1 //===-- ArgumentPromotion.cpp - Promote by-reference 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 promotes "by reference" arguments to be "by value" arguments. In 11 // practice, this means looking for internal functions that have pointer 12 // arguments. If it can prove, through the use of alias analysis, that an 13 // argument is *only* loaded, then it can pass the value into the function 14 // instead of the address of the value. This can cause recursive simplification 15 // of code and lead to the elimination of allocas (especially in C++ template 16 // code like the STL). 17 // 18 // This pass also handles aggregate arguments that are passed into a function, 19 // scalarizing them if the elements of the aggregate are only loaded. Note that 20 // by default it refuses to scalarize aggregates which would require passing in 21 // more than three operands to the function, because passing thousands of 22 // operands for a large array or structure is unprofitable! This limit can be 23 // configured or disabled, however. 24 // 25 // Note that this transformation could also be done for arguments that are only 26 // stored to (returning the value instead), but does not currently. This case 27 // would be best handled when and if LLVM begins supporting multiple return 28 // values from functions. 29 // 30 //===----------------------------------------------------------------------===// 31 32 #define DEBUG_TYPE "argpromotion" 33 #include "llvm/Transforms/IPO.h" 34 #include "llvm/Constants.h" 35 #include "llvm/DerivedTypes.h" 36 #include "llvm/Module.h" 37 #include "llvm/CallGraphSCCPass.h" 38 #include "llvm/Instructions.h" 39 #include "llvm/LLVMContext.h" 40 #include "llvm/Analysis/AliasAnalysis.h" 41 #include "llvm/Analysis/CallGraph.h" 42 #include "llvm/Support/CallSite.h" 43 #include "llvm/Support/CFG.h" 44 #include "llvm/Support/Debug.h" 45 #include "llvm/Support/raw_ostream.h" 46 #include "llvm/ADT/DepthFirstIterator.h" 47 #include "llvm/ADT/Statistic.h" 48 #include "llvm/ADT/StringExtras.h" 49 #include <set> 50 using namespace llvm; 51 52 STATISTIC(NumArgumentsPromoted , "Number of pointer arguments promoted"); 53 STATISTIC(NumAggregatesPromoted, "Number of aggregate arguments promoted"); 54 STATISTIC(NumByValArgsPromoted , "Number of byval arguments promoted"); 55 STATISTIC(NumArgumentsDead , "Number of dead pointer args eliminated"); 56 57 namespace { 58 /// ArgPromotion - The 'by reference' to 'by value' argument promotion pass. 59 /// 60 struct ArgPromotion : public CallGraphSCCPass { 61 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 62 AU.addRequired<AliasAnalysis>(); 63 CallGraphSCCPass::getAnalysisUsage(AU); 64 } 65 66 virtual bool runOnSCC(CallGraphSCC &SCC); 67 static char ID; // Pass identification, replacement for typeid 68 explicit ArgPromotion(unsigned maxElements = 3) 69 : CallGraphSCCPass(ID), maxElements(maxElements) { 70 initializeArgPromotionPass(*PassRegistry::getPassRegistry()); 71 } 72 73 /// A vector used to hold the indices of a single GEP instruction 74 typedef std::vector<uint64_t> IndicesVector; 75 76 private: 77 CallGraphNode *PromoteArguments(CallGraphNode *CGN); 78 bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const; 79 CallGraphNode *DoPromotion(Function *F, 80 SmallPtrSet<Argument*, 8> &ArgsToPromote, 81 SmallPtrSet<Argument*, 8> &ByValArgsToTransform); 82 /// The maximum number of elements to expand, or 0 for unlimited. 83 unsigned maxElements; 84 }; 85 } 86 87 char ArgPromotion::ID = 0; 88 INITIALIZE_PASS_BEGIN(ArgPromotion, "argpromotion", 89 "Promote 'by reference' arguments to scalars", false, false) 90 INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 91 INITIALIZE_AG_DEPENDENCY(CallGraph) 92 INITIALIZE_PASS_END(ArgPromotion, "argpromotion", 93 "Promote 'by reference' arguments to scalars", false, false) 94 95 Pass *llvm::createArgumentPromotionPass(unsigned maxElements) { 96 return new ArgPromotion(maxElements); 97 } 98 99 bool ArgPromotion::runOnSCC(CallGraphSCC &SCC) { 100 bool Changed = false, LocalChange; 101 102 do { // Iterate until we stop promoting from this SCC. 103 LocalChange = false; 104 // Attempt to promote arguments from all functions in this SCC. 105 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) { 106 if (CallGraphNode *CGN = PromoteArguments(*I)) { 107 LocalChange = true; 108 SCC.ReplaceNode(*I, CGN); 109 } 110 } 111 Changed |= LocalChange; // Remember that we changed something. 