ArgumentPromotion.cpp revision 7569322765651f19eea0609fb082e6b267d5d2b5
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 72 /// A vector used to hold the indices of a single GEP instruction 73 typedef std::vector<uint64_t> IndicesVector; 74 75 private: 76 CallGraphNode *PromoteArguments(CallGraphNode *CGN); 77 bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const; 78 CallGraphNode *DoPromotion(Function *F, 79 SmallPtrSet<Argument*, 8> &ArgsToPromote, 80 SmallPtrSet<Argument*, 8> &ByValArgsToTransform); 81 /// The maximum number of elements to expand, or 0 for unlimited. 82 unsigned maxElements; 83 }; 84} 85 86char ArgPromotion::ID = 0; 87INITIALIZE_PASS(ArgPromotion, "argpromotion", 88 "Promote 'by reference' arguments to scalars", false, false); 89 90Pass *llvm::createArgumentPromotionPass(unsigned maxElements) { 91 return new ArgPromotion(maxElements); 92} 93 94bool ArgPromotion::runOnSCC(CallGraphSCC &SCC) { 95 bool Changed = false, LocalChange; 96 97 do { // Iterate until we stop promoting from this SCC. 98 LocalChange = false; 99 // Attempt to promote arguments from all functions in this SCC. 100 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) { 101 if (CallGraphNode *CGN = PromoteArguments(*I)) { 102 LocalChange = true; 103 SCC.ReplaceNode(*I, CGN); 104 } 105 } 106 Changed |= LocalChange; // Remember that we changed something. 107 } while (LocalChange); 108 109 return Changed; 110} 111 112/// PromoteArguments - This method checks the specified function to see if there 113/// are any promotable arguments and if it is safe to promote the function (for 114/// example, all callers are direct). If safe to promote some arguments, it 115/// calls the DoPromotion method. 116/// 117CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) { 118 Function *F = CGN->getFunction(); 119 120 // Make sure that it is local to this module. 121 if (!F || !F->hasLocalLinkage()) return 0; 122 123 // First check: see if there are any pointer arguments! If not, quick exit. 124 SmallVector<std::pair<Argument*, unsigned>, 16> PointerArgs; 125 unsigned ArgNo = 0; 126 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 127 I != E; ++I, ++ArgNo) 128 if (I->getType()->isPointerTy()) 129 PointerArgs.push_back(std::pair<Argument*, unsigned>(I, ArgNo)); 130 if (PointerArgs.empty()) return 0; 131 132 // Second check: make sure that all callers are direct callers. We can't 133 // transform functions that have indirect callers. 134 if (F->hasAddressTaken()) 135 return 0; 136 137 // Check to see which arguments are promotable. If an argument is promotable, 138 // add it to ArgsToPromote. 139 SmallPtrSet<Argument*, 8> ArgsToPromote; 140 SmallPtrSet<Argument*, 8> ByValArgsToTransform; 141 for (unsigned i = 0; i != PointerArgs.size(); ++i) { 142 bool isByVal = F->paramHasAttr(PointerArgs[i].second+1, Attribute::ByVal); 143 144 // If this is a byval argument, and if the aggregate type is small, just 145 // pass the elements, which is always safe. 146 Argument *PtrArg = PointerArgs[i].first; 147 if (isByVal) { 148 const Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType(); 149 if (const StructType *STy = dyn_cast<StructType>(AgTy)) { 150 if (maxElements > 0 && STy->getNumElements() > maxElements) { 151 DEBUG(dbgs() << "argpromotion disable promoting argument '" 152 << PtrArg->getName() << "' because it would require adding more" 153 << " than " << maxElements << " arguments to the function.\n"); 154 } else { 155 // If all the elements are single-value types, we can promote it. 156 bool AllSimple = true; 157 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 158 if (!STy->getElementType(i)->isSingleValueType()) { 159 AllSimple = false; 160 break; 161 } 162 163 // Safe to transform, don't even bother trying to "promote" it. 164 // Passing the elements as a scalar will allow scalarrepl to hack on 165 // the new alloca we introduce. 166 if (AllSimple) { 167 ByValArgsToTransform.insert(PtrArg); 168 continue; 169 } 170 } 171 } 172 } 173 174 // Otherwise, see if we can promote the pointer to its value. 175 if (isSafeToPromoteArgument(PtrArg, isByVal)) 176 ArgsToPromote.insert(PtrArg); 177 } 178 179 // No promotable pointer arguments. 