ArgumentPromotion.cpp revision 32f5cb23caae91a35c354496a9dd3edddbafbc17
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/Target/TargetData.h" 43#include "llvm/Support/CallSite.h" 44#include "llvm/Support/CFG.h" 45#include "llvm/Support/Debug.h" 46#include "llvm/Support/raw_ostream.h" 47#include "llvm/ADT/DepthFirstIterator.h" 48#include "llvm/ADT/Statistic.h" 49#include "llvm/ADT/StringExtras.h" 50#include <set> 51using namespace llvm; 52 53STATISTIC(NumArgumentsPromoted , "Number of pointer arguments promoted"); 54STATISTIC(NumAggregatesPromoted, "Number of aggregate arguments promoted"); 55STATISTIC(NumByValArgsPromoted , "Number of byval arguments promoted"); 56STATISTIC(NumArgumentsDead , "Number of dead pointer args eliminated"); 57 58namespace { 59 /// ArgPromotion - The 'by reference' to 'by value' argument promotion pass. 60 /// 61 struct ArgPromotion : public CallGraphSCCPass { 62 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 63 AU.addRequired<AliasAnalysis>(); 64 CallGraphSCCPass::getAnalysisUsage(AU); 65 } 66 67 virtual bool runOnSCC(CallGraphSCC &SCC); 68 static char ID; // Pass identification, replacement for typeid 69 explicit ArgPromotion(unsigned maxElements = 3) 70 : CallGraphSCCPass(ID), maxElements(maxElements) { 71 initializeArgPromotionPass(*PassRegistry::getPassRegistry()); 72 } 73 74 /// A vector used to hold the indices of a single GEP instruction 75 typedef std::vector<uint64_t> IndicesVector; 76 77 private: 78 CallGraphNode *PromoteArguments(CallGraphNode *CGN); 79 bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const; 80 CallGraphNode *DoPromotion(Function *F, 81 SmallPtrSet<Argument*, 8> &ArgsToPromote, 82 SmallPtrSet<Argument*, 8> &ByValArgsToTransform); 83 /// The maximum number of elements to expand, or 0 for unlimited. 84 unsigned maxElements; 85 }; 86} 87 88char ArgPromotion::ID = 0; 89INITIALIZE_PASS_BEGIN(ArgPromotion, "argpromotion", 90 "Promote 'by reference' arguments to scalars", false, false) 91INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 92INITIALIZE_AG_DEPENDENCY(CallGraph) 93INITIALIZE_PASS_END(ArgPromotion, "argpromotion", 94 "Promote 'by reference' arguments to scalars", false, false) 95 96Pass *llvm::createArgumentPromotionPass(unsigned maxElements) { 97 return new ArgPromotion(maxElements); 98} 99 100bool ArgPromotion::runOnSCC(CallGraphSCC &SCC) { 101 bool Changed = false, LocalChange; 102 103 do { // Iterate until we stop promoting from this SCC. 104 LocalChange = false; 105 // Attempt to promote arguments from all functions in this SCC. 106 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) { 107 if (CallGraphNode *CGN = PromoteArguments(*I)) { 108 LocalChange = true; 109 SCC.ReplaceNode(*I, CGN); 110 } 111 } 112 Changed |= LocalChange; // Remember that we changed something. 113 } while (LocalChange); 114 115 return Changed; 116} 117 118/// PromoteArguments - This method checks the specified function to see if there 119/// are any promotable arguments and if it is safe to promote the function (for 120/// example, all callers are direct). If safe to promote some arguments, it 121/// calls the DoPromotion method. 122/// 123CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) { 124 Function *F = CGN->getFunction(); 125 126 // Make sure that it is local to this module. 127 if (!F || !F->hasLocalLinkage()) return 0; 128 129 // First check: see if there are any pointer arguments! If not, quick exit. 130 SmallVector<std::pair<Argument*, unsigned>, 16> PointerArgs; 131 unsigned ArgNo = 0; 132 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 133 I != E; ++I, ++ArgNo) 134 if (I->getType()->isPointerTy()) 135 PointerArgs.push_back(std::pair<Argument*, unsigned>(I, ArgNo)); 136 if (PointerArgs.empty()) return 0; 137 138 // Second check: make sure that all callers are direct callers. We can't 139 // transform functions that have indirect callers. 140 if (F->hasAddressTaken()) 141 return 0; 142 143 // Check to see which arguments are promotable. If an argument is promotable, 144 // add it to ArgsToPromote. 