remove any arguments and return values from F 732 // that are not in LiveValues. Transform the function and all of the callees of 733 // the function to not have these arguments and return values. 734 // 735 bool DAE::RemoveDeadStuffFromFunction(Function *F) { 736 // Don't modify fully live functions 737 if (LiveFunctions.count(F)) 738 return false; 739 740 // Start by computing a new prototype for the function, which is the same as 741 // the old function, but has fewer arguments and a different return type. 742 FunctionType *FTy = F->getFunctionType(); 743 std::vector<Type*> Params; 744 745 // Keep track of if we have a live 'returned' argument 746 bool HasLiveReturnedArg = false; 747 748 // Set up to build a new list of parameter attributes. 749 SmallVector<AttributeSet, 8> AttributesVec; 750 const AttributeSet &PAL = F->getAttributes(); 751 752 // Remember which arguments are still alive. 753 SmallVector<bool, 10> ArgAlive(FTy->getNumParams(), false); 754 // Construct the new parameter list from non-dead arguments. Also construct 755 // a new set of parameter attributes to correspond. Skip the first parameter 756 // attribute, since that belongs to the return value. 757 unsigned i = 0; 758 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 759 I != E; ++I, ++i) { 760 RetOrArg Arg = CreateArg(F, i); 761 if (LiveValues.erase(Arg)) { 762 Params.push_back(I->getType()); 763 ArgAlive[i] = true; 764 765 // Get the original parameter attributes (skipping the first one, that is 766 // for the return value. 767 if (PAL.hasAttributes(i + 1)) { 768 AttrBuilder B(PAL, i + 1); 769 if (B.contains(Attribute::Returned)) 770 HasLiveReturnedArg = true; 771 AttributesVec. 772 push_back(AttributeSet::get(F->getContext(), Params.size(), B)); 773 } 774 } else { 775 ++NumArgumentsEliminated; 776 DEBUG(dbgs() << "DAE - Removing argument " << i << " (" << I->getName() 777 << ") from " << F->getName() << "\n"); 778 } 779 } 780 781 // Find out the new return value. 782 Type *RetTy = FTy->getReturnType(); 783 Type *NRetTy = nullptr; 784 unsigned RetCount = NumRetVals(F); 785 786 // -1 means unused, other numbers are the new index 787 SmallVector<int, 5> NewRetIdxs(RetCount, -1); 788 std::vector<Type*> RetTypes; 789 790 // If there is a function with a live 'returned' argument but a dead return 791 // value, then there are two possible actions: 792 // 1) Eliminate the return value and take off the 'returned' attribute on the 793 // argument. 794 // 2) Retain the 'returned' attribute and treat the return value (but not the 795 // entire function) as live so that it is not eliminated. 796 // 797 // It's not clear in the general case which option is more profitable because, 798 // even in the absence of explicit uses of the return value, code generation 799 // is free to use the 'returned' attribute to do things like eliding 800 // save/restores of registers across calls. Whether or not this happens is 801 // target and ABI-specific as well as depending on the amount of register 802 // pressure, so there's no good way for an IR-level pass to figure this out. 803 // 804 // Fortunately, the only places where 'returned' is currently generated by 805 // the FE are places where 'returned' is basically free and almost always a 806 // performance win, so the second option can just be used always for now. 807 // 808 // This should be revisited if 'returned' is ever applied more liberally. 809 if (RetTy->isVoidTy() || HasLiveReturnedArg) { 810 NRetTy = RetTy; 811 } else { 812 // Look at each of the original return values individually. 813 for (unsigned i = 0; i != RetCount; ++i) { 814 RetOrArg Ret = CreateRet(F, i); 815 if (LiveValues.erase(Ret)) { 816 RetTypes.push_back(getRetComponentType(F, i)); 817 NewRetIdxs[i] = RetTypes.size() - 1; 818 } else { 819 ++NumRetValsEliminated; 820 DEBUG(dbgs() << "DAE - Removing return value " << i << " from " 821 << F->getName() << "\n"); 822 } 823 } 824 if (RetTypes.size() > 1) { 825 // More than one return type? Reduce it down to size. 826 if (StructType *STy = dyn_cast<StructType>(RetTy)) { 827 // Make the new struct packed if we used to return a packed struct 828 // already. 