180 if (ArgsToPromote.empty() && ByValArgsToTransform.empty()) 181 return 0; 182 183 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform); 184} 185 186/// IsAlwaysValidPointer - Return true if the specified pointer is always legal 187/// to load. 188static bool IsAlwaysValidPointer(Value *V) { 189 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true; 190 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) 191 return IsAlwaysValidPointer(GEP->getOperand(0)); 192 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 193 if (CE->getOpcode() == Instruction::GetElementPtr) 194 return IsAlwaysValidPointer(CE->getOperand(0)); 195 196 return false; 197} 198 199/// AllCalleesPassInValidPointerForArgument - Return true if we can prove that 200/// all callees pass in a valid pointer for the specified function argument. 201static bool AllCalleesPassInValidPointerForArgument(Argument *Arg) { 202 Function *Callee = Arg->getParent(); 203 204 unsigned ArgNo = std::distance(Callee->arg_begin(), 205 Function::arg_iterator(Arg)); 206 207 // Look at all call sites of the function. At this pointer we know we only 208 // have direct callees. 209 for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end(); 210 UI != E; ++UI) { 211 CallSite CS(*UI); 212 assert(CS && "Should only have direct calls!"); 213 214 if (!IsAlwaysValidPointer(CS.getArgument(ArgNo))) 215 return false; 216 } 217 return true; 218} 219 220/// Returns true if Prefix is a prefix of longer. That means, Longer has a size 221/// that is greater than or equal to the size of prefix, and each of the 222/// elements in Prefix is the same as the corresponding elements in Longer. 223/// 224/// This means it also returns true when Prefix and Longer are equal! 225static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix, 226 const ArgPromotion::IndicesVector &Longer) { 227 if (Prefix.size() > Longer.size()) 228 return false; 229 for (unsigned i = 0, e = Prefix.size(); i != e; ++i) 230 if (Prefix[i] != Longer[i]) 231 return false; 232 return true; 233} 234 235 236/// Checks if Indices, or a prefix of Indices, is in Set. 237static bool PrefixIn(const ArgPromotion::IndicesVector &Indices, 238 std::set<ArgPromotion::IndicesVector> &Set) { 239 std::set<ArgPromotion::IndicesVector>::iterator Low; 240 Low = Set.upper_bound(Indices); 241 if (Low != Set.begin()) 242 Low--; 243 // Low is now the last element smaller than or equal to Indices. This means 244 // it points to a prefix of Indices (possibly Indices itself), if such 245 // prefix exists. 246 // 247 // This load is safe if any prefix of its operands is safe to load. 248 return Low != Set.end() && IsPrefix(*Low, Indices); 249} 250 251/// Mark the given indices (ToMark) as safe in the given set of indices 252/// (Safe). Marking safe usually means adding ToMark to Safe. However, if there 253/// is already a prefix of Indices in Safe, Indices are implicitely marked safe 254/// already. Furthermore, any indices that Indices is itself a prefix of, are 255/// removed from Safe (since they are implicitely safe because of Indices now). 256static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark, 257 std::set<ArgPromotion::IndicesVector> &Safe) { 258 std::set<ArgPromotion::IndicesVector>::iterator Low; 259 Low = Safe.upper_bound(ToMark); 260 // Guard against the case where Safe is empty 261 if (Low != Safe.begin()) 262 Low--; 263 // Low is now the last element smaller than or equal to Indices. This 264 // means it points to a prefix of Indices (possibly Indices itself), if 265 // such prefix exists. 266 if (Low != Safe.end()) { 267 if (IsPrefix(*Low, ToMark)) 268 // If there is already a prefix of these indices (or exactly these 269 // indices) marked a safe, don't bother adding these indices 270 return; 271 272 // Increment Low, so we can use it as a "insert before" hint 273 ++Low; 274 } 275 // Insert 276 Low = Safe.insert(Low, ToMark); 277 ++Low; 278 // If there we're a prefix of longer index list(s), remove those 279 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end(); 280 while (Low != End && IsPrefix(ToMark, *Low)) { 281 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low; 282 ++Low; 283 Safe.erase(Remove); 284 } 285} 286 287/// isSafeToPromoteArgument - As you might guess from the name of this method, 288/// it checks to see if it is both safe and useful to promote the argument. 