145 SmallPtrSet<Argument*, 8> ArgsToPromote; 146 SmallPtrSet<Argument*, 8> ByValArgsToTransform; 147 for (unsigned i = 0; i != PointerArgs.size(); ++i) { 148 bool isByVal = F->paramHasAttr(PointerArgs[i].second+1, Attribute::ByVal); 149 150 // If this is a byval argument, and if the aggregate type is small, just 151 // pass the elements, which is always safe. 152 Argument *PtrArg = PointerArgs[i].first; 153 if (isByVal) { 154 const Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType(); 155 if (const StructType *STy = dyn_cast<StructType>(AgTy)) { 156 if (maxElements > 0 && STy->getNumElements() > maxElements) { 157 DEBUG(dbgs() << "argpromotion disable promoting argument '" 158 << PtrArg->getName() << "' because it would require adding more" 159 << " than " << maxElements << " arguments to the function.\n"); 160 } else { 161 // If all the elements are single-value types, we can promote it. 162 bool AllSimple = true; 163 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 164 if (!STy->getElementType(i)->isSingleValueType()) { 165 AllSimple = false; 166 break; 167 } 168 169 // Safe to transform, don't even bother trying to "promote" it. 170 // Passing the elements as a scalar will allow scalarrepl to hack on 171 // the new alloca we introduce. 172 if (AllSimple) { 173 ByValArgsToTransform.insert(PtrArg); 174 continue; 175 } 176 } 177 } 178 } 179 180 // Otherwise, see if we can promote the pointer to its value. 181 if (isSafeToPromoteArgument(PtrArg, isByVal)) 182 ArgsToPromote.insert(PtrArg); 183 } 184 185 // No promotable pointer arguments. 186 if (ArgsToPromote.empty() && ByValArgsToTransform.empty()) 187 return 0; 188 189 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform); 190} 191 192/// IsAlwaysValidPointer - Return true if the specified pointer is always legal 193/// to load. 194static bool IsAlwaysValidPointer(Value *V) { 195 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true; 196 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) 197 return IsAlwaysValidPointer(GEP->getOperand(0)); 198 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 199 if (CE->getOpcode() == Instruction::GetElementPtr) 200 return IsAlwaysValidPointer(CE->getOperand(0)); 201 202 return false; 203} 204 205/// AllCalleesPassInValidPointerForArgument - Return true if we can prove that 206/// all callees pass in a valid pointer for the specified function argument. 207static bool AllCalleesPassInValidPointerForArgument(Argument *Arg) { 208 Function *Callee = Arg->getParent(); 209 210 unsigned ArgNo = std::distance(Callee->arg_begin(), 211 Function::arg_iterator(Arg)); 212 213 // Look at all call sites of the function. At this pointer we know we only 214 // have direct callees. 215 for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end(); 216 UI != E; ++UI) { 217 CallSite CS(*UI); 218 assert(CS && "Should only have direct calls!"); 219 220 if (!IsAlwaysValidPointer(CS.getArgument(ArgNo))) 221 return false; 222 } 223 return true; 224} 225 226/// Returns true if Prefix is a prefix of longer. That means, Longer has a size 227/// that is greater than or equal to the size of prefix, and each of the 228/// elements in Prefix is the same as the corresponding elements in Longer. 229/// 230/// This means it also returns true when Prefix and Longer are equal! 231static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix, 232 const ArgPromotion::IndicesVector &Longer) { 233 if (Prefix.size() > Longer.size()) 234 return false; 235 for (unsigned i = 0, e = Prefix.size(); i != e; ++i) 236 if (Prefix[i] != Longer[i]) 237 return false; 238 return true; 239} 240 241 242/// Checks if Indices, or a prefix of Indices, is in Set. 243static bool PrefixIn(const ArgPromotion::IndicesVector &Indices, 244 std::set<ArgPromotion::IndicesVector> &Set) { 245 std::set<ArgPromotion::IndicesVector>::iterator Low; 246 Low = Set.upper_bound(Indices); 247 if (Low != Set.begin()) 248 Low--; 249 // Low is now the last element smaller than or equal to Indices. This means 250 // it points to a prefix of Indices (possibly Indices itself), if such 251 // prefix exists. 