112 } while (LocalChange); 113 114 return Changed; 115 } 116 117 /// PromoteArguments - This method checks the specified function to see if there 118 /// are any promotable arguments and if it is safe to promote the function (for 119 /// example, all callers are direct). If safe to promote some arguments, it 120 /// calls the DoPromotion method. 121 /// 122 CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) { 123 Function *F = CGN->getFunction(); 124 125 // Make sure that it is local to this module. 126 if (!F || !F->hasLocalLinkage()) return 0; 127 128 // First check: see if there are any pointer arguments! If not, quick exit. 129 SmallVector<std::pair<Argument*, unsigned>, 16> PointerArgs; 130 unsigned ArgNo = 0; 131 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 132 I != E; ++I, ++ArgNo) 133 if (I->getType()->isPointerTy()) 134 PointerArgs.push_back(std::pair<Argument*, unsigned>(I, ArgNo)); 135 if (PointerArgs.empty()) return 0; 136 137 // Second check: make sure that all callers are direct callers. We can't 138 // transform functions that have indirect callers. Also see if the function 139 // is self-recursive. 140 bool isSelfRecursive = false; 141 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); 142 UI != E; ++UI) { 143 CallSite CS(*UI); 144 // Must be a direct call. 145 if (CS.getInstruction() == 0 || !CS.isCallee(UI)) return 0; 146 147 if (CS.getInstruction()->getParent()->getParent() == F) 148 isSelfRecursive = true; 149 } 150 151 // Check to see which arguments are promotable. If an argument is promotable, 152 // add it to ArgsToPromote. 153 SmallPtrSet<Argument*, 8> ArgsToPromote; 154 SmallPtrSet<Argument*, 8> ByValArgsToTransform; 155 for (unsigned i = 0; i != PointerArgs.size(); ++i) { 156 bool isByVal = F->paramHasAttr(PointerArgs[i].second+1, Attribute::ByVal); 157 Argument *PtrArg = PointerArgs[i].first; 158 Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType(); 159 160 // If this is a byval argument, and if the aggregate type is small, just 161 // pass the elements, which is always safe. 162 if (isByVal) { 163 if (StructType *STy = dyn_cast<StructType>(AgTy)) { 164 if (maxElements > 0 && STy->getNumElements() > maxElements) { 165 DEBUG(dbgs() << "argpromotion disable promoting argument '" 166 << PtrArg->getName() << "' because it would require adding more" 167 << " than " << maxElements << " arguments to the function.\n"); 168 continue; 169 } 170 171 // If all the elements are single-value types, we can promote it. 172 bool AllSimple = true; 173 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 174 if (!STy->getElementType(i)->isSingleValueType()) { 175 AllSimple = false; 176 break; 177 } 178 } 179 180 // Safe to transform, don't even bother trying to "promote" it. 181 // Passing the elements as a scalar will allow scalarrepl to hack on 182 // the new alloca we introduce. 183 if (AllSimple) { 184 ByValArgsToTransform.insert(PtrArg); 185 continue; 186 } 187 } 188 } 189 190 // If the argument is a recursive type and we're in a recursive 191 // function, we could end up infinitely peeling the function argument. 192 if (isSelfRecursive) { 193 if (StructType *STy = dyn_cast<StructType>(AgTy)) { 194 bool RecursiveType = false; 195 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 196 if (STy->getElementType(i) == PtrArg->getType()) { 197 RecursiveType = true; 198 break; 199 } 200 } 201 if (RecursiveType) 202 continue; 203 } 204 } 205 206 // Otherwise, see if we can promote the pointer to its value. 207 if (isSafeToPromoteArgument(PtrArg, isByVal)) 208 ArgsToPromote.insert(PtrArg); 209 } 210 211 // No promotable pointer arguments. 212 if (ArgsToPromote.empty() && ByValArgsToTransform.empty()) 213 return 0; 214 215 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform); 216 } 217 218 /// AllCallersPassInValidPointerForArgument - Return true if we can prove that 219 /// all callees pass in a valid pointer for the specified function argument. 220 static bool AllCallersPassInValidPointerForArgument(Argument *Arg) { 221 Function *Callee = Arg->getParent(); 222 223 unsigned ArgNo = std::distance(Callee->arg_begin(), 224 Function::arg_iterator(Arg)); 225 226 // Look at all call sites of the function. At this pointer we know we only 227 // have direct callees. 