289/// This method limits promotion of aggregates to only promote up to three 290/// elements of the aggregate in order to avoid exploding the number of 291/// arguments passed in. 292bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { 293 typedef std::set<IndicesVector> GEPIndicesSet; 294 295 // Quick exit for unused arguments 296 if (Arg->use_empty()) 297 return true; 298 299 // We can only promote this argument if all of the uses are loads, or are GEP 300 // instructions (with constant indices) that are subsequently loaded. 301 // 302 // Promoting the argument causes it to be loaded in the caller 303 // unconditionally. This is only safe if we can prove that either the load 304 // would have happened in the callee anyway (ie, there is a load in the entry 305 // block) or the pointer passed in at every call site is guaranteed to be 306 // valid. 307 // In the former case, invalid loads can happen, but would have happened 308 // anyway, in the latter case, invalid loads won't happen. This prevents us 309 // from introducing an invalid load that wouldn't have happened in the 310 // original code. 311 // 312 // This set will contain all sets of indices that are loaded in the entry 313 // block, and thus are safe to unconditionally load in the caller. 314 GEPIndicesSet SafeToUnconditionallyLoad; 315 316 // This set contains all the sets of indices that we are planning to promote. 317 // This makes it possible to limit the number of arguments added. 318 GEPIndicesSet ToPromote; 319 320 // If the pointer is always valid, any load with first index 0 is valid. 321 if (isByVal || AllCalleesPassInValidPointerForArgument(Arg)) 322 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0)); 323 324 // First, iterate the entry block and mark loads of (geps of) arguments as 325 // safe. 326 BasicBlock *EntryBlock = Arg->getParent()->begin(); 327 // Declare this here so we can reuse it 328 IndicesVector Indices; 329 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end(); 330 I != E; ++I) 331 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 332 Value *V = LI->getPointerOperand(); 333 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) { 334 V = GEP->getPointerOperand(); 335 if (V == Arg) { 336 // This load actually loads (part of) Arg? Check the indices then. 337 Indices.reserve(GEP->getNumIndices()); 338 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 339 II != IE; ++II) 340 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II)) 341 Indices.push_back(CI->getSExtValue()); 342 else 343 // We found a non-constant GEP index for this argument? Bail out 344 // right away, can't promote this argument at all. 345 return false; 346 347 // Indices checked out, mark them as safe 348 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad); 349 Indices.clear(); 350 } 351 } else if (V == Arg) { 352 // Direct loads are equivalent to a GEP with a single 0 index. 353 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad); 354 } 355 } 356 357 // Now, iterate all uses of the argument to see if there are any uses that are 358 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote. 359 SmallVector<LoadInst*, 16> Loads; 360 IndicesVector Operands; 361 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end(); 362 UI != E; ++UI) { 363 User *U = *UI; 364 Operands.clear(); 365 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 366 if (LI->isVolatile()) return false; // Don't hack volatile loads 367 Loads.push_back(LI); 368 // Direct loads are equivalent to a GEP with a zero index and then a load. 369 Operands.push_back(0); 370 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 371 if (GEP->use_empty()) { 372 // Dead GEP's cause trouble later. Just remove them if we run into 373 // them. 374 getAnalysis<AliasAnalysis>().deleteValue(GEP); 375 GEP->eraseFromParent(); 376 // TODO: This runs the above loop over and over again for dead GEPs 377 // Couldn't we just do increment the UI iterator earlier and erase the 378 // use? 379 return isSafeToPromoteArgument(Arg, isByVal); 380 } 381 382 // Ensure that all of the indices are constants. 