252 // 253 // This load is safe if any prefix of its operands is safe to load. 254 return Low != Set.end() && IsPrefix(*Low, Indices); 255} 256 257/// Mark the given indices (ToMark) as safe in the given set of indices 258/// (Safe). Marking safe usually means adding ToMark to Safe. However, if there 259/// is already a prefix of Indices in Safe, Indices are implicitely marked safe 260/// already. Furthermore, any indices that Indices is itself a prefix of, are 261/// removed from Safe (since they are implicitely safe because of Indices now). 262static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark, 263 std::set<ArgPromotion::IndicesVector> &Safe) { 264 std::set<ArgPromotion::IndicesVector>::iterator Low; 265 Low = Safe.upper_bound(ToMark); 266 // Guard against the case where Safe is empty 267 if (Low != Safe.begin()) 268 Low--; 269 // Low is now the last element smaller than or equal to Indices. This 270 // means it points to a prefix of Indices (possibly Indices itself), if 271 // such prefix exists. 272 if (Low != Safe.end()) { 273 if (IsPrefix(*Low, ToMark)) 274 // If there is already a prefix of these indices (or exactly these 275 // indices) marked a safe, don't bother adding these indices 276 return; 277 278 // Increment Low, so we can use it as a "insert before" hint 279 ++Low; 280 } 281 // Insert 282 Low = Safe.insert(Low, ToMark); 283 ++Low; 284 // If there we're a prefix of longer index list(s), remove those 285 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end(); 286 while (Low != End && IsPrefix(ToMark, *Low)) { 287 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low; 288 ++Low; 289 Safe.erase(Remove); 290 } 291} 292 293/// isSafeToPromoteArgument - As you might guess from the name of this method, 294/// it checks to see if it is both safe and useful to promote the argument. 295/// This method limits promotion of aggregates to only promote up to three 296/// elements of the aggregate in order to avoid exploding the number of 297/// arguments passed in. 298bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { 299 typedef std::set<IndicesVector> GEPIndicesSet; 300 301 // Quick exit for unused arguments 302 if (Arg->use_empty()) 303 return true; 304 305 // We can only promote this argument if all of the uses are loads, or are GEP 306 // instructions (with constant indices) that are subsequently loaded. 307 // 308 // Promoting the argument causes it to be loaded in the caller 309 // unconditionally. This is only safe if we can prove that either the load 310 // would have happened in the callee anyway (ie, there is a load in the entry 311 // block) or the pointer passed in at every call site is guaranteed to be 312 // valid. 313 // In the former case, invalid loads can happen, but would have happened 314 // anyway, in the latter case, invalid loads won't happen. This prevents us 315 // from introducing an invalid load that wouldn't have happened in the 316 // original code. 317 // 318 // This set will contain all sets of indices that are loaded in the entry 319 // block, and thus are safe to unconditionally load in the caller. 320 GEPIndicesSet SafeToUnconditionallyLoad; 321 322 // This set contains all the sets of indices that we are planning to promote. 323 // This makes it possible to limit the number of arguments added. 324 GEPIndicesSet ToPromote; 325 326 // If the pointer is always valid, any load with first index 0 is valid. 327 if (isByVal || AllCalleesPassInValidPointerForArgument(Arg)) 328 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0)); 329 330 // First, iterate the entry block and mark loads of (geps of) arguments as 331 // safe. 332 BasicBlock *EntryBlock = Arg->getParent()->begin(); 333 // Declare this here so we can reuse it 334 IndicesVector Indices; 335 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end(); 336 I != E; ++I) 337 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 338 Value *V = LI->getPointerOperand(); 339 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) { 340 V = GEP->getPointerOperand(); 341 if (V == Arg) { 342 // This load actually loads (part of) Arg? Check the indices then. 343 Indices.reserve(GEP->getNumIndices()); 344 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 345 II != IE; ++II) 346 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II)) 347 Indices.push_back(CI->getSExtValue()); 348 else 349 // We found a non-constant GEP index for this argument? Bail out 350 // right away, can't promote this argument at all. 351 return false; 352 353 // Indices checked out, mark them as safe 354 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad); 355 Indices.clear(); 356 } 357 } else if (V == Arg) { 358 // Direct loads are equivalent to a GEP with a single 0 index. 