228 for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end(); 229 UI != E; ++UI) { 230 CallSite CS(*UI); 231 assert(CS && "Should only have direct calls!"); 232 233 if (!CS.getArgument(ArgNo)->isDereferenceablePointer()) 234 return false; 235 } 236 return true; 237 } 238 239 /// Returns true if Prefix is a prefix of longer. That means, Longer has a size 240 /// that is greater than or equal to the size of prefix, and each of the 241 /// elements in Prefix is the same as the corresponding elements in Longer. 242 /// 243 /// This means it also returns true when Prefix and Longer are equal! 244 static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix, 245 const ArgPromotion::IndicesVector &Longer) { 246 if (Prefix.size() > Longer.size()) 247 return false; 248 return std::equal(Prefix.begin(), Prefix.end(), Longer.begin()); 249 } 250 251 252 /// Checks if Indices, or a prefix of Indices, is in Set. 253 static bool PrefixIn(const ArgPromotion::IndicesVector &Indices, 254 std::set<ArgPromotion::IndicesVector> &Set) { 255 std::set<ArgPromotion::IndicesVector>::iterator Low; 256 Low = Set.upper_bound(Indices); 257 if (Low != Set.begin()) 258 Low--; 259 // Low is now the last element smaller than or equal to Indices. This means 260 // it points to a prefix of Indices (possibly Indices itself), if such 261 // prefix exists. 262 // 263 // This load is safe if any prefix of its operands is safe to load. 264 return Low != Set.end() && IsPrefix(*Low, Indices); 265 } 266 267 /// Mark the given indices (ToMark) as safe in the given set of indices 268 /// (Safe). Marking safe usually means adding ToMark to Safe. However, if there 269 /// is already a prefix of Indices in Safe, Indices are implicitely marked safe 270 /// already. Furthermore, any indices that Indices is itself a prefix of, are 271 /// removed from Safe (since they are implicitely safe because of Indices now). 272 static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark, 273 std::set<ArgPromotion::IndicesVector> &Safe) { 274 std::set<ArgPromotion::IndicesVector>::iterator Low; 275 Low = Safe.upper_bound(ToMark); 276 // Guard against the case where Safe is empty 277 if (Low != Safe.begin()) 278 Low--; 279 // Low is now the last element smaller than or equal to Indices. This 280 // means it points to a prefix of Indices (possibly Indices itself), if 281 // such prefix exists. 282 if (Low != Safe.end()) { 283 if (IsPrefix(*Low, ToMark)) 284 // If there is already a prefix of these indices (or exactly these 285 // indices) marked a safe, don't bother adding these indices 286 return; 287 288 // Increment Low, so we can use it as a "insert before" hint 289 ++Low; 290 } 291 // Insert 292 Low = Safe.insert(Low, ToMark); 293 ++Low; 294 // If there we're a prefix of longer index list(s), remove those 295 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end(); 296 while (Low != End && IsPrefix(ToMark, *Low)) { 297 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low; 298 ++Low; 299 Safe.erase(Remove); 300 } 301 } 302 303 /// isSafeToPromoteArgument - As you might guess from the name of this method, 304 /// it checks to see if it is both safe and useful to promote the argument. 305 /// This method limits promotion of aggregates to only promote up to three 306 /// elements of the aggregate in order to avoid exploding the number of 307 /// arguments passed in. 308 bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { 309 typedef std::set<IndicesVector> GEPIndicesSet; 310 311 // Quick exit for unused arguments 312 if (Arg->use_empty()) 313 return true; 314 315 // We can only promote this argument if all of the uses are loads, or are GEP 316 // instructions (with constant indices) that are subsequently loaded. 317 // 318 // Promoting the argument causes it to be loaded in the caller 319 // unconditionally. This is only safe if we can prove that either the load 320 // would have happened in the callee anyway (ie, there is a load in the entry 321 // block) or the pointer passed in at every call site is guaranteed to be 322 // valid. 323 // In the former case, invalid loads can happen, but would have happened 324 // anyway, in the latter case, invalid loads won't happen. This prevents us 325 // from introducing an invalid load that wouldn't have happened in the 326 // original code. 327 // 328 // This set will contain all sets of indices that are loaded in the entry 329 // block, and thus are safe to unconditionally load in the caller. 