829 NRetTy = StructType::get(STy->getContext(), RetTypes, STy->isPacked()); 830 } else { 831 assert(isa<ArrayType>(RetTy) && "unexpected multi-value return"); 832 NRetTy = ArrayType::get(RetTypes[0], RetTypes.size()); 833 } 834 } else if (RetTypes.size() == 1) 835 // One return type? Just a simple value then, but only if we didn't use to 836 // return a struct with that simple value before. 837 NRetTy = RetTypes.front(); 838 else if (RetTypes.size() == 0) 839 // No return types? Make it void, but only if we didn't use to return {}. 840 NRetTy = Type::getVoidTy(F->getContext()); 841 } 842 843 assert(NRetTy && "No new return type found?"); 844 845 // The existing function return attributes. 846 AttributeSet RAttrs = PAL.getRetAttributes(); 847 848 // Remove any incompatible attributes, but only if we removed all return 849 // values. Otherwise, ensure that we don't have any conflicting attributes 850 // here. Currently, this should not be possible, but special handling might be 851 // required when new return value attributes are added. 852 if (NRetTy->isVoidTy()) 853 RAttrs = RAttrs.removeAttributes(NRetTy->getContext(), 854 AttributeSet::ReturnIndex, 855 AttributeFuncs::typeIncompatible(NRetTy)); 856 else 857 assert(!AttrBuilder(RAttrs, AttributeSet::ReturnIndex). 858 overlaps(AttributeFuncs::typeIncompatible(NRetTy)) && 859 "Return attributes no longer compatible?"); 860 861 if (RAttrs.hasAttributes(AttributeSet::ReturnIndex)) 862 AttributesVec.push_back(AttributeSet::get(NRetTy->getContext(), RAttrs)); 863 864 if (PAL.hasAttributes(AttributeSet::FunctionIndex)) 865 AttributesVec.push_back(AttributeSet::get(F->getContext(), 866 PAL.getFnAttributes())); 867 868 // Reconstruct the AttributesList based on the vector we constructed. 869 AttributeSet NewPAL = AttributeSet::get(F->getContext(), AttributesVec); 870 871 // Create the new function type based on the recomputed parameters. 872 FunctionType *NFTy = FunctionType::get(NRetTy, Params, FTy->isVarArg()); 873 874 // No change? 875 if (NFTy == FTy) 876 return false; 877 878 // Create the new function body and insert it into the module... 879 Function *NF = Function::Create(NFTy, F->getLinkage()); 880 NF->copyAttributesFrom(F); 881 NF->setAttributes(NewPAL); 882 // Insert the new function before the old function, so we won't be processing 883 // it again. 884 F->getParent()->getFunctionList().insert(F->getIterator(), NF); 885 NF->takeName(F); 886 887 // Loop over all of the callers of the function, transforming the call sites 888 // to pass in a smaller number of arguments into the new function. 889 // 890 std::vector<Value*> Args; 891 while (!F->use_empty()) { 892 CallSite CS(F->user_back()); 893 Instruction *Call = CS.getInstruction(); 894 895 AttributesVec.clear(); 896 const AttributeSet &CallPAL = CS.getAttributes(); 897 898 // The call return attributes. 899 AttributeSet RAttrs = CallPAL.getRetAttributes(); 900 901 // Adjust in case the function was changed to return void. 902 RAttrs = RAttrs.removeAttributes(NRetTy->getContext(), 903 AttributeSet::ReturnIndex, 904 AttributeFuncs::typeIncompatible(NF->getReturnType())); 905 if (RAttrs.hasAttributes(AttributeSet::ReturnIndex)) 906 AttributesVec.push_back(AttributeSet::get(NF->getContext(), RAttrs)); 907 908 // Declare these outside of the loops, so we can reuse them for the second 909 // loop, which loops the varargs. 910 CallSite::arg_iterator I = CS.arg_begin(); 911 unsigned i = 0; 912 // Loop over those operands, corresponding to the normal arguments to the 913 // original function, and add those that are still alive. 914 for (unsigned e = FTy->getNumParams(); i != e; ++I, ++i) 915 if (ArgAlive[i]) { 916 Args.push_back(*I); 917 // Get original parameter attributes, but skip return attributes. 