359 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad); 360 } 361 } 362 363 // Now, iterate all uses of the argument to see if there are any uses that are 364 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote. 365 SmallVector<LoadInst*, 16> Loads; 366 IndicesVector Operands; 367 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end(); 368 UI != E; ++UI) { 369 User *U = *UI; 370 Operands.clear(); 371 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 372 if (LI->isVolatile()) return false; // Don't hack volatile loads 373 Loads.push_back(LI); 374 // Direct loads are equivalent to a GEP with a zero index and then a load. 375 Operands.push_back(0); 376 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 377 if (GEP->use_empty()) { 378 // Dead GEP's cause trouble later. Just remove them if we run into 379 // them. 380 getAnalysis<AliasAnalysis>().deleteValue(GEP); 381 GEP->eraseFromParent(); 382 // TODO: This runs the above loop over and over again for dead GEPs 383 // Couldn't we just do increment the UI iterator earlier and erase the 384 // use? 385 return isSafeToPromoteArgument(Arg, isByVal); 386 } 387 388 // Ensure that all of the indices are constants. 389 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end(); 390 i != e; ++i) 391 if (ConstantInt *C = dyn_cast<ConstantInt>(*i)) 392 Operands.push_back(C->getSExtValue()); 393 else 394 return false; // Not a constant operand GEP! 395 396 // Ensure that the only users of the GEP are load instructions. 397 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end(); 398 UI != E; ++UI) 399 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 400 if (LI->isVolatile()) return false; // Don't hack volatile loads 401 Loads.push_back(LI); 402 } else { 403 // Other uses than load? 404 return false; 405 } 406 } else { 407 return false; // Not a load or a GEP. 408 } 409 410 // Now, see if it is safe to promote this load / loads of this GEP. Loading 411 // is safe if Operands, or a prefix of Operands, is marked as safe. 412 if (!PrefixIn(Operands, SafeToUnconditionallyLoad)) 413 return false; 414 415 // See if we are already promoting a load with these indices. If not, check 416 // to make sure that we aren't promoting too many elements. If so, nothing 417 // to do. 418 if (ToPromote.find(Operands) == ToPromote.end()) { 419 if (maxElements > 0 && ToPromote.size() == maxElements) { 420 DEBUG(dbgs() << "argpromotion not promoting argument '" 421 << Arg->getName() << "' because it would require adding more " 422 << "than " << maxElements << " arguments to the function.\n"); 423 // We limit aggregate promotion to only promoting up to a fixed number 424 // of elements of the aggregate. 425 return false; 426 } 427 ToPromote.insert(Operands); 428 } 429 } 430 431 if (Loads.empty()) return true; // No users, this is a dead argument. 432 433 // Okay, now we know that the argument is only used by load instructions and 434 // it is safe to unconditionally perform all of them. Use alias analysis to 435 // check to see if the pointer is guaranteed to not be modified from entry of 436 // the function to each of the load instructions. 437 438 // Because there could be several/many load instructions, remember which 439 // blocks we know to be transparent to the load. 440 SmallPtrSet<BasicBlock*, 16> TranspBlocks; 441 442 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 443 TargetData *TD = getAnalysisIfAvailable<TargetData>(); 444 if (!TD) return false; // Without TargetData, assume the worst. 445 446 for (unsigned i = 0, e = Loads.size(); i != e; ++i) { 447 // Check to see if the load is invalidated from the start of the block to 448 // the load itself. 449 LoadInst *Load = Loads[i]; 450 BasicBlock *BB = Load->getParent(); 451 452 const PointerType *LoadTy = 453 cast<PointerType>(Load->getPointerOperand()->getType()); 454 uint64_t LoadSize = TD->getTypeStoreSize(LoadTy->getElementType()); 455 456 if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize)) 457 return false; // Pointer is invalidated! 