330 GEPIndicesSet SafeToUnconditionallyLoad; 331 332 // This set contains all the sets of indices that we are planning to promote. 333 // This makes it possible to limit the number of arguments added. 334 GEPIndicesSet ToPromote; 335 336 // If the pointer is always valid, any load with first index 0 is valid. 337 if (isByVal || AllCallersPassInValidPointerForArgument(Arg)) 338 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0)); 339 340 // First, iterate the entry block and mark loads of (geps of) arguments as 341 // safe. 342 BasicBlock *EntryBlock = Arg->getParent()->begin(); 343 // Declare this here so we can reuse it 344 IndicesVector Indices; 345 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end(); 346 I != E; ++I) 347 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 348 Value *V = LI->getPointerOperand(); 349 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) { 350 V = GEP->getPointerOperand(); 351 if (V == Arg) { 352 // This load actually loads (part of) Arg? Check the indices then. 353 Indices.reserve(GEP->getNumIndices()); 354 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 355 II != IE; ++II) 356 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II)) 357 Indices.push_back(CI->getSExtValue()); 358 else 359 // We found a non-constant GEP index for this argument? Bail out 360 // right away, can't promote this argument at all. 361 return false; 362 363 // Indices checked out, mark them as safe 364 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad); 365 Indices.clear(); 366 } 367 } else if (V == Arg) { 368 // Direct loads are equivalent to a GEP with a single 0 index. 369 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad); 370 } 371 } 372 373 // Now, iterate all uses of the argument to see if there are any uses that are 374 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote. 375 SmallVector<LoadInst*, 16> Loads; 376 IndicesVector Operands; 377 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end(); 378 UI != E; ++UI) { 379 User *U = *UI; 380 Operands.clear(); 381 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 382 // Don't hack volatile/atomic loads 383 if (!LI->isSimple()) return false; 384 Loads.push_back(LI); 385 // Direct loads are equivalent to a GEP with a zero index and then a load. 386 Operands.push_back(0); 387 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 388 if (GEP->use_empty()) { 389 // Dead GEP's cause trouble later. Just remove them if we run into 390 // them. 391 getAnalysis<AliasAnalysis>().deleteValue(GEP); 392 GEP->eraseFromParent(); 393 // TODO: This runs the above loop over and over again for dead GEPs 394 // Couldn't we just do increment the UI iterator earlier and erase the 395 // use? 396 return isSafeToPromoteArgument(Arg, isByVal); 397 } 398 399 // Ensure that all of the indices are constants. 400 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end(); 401 i != e; ++i) 402 if (ConstantInt *C = dyn_cast<ConstantInt>(*i)) 403 Operands.push_back(C->getSExtValue()); 404 else 405 return false; // Not a constant operand GEP! 406 407 // Ensure that the only users of the GEP are load instructions. 408 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end(); 409 UI != E; ++UI) 410 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 411 // Don't hack volatile/atomic loads 412 if (!LI->isSimple()) return false; 413 Loads.push_back(LI); 414 } else { 415 // Other uses than load? 416 return false; 417 } 418 } else { 419 return false; // Not a load or a GEP. 420 } 421 422 // Now, see if it is safe to promote this load / loads of this GEP. Loading 423 // is safe if Operands, or a prefix of Operands, is marked as safe. 424 if (!PrefixIn(Operands, SafeToUnconditionallyLoad)) 425 return false; 426 427 // See if we are already promoting a load with these indices. If not, check 428 // to make sure that we aren't promoting too many elements. If so, nothing 429 // to do. 430 if (ToPromote.find(Operands) == ToPromote.end()) { 431 if (maxElements > 0 && ToPromote.size() == maxElements) { 432 DEBUG(dbgs() << "argpromotion not promoting argument '" 433 << Arg->getName() << "' because it would require adding more " 434 << "than " << maxElements << " arguments to the function.\n"); 435 // We limit aggregate promotion to only promoting up to a fixed number 436 // of elements of the aggregate. 