383 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end(); 384 i != e; ++i) 385 if (ConstantInt *C = dyn_cast<ConstantInt>(*i)) 386 Operands.push_back(C->getSExtValue()); 387 else 388 return false; // Not a constant operand GEP! 389 390 // Ensure that the only users of the GEP are load instructions. 391 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end(); 392 UI != E; ++UI) 393 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 394 if (LI->isVolatile()) return false; // Don't hack volatile loads 395 Loads.push_back(LI); 396 } else { 397 // Other uses than load? 398 return false; 399 } 400 } else { 401 return false; // Not a load or a GEP. 402 } 403 404 // Now, see if it is safe to promote this load / loads of this GEP. Loading 405 // is safe if Operands, or a prefix of Operands, is marked as safe. 406 if (!PrefixIn(Operands, SafeToUnconditionallyLoad)) 407 return false; 408 409 // See if we are already promoting a load with these indices. If not, check 410 // to make sure that we aren't promoting too many elements. If so, nothing 411 // to do. 412 if (ToPromote.find(Operands) == ToPromote.end()) { 413 if (maxElements > 0 && ToPromote.size() == maxElements) { 414 DEBUG(dbgs() << "argpromotion not promoting argument '" 415 << Arg->getName() << "' because it would require adding more " 416 << "than " << maxElements << " arguments to the function.\n"); 417 // We limit aggregate promotion to only promoting up to a fixed number 418 // of elements of the aggregate. 419 return false; 420 } 421 ToPromote.insert(Operands); 422 } 423 } 424 425 if (Loads.empty()) return true; // No users, this is a dead argument. 426 427 // Okay, now we know that the argument is only used by load instructions and 428 // it is safe to unconditionally perform all of them. Use alias analysis to 429 // check to see if the pointer is guaranteed to not be modified from entry of 430 // the function to each of the load instructions. 431 432 // Because there could be several/many load instructions, remember which 433 // blocks we know to be transparent to the load. 434 SmallPtrSet<BasicBlock*, 16> TranspBlocks; 435 436 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 437 TargetData *TD = getAnalysisIfAvailable<TargetData>(); 438 if (!TD) return false; // Without TargetData, assume the worst. 439 440 for (unsigned i = 0, e = Loads.size(); i != e; ++i) { 441 // Check to see if the load is invalidated from the start of the block to 442 // the load itself. 443 LoadInst *Load = Loads[i]; 444 BasicBlock *BB = Load->getParent(); 445 446 const PointerType *LoadTy = 447 cast<PointerType>(Load->getPointerOperand()->getType()); 448 unsigned LoadSize =(unsigned)TD->getTypeStoreSize(LoadTy->getElementType()); 449 450 if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize)) 451 return false; // Pointer is invalidated! 452 453 // Now check every path from the entry block to the load for transparency. 454 // To do this, we perform a depth first search on the inverse CFG from the 455 // loading block. 456 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 457 BasicBlock *P = *PI; 458 for (idf_ext_iterator<BasicBlock*, SmallPtrSet<BasicBlock*, 16> > 459 I = idf_ext_begin(P, TranspBlocks), 460 E = idf_ext_end(P, TranspBlocks); I != E; ++I) 461 if (AA.canBasicBlockModify(**I, Arg, LoadSize)) 462 return false; 463 } 464 } 465 466 // If the path from the entry of the function to each load is free of 467 // instructions that potentially invalidate the load, we can make the 468 // transformation! 469 return true; 470} 471 472/// DoPromotion - This method actually performs the promotion of the specified 473/// arguments, and returns the new function. At this point, we know that it's 474/// safe to do so. 475CallGraphNode *ArgPromotion::DoPromotion(Function *F, 476 SmallPtrSet<Argument*, 8> &ArgsToPromote, 477 SmallPtrSet<Argument*, 8> &ByValArgsToTransform) { 478 479 // Start by computing a new prototype for the function, which is the same as 480 // the old function, but has modified arguments. 