458 459 // Now check every path from the entry block to the load for transparency. 460 // To do this, we perform a depth first search on the inverse CFG from the 461 // loading block. 462 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 463 BasicBlock *P = *PI; 464 for (idf_ext_iterator<BasicBlock*, SmallPtrSet<BasicBlock*, 16> > 465 I = idf_ext_begin(P, TranspBlocks), 466 E = idf_ext_end(P, TranspBlocks); I != E; ++I) 467 if (AA.canBasicBlockModify(**I, Arg, LoadSize)) 468 return false; 469 } 470 } 471 472 // If the path from the entry of the function to each load is free of 473 // instructions that potentially invalidate the load, we can make the 474 // transformation! 475 return true; 476} 477 478/// DoPromotion - This method actually performs the promotion of the specified 479/// arguments, and returns the new function. At this point, we know that it's 480/// safe to do so. 481CallGraphNode *ArgPromotion::DoPromotion(Function *F, 482 SmallPtrSet<Argument*, 8> &ArgsToPromote, 483 SmallPtrSet<Argument*, 8> &ByValArgsToTransform) { 484 485 // Start by computing a new prototype for the function, which is the same as 486 // the old function, but has modified arguments. 487 const FunctionType *FTy = F->getFunctionType(); 488 std::vector<const Type*> Params; 489 490 typedef std::set<IndicesVector> ScalarizeTable; 491 492 // ScalarizedElements - If we are promoting a pointer that has elements 493 // accessed out of it, keep track of which elements are accessed so that we 494 // can add one argument for each. 495 // 496 // Arguments that are directly loaded will have a zero element value here, to 497 // handle cases where there are both a direct load and GEP accesses. 498 // 499 std::map<Argument*, ScalarizeTable> ScalarizedElements; 500 501 // OriginalLoads - Keep track of a representative load instruction from the 502 // original function so that we can tell the alias analysis implementation 503 // what the new GEP/Load instructions we are inserting look like. 504 std::map<IndicesVector, LoadInst*> OriginalLoads; 505 506 // Attributes - Keep track of the parameter attributes for the arguments 507 // that we are *not* promoting. For the ones that we do promote, the parameter 508 // attributes are lost 509 SmallVector<AttributeWithIndex, 8> AttributesVec; 510 const AttrListPtr &PAL = F->getAttributes(); 511 512 // Add any return attributes. 513 if (Attributes attrs = PAL.getRetAttributes()) 514 AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); 515 516 // First, determine the new argument list 517 unsigned ArgIndex = 1; 518 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; 519 ++I, ++ArgIndex) { 520 if (ByValArgsToTransform.count(I)) { 521 // Simple byval argument? Just add all the struct element types. 522 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 523 const StructType *STy = cast<StructType>(AgTy); 524 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 525 Params.push_back(STy->getElementType(i)); 526 ++NumByValArgsPromoted; 527 } else if (!ArgsToPromote.count(I)) { 528 // Unchanged argument 529 Params.push_back(I->getType()); 530 if (Attributes attrs = PAL.getParamAttributes(ArgIndex)) 531 AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs)); 532 } else if (I->use_empty()) { 533 // Dead argument (which are always marked as promotable) 534 ++NumArgumentsDead; 535 } else { 536 // Okay, this is being promoted. This means that the only uses are loads 537 // or GEPs which are only used by loads 538 539 // In this table, we will track which indices are loaded from the argument 540 // (where direct loads are tracked as no indices). 541 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 542 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 543 ++UI) { 544 Instruction *User = cast<Instruction>(*UI); 545 assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User)); 546 IndicesVector Indices; 547 Indices.reserve(User->getNumOperands() - 1); 548 // Since loads will only have a single operand, and GEPs only a single 549 // non-index operand, this will record direct loads without any indices, 550 // and gep+loads with the GEP indices. 