437 return false; 438 } 439 ToPromote.insert(Operands); 440 } 441 } 442 443 if (Loads.empty()) return true; // No users, this is a dead argument. 444 445 // Okay, now we know that the argument is only used by load instructions and 446 // it is safe to unconditionally perform all of them. Use alias analysis to 447 // check to see if the pointer is guaranteed to not be modified from entry of 448 // the function to each of the load instructions. 449 450 // Because there could be several/many load instructions, remember which 451 // blocks we know to be transparent to the load. 452 SmallPtrSet<BasicBlock*, 16> TranspBlocks; 453 454 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 455 456 for (unsigned i = 0, e = Loads.size(); i != e; ++i) { 457 // Check to see if the load is invalidated from the start of the block to 458 // the load itself. 459 LoadInst *Load = Loads[i]; 460 BasicBlock *BB = Load->getParent(); 461 462 AliasAnalysis::Location Loc = AA.getLocation(Load); 463 if (AA.canInstructionRangeModify(BB->front(), *Load, Loc)) 464 return false; // Pointer is invalidated! 465 466 // Now check every path from the entry block to the load for transparency. 467 // To do this, we perform a depth first search on the inverse CFG from the 468 // loading block. 469 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 470 BasicBlock *P = *PI; 471 for (idf_ext_iterator<BasicBlock*, SmallPtrSet<BasicBlock*, 16> > 472 I = idf_ext_begin(P, TranspBlocks), 473 E = idf_ext_end(P, TranspBlocks); I != E; ++I) 474 if (AA.canBasicBlockModify(**I, Loc)) 475 return false; 476 } 477 } 478 479 // If the path from the entry of the function to each load is free of 480 // instructions that potentially invalidate the load, we can make the 481 // transformation! 482 return true; 483 } 484 485 /// DoPromotion - This method actually performs the promotion of the specified 486 /// arguments, and returns the new function. At this point, we know that it's 487 /// safe to do so. 488 CallGraphNode *ArgPromotion::DoPromotion(Function *F, 489 SmallPtrSet<Argument*, 8> &ArgsToPromote, 490 SmallPtrSet<Argument*, 8> &ByValArgsToTransform) { 491 492 // Start by computing a new prototype for the function, which is the same as 493 // the old function, but has modified arguments. 494 FunctionType *FTy = F->getFunctionType(); 495 std::vector<Type*> Params; 496 497 typedef std::set<IndicesVector> ScalarizeTable; 498 499 // ScalarizedElements - If we are promoting a pointer that has elements 500 // accessed out of it, keep track of which elements are accessed so that we 501 // can add one argument for each. 502 // 503 // Arguments that are directly loaded will have a zero element value here, to 504 // handle cases where there are both a direct load and GEP accesses. 505 // 506 std::map<Argument*, ScalarizeTable> ScalarizedElements; 507 508 // OriginalLoads - Keep track of a representative load instruction from the 509 // original function so that we can tell the alias analysis implementation 510 // what the new GEP/Load instructions we are inserting look like. 511 std::map<IndicesVector, LoadInst*> OriginalLoads; 512 513 // Attributes - Keep track of the parameter attributes for the arguments 514 // that we are *not* promoting. For the ones that we do promote, the parameter 515 // attributes are lost 516 SmallVector<AttributeWithIndex, 8> AttributesVec; 517 const AttrListPtr &PAL = F->getAttributes(); 518 519 // Add any return attributes. 520 if (Attributes attrs = PAL.getRetAttributes()) 521 AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); 522 523 // First, determine the new argument list 524 unsigned ArgIndex = 1; 525 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; 526 ++I, ++ArgIndex) { 527 if (ByValArgsToTransform.count(I)) { 528 // Simple byval argument? Just add all the struct element types. 529 Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 530 StructType *STy = cast<StructType>(AgTy); 531 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 532 Params.push_back(STy->getElementType(i)); 533 ++NumByValArgsPromoted; 534 } else if (!ArgsToPromote.count(I)) { 535 // Unchanged argument 536 Params.push_back(I->getType()); 537 if (Attributes attrs = PAL.getParamAttributes(ArgIndex)) 538 AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs)); 539 } else if (I->use_empty()) { 540 // Dead argument (which are always marked as promotable) 541 ++NumArgumentsDead; 542 } else { 543 // Okay, this is being promoted. This means that the only uses are loads 544 // or GEPs which are only used by loads 545 546 // In this table, we will track which indices are loaded from the argument 547 // (where direct loads are tracked as no indices). 