918 if (CallPAL.hasAttributes(i + 1)) { 919 AttrBuilder B(CallPAL, i + 1); 920 // If the return type has changed, then get rid of 'returned' on the 921 // call site. The alternative is to make all 'returned' attributes on 922 // call sites keep the return value alive just like 'returned' 923 // attributes on function declaration but it's less clearly a win 924 // and this is not an expected case anyway 925 if (NRetTy != RetTy && B.contains(Attribute::Returned)) 926 B.removeAttribute(Attribute::Returned); 927 AttributesVec. 928 push_back(AttributeSet::get(F->getContext(), Args.size(), B)); 929 } 930 } 931 932 // Push any varargs arguments on the list. Don't forget their attributes. 933 for (CallSite::arg_iterator E = CS.arg_end(); I != E; ++I, ++i) { 934 Args.push_back(*I); 935 if (CallPAL.hasAttributes(i + 1)) { 936 AttrBuilder B(CallPAL, i + 1); 937 AttributesVec. 938 push_back(AttributeSet::get(F->getContext(), Args.size(), B)); 939 } 940 } 941 942 if (CallPAL.hasAttributes(AttributeSet::FunctionIndex)) 943 AttributesVec.push_back(AttributeSet::get(Call->getContext(), 944 CallPAL.getFnAttributes())); 945 946 // Reconstruct the AttributesList based on the vector we constructed. 947 AttributeSet NewCallPAL = AttributeSet::get(F->getContext(), AttributesVec); 948 949 Instruction *New; 950 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 951 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 952 Args, "", Call->getParent()); 953 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 954 cast<InvokeInst>(New)->setAttributes(NewCallPAL); 955 } else { 956 New = CallInst::Create(NF, Args, "", Call); 957 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 958 cast<CallInst>(New)->setAttributes(NewCallPAL); 959 if (cast<CallInst>(Call)->isTailCall()) 960 cast<CallInst>(New)->setTailCall(); 961 } 962 New->setDebugLoc(Call->getDebugLoc()); 963 964 Args.clear(); 965 966 if (!Call->use_empty()) { 967 if (New->getType() == Call->getType()) { 968 // Return type not changed? Just replace users then. 969 Call->replaceAllUsesWith(New); 970 New->takeName(Call); 971 } else if (New->getType()->isVoidTy()) { 972 // Our return value has uses, but they will get removed later on. 973 // Replace by null for now. 974 if (!Call->getType()->isX86_MMXTy()) 975 Call->replaceAllUsesWith(Constant::getNullValue(Call->getType())); 976 } else { 977 assert((RetTy->isStructTy() || RetTy->isArrayTy()) && 978 "Return type changed, but not into a void. The old return type" 979 " must have been a struct or an array!"); 980 Instruction *InsertPt = Call; 981 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 982 BasicBlock *NewEdge = SplitEdge(New->getParent(), II->getNormalDest()); 983 InsertPt = &*NewEdge->getFirstInsertionPt(); 984 } 985 986 // We used to return a struct or array. Instead of doing smart stuff 987 // with all the uses, we will just rebuild it using extract/insertvalue 988 // chaining and let instcombine clean that up. 989 // 990 // Start out building up our return value from undef 991 Value *RetVal = UndefValue::get(RetTy); 992 for (unsigned i = 0; i != RetCount; ++i) 993 if (NewRetIdxs[i] != -1) { 994 Value *V; 995 if (RetTypes.size() > 1) 996 // We are still returning a struct, so extract the value from our 997 // return value 998 V = ExtractValueInst::Create(New, NewRetIdxs[i], "newret", 999 InsertPt); 1000 else 1001 // We are now returning a single element, so just insert that 1002 V = New; 1003 // Insert the value at the old position 1004 RetVal = InsertValueInst::Create(RetVal, V, i, "oldret", InsertPt); 1005 } 1006 // Now, replace all uses of the old call instruction with the return 1007 // struct we built 1008 Call->replaceAllUsesWith(RetVal); 1009 New->takeName(Call); 1010 } 1011 } 1012 1013 // Finally, remove the old call from the program, reducing the use-count of 1014 // F. 1015 Call->eraseFromParent(); 1016 } 1017 1018 // Since we have now created the new function, splice the body of the old 1019 // function right into the new function, leaving the old rotting hulk of the 1020 // function empty. 