551 for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end(); 552 II != IE; ++II) 553 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue()); 554 // GEPs with a single 0 index can be merged with direct loads 555 if (Indices.size() == 1 && Indices.front() == 0) 556 Indices.clear(); 557 ArgIndices.insert(Indices); 558 LoadInst *OrigLoad; 559 if (LoadInst *L = dyn_cast<LoadInst>(User)) 560 OrigLoad = L; 561 else 562 // Take any load, we will use it only to update Alias Analysis 563 OrigLoad = cast<LoadInst>(User->use_back()); 564 OriginalLoads[Indices] = OrigLoad; 565 } 566 567 // Add a parameter to the function for each element passed in. 568 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 569 E = ArgIndices.end(); SI != E; ++SI) { 570 // not allowed to dereference ->begin() if size() is 0 571 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), 572 SI->begin(), 573 SI->end())); 574 assert(Params.back()); 575 } 576 577 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty()) 578 ++NumArgumentsPromoted; 579 else 580 ++NumAggregatesPromoted; 581 } 582 } 583 584 // Add any function attributes. 585 if (Attributes attrs = PAL.getFnAttributes()) 586 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); 587 588 const Type *RetTy = FTy->getReturnType(); 589 590 // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which 591 // have zero fixed arguments. 592 bool ExtraArgHack = false; 593 if (Params.empty() && FTy->isVarArg()) { 594 ExtraArgHack = true; 595 Params.push_back(Type::getInt32Ty(F->getContext())); 596 } 597 598 // Construct the new function type using the new arguments. 599 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg()); 600 601 // Create the new function body and insert it into the module. 602 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName()); 603 NF->copyAttributesFrom(F); 604 605 606 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n" 607 << "From: " << *F); 608 609 // Recompute the parameter attributes list based on the new arguments for 610 // the function. 611 NF->setAttributes(AttrListPtr::get(AttributesVec.begin(), 612 AttributesVec.end())); 613 AttributesVec.clear(); 614 615 F->getParent()->getFunctionList().insert(F, NF); 616 NF->takeName(F); 617 618 // Get the alias analysis information that we need to update to reflect our 619 // changes. 620 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 621 622 // Get the callgraph information that we need to update to reflect our 623 // changes. 624 CallGraph &CG = getAnalysis<CallGraph>(); 625 626 // Get a new callgraph node for NF. 627 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF); 628 629 // Loop over all of the callers of the function, transforming the call sites 630 // to pass in the loaded pointers. 631 // 632 SmallVector<Value*, 16> Args; 633 while (!F->use_empty()) { 634 CallSite CS(F->use_back()); 635 assert(CS.getCalledFunction() == F); 636 Instruction *Call = CS.getInstruction(); 637 const AttrListPtr &CallPAL = CS.getAttributes(); 638 639 // Add any return attributes. 640 if (Attributes attrs = CallPAL.getRetAttributes()) 641 AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); 642 643 // Loop over the operands, inserting GEP and loads in the caller as 644 // appropriate. 645 CallSite::arg_iterator AI = CS.arg_begin(); 646 ArgIndex = 1; 647 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 648 I != E; ++I, ++AI, ++ArgIndex) 649 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 650 Args.push_back(*AI); // Unmodified argument 651 652 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) 653 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 654 655 } else if (ByValArgsToTransform.count(I)) { 656 // Emit a GEP and load for each element of the struct. 657 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 658 const StructType *STy = cast<StructType>(AgTy); 659 Value *Idxs[2] = { 660 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; 661 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 662 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); 663 Value *Idx = GetElementPtrInst::Create(*AI, Idxs, Idxs+2, 664 (*AI)->getName()+"."