548 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 549 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 550 ++UI) { 551 Instruction *User = cast<Instruction>(*UI); 552 assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User)); 553 IndicesVector Indices; 554 Indices.reserve(User->getNumOperands() - 1); 555 // Since loads will only have a single operand, and GEPs only a single 556 // non-index operand, this will record direct loads without any indices, 557 // and gep+loads with the GEP indices. 558 for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end(); 559 II != IE; ++II) 560 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue()); 561 // GEPs with a single 0 index can be merged with direct loads 562 if (Indices.size() == 1 && Indices.front() == 0) 563 Indices.clear(); 564 ArgIndices.insert(Indices); 565 LoadInst *OrigLoad; 566 if (LoadInst *L = dyn_cast<LoadInst>(User)) 567 OrigLoad = L; 568 else 569 // Take any load, we will use it only to update Alias Analysis 570 OrigLoad = cast<LoadInst>(User->use_back()); 571 OriginalLoads[Indices] = OrigLoad; 572 } 573 574 // Add a parameter to the function for each element passed in. 575 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 576 E = ArgIndices.end(); SI != E; ++SI) { 577 // not allowed to dereference ->begin() if size() is 0 578 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), *SI)); 579 assert(Params.back()); 580 } 581 582 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty()) 583 ++NumArgumentsPromoted; 584 else 585 ++NumAggregatesPromoted; 586 } 587 } 588 589 // Add any function attributes. 590 if (Attributes attrs = PAL.getFnAttributes()) 591 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); 592 593 Type *RetTy = FTy->getReturnType(); 594 595 // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which 596 // have zero fixed arguments. 597 bool ExtraArgHack = false; 598 if (Params.empty() && FTy->isVarArg()) { 599 ExtraArgHack = true; 600 Params.push_back(Type::getInt32Ty(F->getContext())); 601 } 602 603 // Construct the new function type using the new arguments. 604 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg()); 605 606 // Create the new function body and insert it into the module. 607 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName()); 608 NF->copyAttributesFrom(F); 609 610 611 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n" 612 << "From: " << *F); 613 614 // Recompute the parameter attributes list based on the new arguments for 615 // the function. 616 NF->setAttributes(AttrListPtr::get(AttributesVec)); 617 AttributesVec.clear(); 618 619 F->getParent()->getFunctionList().insert(F, NF); 620 NF->takeName(F); 621 622 // Get the alias analysis information that we need to update to reflect our 623 // changes. 624 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 625 626 // Get the callgraph information that we need to update to reflect our 627 // changes. 628 CallGraph &CG = getAnalysis<CallGraph>(); 629 630 // Get a new callgraph node for NF. 631 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF); 632 633 // Loop over all of the callers of the function, transforming the call sites 634 // to pass in the loaded pointers. 635 // 636 SmallVector<Value*, 16> Args; 637 while (!F->use_empty()) { 638 CallSite CS(F->use_back()); 639 assert(CS.getCalledFunction() == F); 640 Instruction *Call = CS.getInstruction(); 641 const AttrListPtr &CallPAL = CS.getAttributes(); 642 643 // Add any return attributes. 644 if (Attributes attrs = CallPAL.getRetAttributes()) 645 AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); 646 647 // Loop over the operands, inserting GEP and loads in the caller as 648 // appropriate. 649 CallSite::arg_iterator AI = CS.arg_begin(); 650 ArgIndex = 1; 651 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 652 I != E; ++I, ++AI, ++ArgIndex) 653 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 654 Args.push_back(*AI); // Unmodified argument 655 656 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) 657 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 658 659 } else if (ByValArgsToTransform.count(I)) { 660 // Emit a GEP and load for each element of the struct. 