481 const FunctionType *FTy = F->getFunctionType(); 482 std::vector<const Type*> Params; 483 484 typedef std::set<IndicesVector> ScalarizeTable; 485 486 // ScalarizedElements - If we are promoting a pointer that has elements 487 // accessed out of it, keep track of which elements are accessed so that we 488 // can add one argument for each. 489 // 490 // Arguments that are directly loaded will have a zero element value here, to 491 // handle cases where there are both a direct load and GEP accesses. 492 // 493 std::map<Argument*, ScalarizeTable> ScalarizedElements; 494 495 // OriginalLoads - Keep track of a representative load instruction from the 496 // original function so that we can tell the alias analysis implementation 497 // what the new GEP/Load instructions we are inserting look like. 498 std::map<IndicesVector, LoadInst*> OriginalLoads; 499 500 // Attributes - Keep track of the parameter attributes for the arguments 501 // that we are *not* promoting. For the ones that we do promote, the parameter 502 // attributes are lost 503 SmallVector<AttributeWithIndex, 8> AttributesVec; 504 const AttrListPtr &PAL = F->getAttributes(); 505 506 // Add any return attributes. 507 if (Attributes attrs = PAL.getRetAttributes()) 508 AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); 509 510 // First, determine the new argument list 511 unsigned ArgIndex = 1; 512 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; 513 ++I, ++ArgIndex) { 514 if (ByValArgsToTransform.count(I)) { 515 // Simple byval argument? Just add all the struct element types. 516 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 517 const StructType *STy = cast<StructType>(AgTy); 518 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 519 Params.push_back(STy->getElementType(i)); 520 ++NumByValArgsPromoted; 521 } else if (!ArgsToPromote.count(I)) { 522 // Unchanged argument 523 Params.push_back(I->getType()); 524 if (Attributes attrs = PAL.getParamAttributes(ArgIndex)) 525 AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs)); 526 } else if (I->use_empty()) { 527 // Dead argument (which are always marked as promotable) 528 ++NumArgumentsDead; 529 } else { 530 // Okay, this is being promoted. This means that the only uses are loads 531 // or GEPs which are only used by loads 532 533 // In this table, we will track which indices are loaded from the argument 534 // (where direct loads are tracked as no indices). 535 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 536 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 537 ++UI) { 538 Instruction *User = cast<Instruction>(*UI); 539 assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User)); 540 IndicesVector Indices; 541 Indices.reserve(User->getNumOperands() - 1); 542 // Since loads will only have a single operand, and GEPs only a single 543 // non-index operand, this will record direct loads without any indices, 544 // and gep+loads with the GEP indices. 545 for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end(); 546 II != IE; ++II) 547 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue()); 548 // GEPs with a single 0 index can be merged with direct loads 549 if (Indices.size() == 1 && Indices.front() == 0) 550 Indices.clear(); 551 ArgIndices.insert(Indices); 552 LoadInst *OrigLoad; 553 if (LoadInst *L = dyn_cast<LoadInst>(User)) 554 OrigLoad = L; 555 else 556 // Take any load, we will use it only to update Alias Analysis 557 OrigLoad = cast<LoadInst>(User->use_back()); 558 OriginalLoads[Indices] = OrigLoad; 559 } 560 561 // Add a parameter to the function for each element passed in. 562 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 563 E = ArgIndices.end(); SI != E; ++SI) { 564 // not allowed to dereference ->begin() if size() is 0 565 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), 566 SI->begin(), 567 SI->end())); 568 assert(Params.back()); 569 } 570 571 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty()) 572 ++NumArgumentsPromoted; 573 else 574 ++NumAggregatesPromoted; 575 } 576 } 577 578 // Add any function attributes. 579 if (Attributes attrs = PAL.getFnAttributes()) 580 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); 581 582 const Type *RetTy = FTy->getReturnType(); 583 584 // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which 585 // have zero fixed arguments. 