+utostr(i), 665 Call); 666 // TODO: Tell AA about the new values? 667 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call)); 668 } 669 } else if (!I->use_empty()) { 670 // Non-dead argument: insert GEPs and loads as appropriate. 671 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 672 // Store the Value* version of the indices in here, but declare it now 673 // for reuse. 674 std::vector<Value*> Ops; 675 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 676 E = ArgIndices.end(); SI != E; ++SI) { 677 Value *V = *AI; 678 LoadInst *OrigLoad = OriginalLoads[*SI]; 679 if (!SI->empty()) { 680 Ops.reserve(SI->size()); 681 const Type *ElTy = V->getType(); 682 for (IndicesVector::const_iterator II = SI->begin(), 683 IE = SI->end(); II != IE; ++II) { 684 // Use i32 to index structs, and i64 for others (pointers/arrays). 685 // This satisfies GEP constraints. 686 const Type *IdxTy = (ElTy->isStructTy() ? 687 Type::getInt32Ty(F->getContext()) : 688 Type::getInt64Ty(F->getContext())); 689 Ops.push_back(ConstantInt::get(IdxTy, *II)); 690 // Keep track of the type we're currently indexing. 691 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II); 692 } 693 // And create a GEP to extract those indices. 694 V = GetElementPtrInst::Create(V, Ops.begin(), Ops.end(), 695 V->getName()+".idx", Call); 696 Ops.clear(); 697 AA.copyValue(OrigLoad->getOperand(0), V); 698 } 699 // Since we're replacing a load make sure we take the alignment 700 // of the previous load. 701 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call); 702 newLoad->setAlignment(OrigLoad->getAlignment()); 703 Args.push_back(newLoad); 704 AA.copyValue(OrigLoad, Args.back()); 705 } 706 } 707 708 if (ExtraArgHack) 709 Args.push_back(Constant::getNullValue(Type::getInt32Ty(F->getContext()))); 710 711 // Push any varargs arguments on the list. 712 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) { 713 Args.push_back(*AI); 714 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) 715 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 716 } 717 718 // Add any function attributes. 719 if (Attributes attrs = CallPAL.getFnAttributes()) 720 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); 721 722 Instruction *New; 723 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 724 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 725 Args.begin(), Args.end(), "", Call); 726 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 727 cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(), 728 AttributesVec.end())); 729 } else { 730 New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call); 731 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 732 cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(), 733 AttributesVec.end())); 734 if (cast<CallInst>(Call)->isTailCall()) 735 cast<CallInst>(New)->setTailCall(); 736 } 737 Args.clear(); 738 AttributesVec.clear(); 739 740 // Update the alias analysis implementation to know that we are replacing 741 // the old call with a new one. 742 AA.replaceWithNewValue(Call, New); 743 744 // Update the callgraph to know that the callsite has been transformed. 745 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()]; 746 CalleeNode->replaceCallEdge(Call, New, NF_CGN); 747 748 if (!Call->use_empty()) { 749 Call->replaceAllUsesWith(New); 750 New->takeName(Call); 751 } 752 753 // Finally, remove the old call from the program, reducing the use-count of 754 // F. 755 Call->eraseFromParent(); 756 } 757 758 // Since we have now created the new function, splice the body of the old 759 // function right into the new function, leaving the old rotting hulk of the 760 // function empty. 761 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 762 763 // Loop over the argument list, transfering uses of the old arguments over to 764 // the new arguments, also transfering over the names as well. 765 // 766 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), 767 I2 = NF->arg_begin(); I != E; ++I) { 768 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 769 // If this is an unmodified argument, move the name and users over to the 770 // new version. 