661 Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 662 StructType *STy = cast<StructType>(AgTy); 663 Value *Idxs[2] = { 664 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; 665 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 666 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); 667 Value *Idx = GetElementPtrInst::Create(*AI, Idxs, 668 (*AI)->getName()+"."+utostr(i), 669 Call); 670 // TODO: Tell AA about the new values? 671 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call)); 672 } 673 } else if (!I->use_empty()) { 674 // Non-dead argument: insert GEPs and loads as appropriate. 675 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 676 // Store the Value* version of the indices in here, but declare it now 677 // for reuse. 678 std::vector<Value*> Ops; 679 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 680 E = ArgIndices.end(); SI != E; ++SI) { 681 Value *V = *AI; 682 LoadInst *OrigLoad = OriginalLoads[*SI]; 683 if (!SI->empty()) { 684 Ops.reserve(SI->size()); 685 Type *ElTy = V->getType(); 686 for (IndicesVector::const_iterator II = SI->begin(), 687 IE = SI->end(); II != IE; ++II) { 688 // Use i32 to index structs, and i64 for others (pointers/arrays). 689 // This satisfies GEP constraints. 690 Type *IdxTy = (ElTy->isStructTy() ? 691 Type::getInt32Ty(F->getContext()) : 692 Type::getInt64Ty(F->getContext())); 693 Ops.push_back(ConstantInt::get(IdxTy, *II)); 694 // Keep track of the type we're currently indexing. 695 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II); 696 } 697 // And create a GEP to extract those indices. 698 V = GetElementPtrInst::Create(V, Ops, V->getName()+".idx", Call); 699 Ops.clear(); 700 AA.copyValue(OrigLoad->getOperand(0), V); 701 } 702 // Since we're replacing a load make sure we take the alignment 703 // of the previous load. 704 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call); 705 newLoad->setAlignment(OrigLoad->getAlignment()); 706 // Transfer the TBAA info too. 707 newLoad->setMetadata(LLVMContext::MD_tbaa, 708 OrigLoad->getMetadata(LLVMContext::MD_tbaa)); 709 Args.push_back(newLoad); 710 AA.copyValue(OrigLoad, Args.back()); 711 } 712 } 713 714 if (ExtraArgHack) 715 Args.push_back(Constant::getNullValue(Type::getInt32Ty(F->getContext()))); 716 717 // Push any varargs arguments on the list. 718 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) { 719 Args.push_back(*AI); 720 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) 721 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 722 } 723 724 // Add any function attributes. 725 if (Attributes attrs = CallPAL.getFnAttributes()) 726 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); 727 728 Instruction *New; 729 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 730 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 731 Args, "", Call); 732 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 733 cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec)); 734 } else { 735 New = CallInst::Create(NF, Args, "", Call); 736 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 737 cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec)); 738 if (cast<CallInst>(Call)->isTailCall()) 739 cast<CallInst>(New)->setTailCall(); 740 } 741 Args.clear(); 742 AttributesVec.clear(); 743 744 // Update the alias analysis implementation to know that we are replacing 745 // the old call with a new one. 746 AA.replaceWithNewValue(Call, New); 747 748 // Update the callgraph to know that the callsite has been transformed. 749 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()]; 750 CalleeNode->replaceCallEdge(Call, New, NF_CGN); 751 752 if (!Call->use_empty()) { 753 Call->replaceAllUsesWith(New); 754 New->takeName(Call); 755 } 756 757 // Finally, remove the old call from the program, reducing the use-count of 758 // F. 759 Call->eraseFromParent(); 760 } 761 762 // Since we have now created the new function, splice the body of the old 763 // function right into the new function, leaving the old rotting hulk of the 764 // function empty. 765 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 766 767 // Loop over the argument list, transferring uses of the old arguments over to 768 // the new arguments, also transferring over the names as well. 769 // 770 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), 771 I2 = NF->arg_begin(); I != E; ++I) { 772 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 773 // If this is an unmodified argument, move the name and users over to the 774 // new version. 