1021 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 1022 1023 // Loop over the argument list, transferring uses of the old arguments over to 1024 // the new arguments, also transferring over the names as well. 1025 i = 0; 1026 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), 1027 I2 = NF->arg_begin(); I != E; ++I, ++i) 1028 if (ArgAlive[i]) { 1029 // If this is a live argument, move the name and users over to the new 1030 // version. 1031 I->replaceAllUsesWith(&*I2); 1032 I2->takeName(&*I); 1033 ++I2; 1034 } else { 1035 // If this argument is dead, replace any uses of it with null constants 1036 // (these are guaranteed to become unused later on). 1037 if (!I->getType()->isX86_MMXTy()) 1038 I->replaceAllUsesWith(Constant::getNullValue(I->getType())); 1039 } 1040 1041 // If we change the return value of the function we must rewrite any return 1042 // instructions. Check this now. 1043 if (F->getReturnType() != NF->getReturnType()) 1044 for (Function::iterator BB = NF->begin(), E = NF->end(); BB != E; ++BB) 1045 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 1046 Value *RetVal; 1047 1048 if (NFTy->getReturnType()->isVoidTy()) { 1049 RetVal = nullptr; 1050 } else { 1051 assert(RetTy->isStructTy() || RetTy->isArrayTy()); 1052 // The original return value was a struct or array, insert 1053 // extractvalue/insertvalue chains to extract only the values we need 1054 // to return and insert them into our new result. 1055 // This does generate messy code, but we'll let it to instcombine to 1056 // clean that up. 1057 Value *OldRet = RI->getOperand(0); 1058 // Start out building up our return value from undef 1059 RetVal = UndefValue::get(NRetTy); 1060 for (unsigned i = 0; i != RetCount; ++i) 1061 if (NewRetIdxs[i] != -1) { 1062 ExtractValueInst *EV = ExtractValueInst::Create(OldRet, i, 1063 "oldret", RI); 1064 if (RetTypes.size() > 1) { 1065 // We're still returning a struct, so reinsert the value into 1066 // our new return value at the new index 1067 1068 RetVal = InsertValueInst::Create(RetVal, EV, NewRetIdxs[i], 1069 "newret", RI); 1070 } else { 1071 // We are now only returning a simple value, so just return the 1072 // extracted value. 1073 RetVal = EV; 1074 } 1075 } 1076 } 1077 // Replace the return instruction with one returning the new return 1078 // value (possibly 0 if we became void). 1079 ReturnInst::Create(F->getContext(), RetVal, RI); 1080 BB->getInstList().erase(RI); 1081 } 1082 1083 // Patch the pointer to LLVM function in debug info descriptor. 1084 NF->setSubprogram(F->getSubprogram()); 1085 1086 // Now that the old function is dead, delete it. 1087 F->eraseFromParent(); 1088 1089 return true; 1090 } 1091 1092 bool DAE::runOnModule(Module &M) { 1093 bool Changed = false; 1094 1095 // First pass: Do a simple check to see if any functions can have their "..." 1096 // removed. We can do this if they never call va_start. This loop cannot be 1097 // fused with the next loop, because deleting a function invalidates 1098 // information computed while surveying other functions. 1099 DEBUG(dbgs() << "DAE - Deleting dead varargs\n"); 1100 for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { 1101 Function &F = *I++; 1102 if (F.getFunctionType()->isVarArg()) 1103 Changed |= DeleteDeadVarargs(F); 1104 } 1105 1106 // Second phase:loop through the module, determining which arguments are live. 1107 // We assume all arguments are dead unless proven otherwise (allowing us to 1108 // determine that dead arguments passed into recursive functions are dead). 1109 // 1110 DEBUG(dbgs() << "DAE - Determining liveness\n"); 1111 for (auto &F : M) 1112 SurveyFunction(F); 1113 1114 // Now, remove all dead arguments and return values from each function in 1115 // turn. 1116 for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { 1117 // Increment now, because the function will probably get removed (ie. 1118 // replaced by a new one). 1119 Function *F = &*I++; 1120 Changed |= RemoveDeadStuffFromFunction(F); 1121 } 1122 1123 // Finally, look for any unused parameters in functions with non-local 1124 // linkage and replace the passed in parameters with undef. 1125 for (auto &F : M) 1126 Changed |= RemoveDeadArgumentsFromCallers(F); 1127 1128 return Changed; 1129 } 1130