771 I->replaceAllUsesWith(I2); 772 I2->takeName(I); 773 AA.replaceWithNewValue(I, I2); 774 ++I2; 775 continue; 776 } 777 778 if (ByValArgsToTransform.count(I)) { 779 // In the callee, we create an alloca, and store each of the new incoming 780 // arguments into the alloca. 781 Instruction *InsertPt = NF->begin()->begin(); 782 783 // Just add all the struct element types. 784 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 785 Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt); 786 const StructType *STy = cast<StructType>(AgTy); 787 Value *Idxs[2] = { 788 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; 789 790 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 791 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); 792 Value *Idx = 793 GetElementPtrInst::Create(TheAlloca, Idxs, Idxs+2, 794 TheAlloca->getName()+"."+Twine(i), 795 InsertPt); 796 I2->setName(I->getName()+"."+Twine(i)); 797 new StoreInst(I2++, Idx, InsertPt); 798 } 799 800 // Anything that used the arg should now use the alloca. 801 I->replaceAllUsesWith(TheAlloca); 802 TheAlloca->takeName(I); 803 AA.replaceWithNewValue(I, TheAlloca); 804 continue; 805 } 806 807 if (I->use_empty()) { 808 AA.deleteValue(I); 809 continue; 810 } 811 812 // Otherwise, if we promoted this argument, then all users are load 813 // instructions (or GEPs with only load users), and all loads should be 814 // using the new argument that we added. 815 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 816 817 while (!I->use_empty()) { 818 if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) { 819 assert(ArgIndices.begin()->empty() && 820 "Load element should sort to front!"); 821 I2->setName(I->getName()+".val"); 822 LI->replaceAllUsesWith(I2); 823 AA.replaceWithNewValue(LI, I2); 824 LI->eraseFromParent(); 825 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName() 826 << "' in function '" << F->getName() << "'\n"); 827 } else { 828 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back()); 829 IndicesVector Operands; 830 Operands.reserve(GEP->getNumIndices()); 831 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 832 II != IE; ++II) 833 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue()); 834 835 // GEPs with a single 0 index can be merged with direct loads 836 if (Operands.size() == 1 && Operands.front() == 0) 837 Operands.clear(); 838 839 Function::arg_iterator TheArg = I2; 840 for (ScalarizeTable::iterator It = ArgIndices.begin(); 841 *It != Operands; ++It, ++TheArg) { 842 assert(It != ArgIndices.end() && "GEP not handled??"); 843 } 844 845 std::string NewName = I->getName(); 846 for (unsigned i = 0, e = Operands.size(); i != e; ++i) { 847 NewName += "." + utostr(Operands[i]); 848 } 849 NewName += ".val"; 850 TheArg->setName(NewName); 851 852 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName() 853 << "' of function '" << NF->getName() << "'\n"); 854 855 // All of the uses must be load instructions. Replace them all with 856 // the argument specified by ArgNo. 857 while (!GEP->use_empty()) { 858 LoadInst *L = cast<LoadInst>(GEP->use_back()); 859 L->replaceAllUsesWith(TheArg); 860 AA.replaceWithNewValue(L, TheArg); 861 L->eraseFromParent(); 862 } 863 AA.deleteValue(GEP); 864 GEP->eraseFromParent(); 865 } 866 } 867 868 // Increment I2 past all of the arguments added for this promoted pointer. 869 for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i) 870 ++I2; 871 } 872 873 // Notify the alias analysis implementation that we inserted a new argument. 874 if (ExtraArgHack) 875 AA.copyValue(Constant::getNullValue(Type::getInt32Ty(F->getContext())), 876 NF->arg_begin()); 877 878 879 // Tell the alias analysis that the old function is about to disappear. 880 AA.replaceWithNewValue(F, NF); 881 882 883 NF_CGN->stealCalledFunctionsFrom(CG[F]); 884 885 // Now that the old function is dead, delete it. If there is a dangling 886 // reference to the CallgraphNode, just leave the dead function around for 887 // someone else to nuke. 888 CallGraphNode *CGN = CG[F]; 889 if (CGN->getNumReferences() == 0) 890 delete CG.removeFunctionFromModule(CGN); 891 else 892 F->setLinkage(Function::ExternalLinkage); 893 894 return NF_CGN; 895} 896