775 I->replaceAllUsesWith(I2); 776 I2->takeName(I); 777 AA.replaceWithNewValue(I, I2); 778 ++I2; 779 continue; 780 } 781 782 if (ByValArgsToTransform.count(I)) { 783 // In the callee, we create an alloca, and store each of the new incoming 784 // arguments into the alloca. 785 Instruction *InsertPt = NF->begin()->begin(); 786 787 // Just add all the struct element types. 788 Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 789 Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt); 790 StructType *STy = cast<StructType>(AgTy); 791 Value *Idxs[2] = { 792 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; 793 794 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 795 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); 796 Value *Idx = 797 GetElementPtrInst::Create(TheAlloca, Idxs, 798 TheAlloca->getName()+"."+Twine(i), 799 InsertPt); 800 I2->setName(I->getName()+"."+Twine(i)); 801 new StoreInst(I2++, Idx, InsertPt); 802 } 803 804 // Anything that used the arg should now use the alloca. 805 I->replaceAllUsesWith(TheAlloca); 806 TheAlloca->takeName(I); 807 AA.replaceWithNewValue(I, TheAlloca); 808 continue; 809 } 810 811 if (I->use_empty()) { 812 AA.deleteValue(I); 813 continue; 814 } 815 816 // Otherwise, if we promoted this argument, then all users are load 817 // instructions (or GEPs with only load users), and all loads should be 818 // using the new argument that we added. 819 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 820 821 while (!I->use_empty()) { 822 if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) { 823 assert(ArgIndices.begin()->empty() && 824 "Load element should sort to front!"); 825 I2->setName(I->getName()+".val"); 826 LI->replaceAllUsesWith(I2); 827 AA.replaceWithNewValue(LI, I2); 828 LI->eraseFromParent(); 829 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName() 830 << "' in function '" << F->getName() << "'\n"); 831 } else { 832 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back()); 833 IndicesVector Operands; 834 Operands.reserve(GEP->getNumIndices()); 835 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 836 II != IE; ++II) 837 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue()); 838 839 // GEPs with a single 0 index can be merged with direct loads 840 if (Operands.size() == 1 && Operands.front() == 0) 841 Operands.clear(); 842 843 Function::arg_iterator TheArg = I2; 844 for (ScalarizeTable::iterator It = ArgIndices.begin(); 845 *It != Operands; ++It, ++TheArg) { 846 assert(It != ArgIndices.end() && "GEP not handled??"); 847 } 848 849 std::string NewName = I->getName(); 850 for (unsigned i = 0, e = Operands.size(); i != e; ++i) { 851 NewName += "." + utostr(Operands[i]); 852 } 853 NewName += ".val"; 854 TheArg->setName(NewName); 855 856 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName() 857 << "' of function '" << NF->getName() << "'\n"); 858 859 // All of the uses must be load instructions. Replace them all with 860 // the argument specified by ArgNo. 861 while (!GEP->use_empty()) { 862 LoadInst *L = cast<LoadInst>(GEP->use_back()); 863 L->replaceAllUsesWith(TheArg); 864 AA.replaceWithNewValue(L, TheArg); 865 L->eraseFromParent(); 866 } 867 AA.deleteValue(GEP); 868 GEP->eraseFromParent(); 869 } 870 } 871 872 // Increment I2 past all of the arguments added for this promoted pointer. 873 for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i) 874 ++I2; 875 } 876 877 // Notify the alias analysis implementation that we inserted a new argument. 878 if (ExtraArgHack) 879 AA.copyValue(Constant::getNullValue(Type::getInt32Ty(F->getContext())), 880 NF->arg_begin()); 881 882 883 // Tell the alias analysis that the old function is about to disappear. 884 AA.replaceWithNewValue(F, NF); 885 886 887 NF_CGN->stealCalledFunctionsFrom(CG[F]); 888 889 // Now that the old function is dead, delete it. If there is a dangling 890 // reference to the CallgraphNode, just leave the dead function around for 891 // someone else to nuke. 892 CallGraphNode *CGN = CG[F]; 893 if (CGN->getNumReferences() == 0) 894 delete CG.removeFunctionFromModule(CGN); 895 else 896 F->setLinkage(Function::ExternalLinkage); 897 898 return NF_CGN; 899 } 900