586 bool ExtraArgHack = false; 587 if (Params.empty() && FTy->isVarArg()) { 588 ExtraArgHack = true; 589 Params.push_back(Type::getInt32Ty(F->getContext())); 590 } 591 592 // Construct the new function type using the new arguments. 593 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg()); 594 595 // Create the new function body and insert it into the module. 596 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName()); 597 NF->copyAttributesFrom(F); 598 599 600 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n" 601 << "From: " << *F); 602 603 // Recompute the parameter attributes list based on the new arguments for 604 // the function. 605 NF->setAttributes(AttrListPtr::get(AttributesVec.begin(), 606 AttributesVec.end())); 607 AttributesVec.clear(); 608 609 F->getParent()->getFunctionList().insert(F, NF); 610 NF->takeName(F); 611 612 // Get the alias analysis information that we need to update to reflect our 613 // changes. 614 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 615 616 // Get the callgraph information that we need to update to reflect our 617 // changes. 618 CallGraph &CG = getAnalysis<CallGraph>(); 619 620 // Get a new callgraph node for NF. 621 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF); 622 623 // Loop over all of the callers of the function, transforming the call sites 624 // to pass in the loaded pointers. 625 // 626 SmallVector<Value*, 16> Args; 627 while (!F->use_empty()) { 628 CallSite CS(F->use_back()); 629 assert(CS.getCalledFunction() == F); 630 Instruction *Call = CS.getInstruction(); 631 const AttrListPtr &CallPAL = CS.getAttributes(); 632 633 // Add any return attributes. 634 if (Attributes attrs = CallPAL.getRetAttributes()) 635 AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); 636 637 // Loop over the operands, inserting GEP and loads in the caller as 638 // appropriate. 639 CallSite::arg_iterator AI = CS.arg_begin(); 640 ArgIndex = 1; 641 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 642 I != E; ++I, ++AI, ++ArgIndex) 643 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 644 Args.push_back(*AI); // Unmodified argument 645 646 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) 647 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 648 649 } else if (ByValArgsToTransform.count(I)) { 650 // Emit a GEP and load for each element of the struct. 651 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 652 const StructType *STy = cast<StructType>(AgTy); 653 Value *Idxs[2] = { 654 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; 655 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 656 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); 657 Value *Idx = GetElementPtrInst::Create(*AI, Idxs, Idxs+2, 658 (*AI)->getName()+"."+utostr(i), 659 Call); 660 // TODO: Tell AA about the new values? 661 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call)); 662 } 663 } else if (!I->use_empty()) { 664 // Non-dead argument: insert GEPs and loads as appropriate. 665 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 666 // Store the Value* version of the indices in here, but declare it now 667 // for reuse. 668 std::vector<Value*> Ops; 669 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 670 E = ArgIndices.end(); SI != E; ++SI) { 671 Value *V = *AI; 672 LoadInst *OrigLoad = OriginalLoads[*SI]; 673 if (!SI->empty()) { 674 Ops.reserve(SI->size()); 675 const Type *ElTy = V->getType(); 676 for (IndicesVector::const_iterator II = SI->begin(), 677 IE = SI->end(); II != IE; ++II) { 678 // Use i32 to index structs, and i64 for others (pointers/arrays). 679 // This satisfies GEP constraints. 680 const Type *IdxTy = (ElTy->isStructTy() ? 681 Type::getInt32Ty(F->getContext()) : 682 Type::getInt64Ty(F->getContext())); 683 Ops.push_back(ConstantInt::get(IdxTy, *II)); 684 // Keep track of the type we're currently indexing. 685 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II); 686 } 687 // And create a GEP to extract those indices. 688 V = GetElementPtrInst::Create(V, Ops.begin(), Ops.end(), 689 V->getName()+".idx", Call); 690 Ops.clear(); 691 AA.copyValue(OrigLoad->getOperand(0), V); 692 } 693 // Since we're replacing a load make sure we take the alignment 694 // of the previous load. 695 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call); 696 newLoad->setAlignment(OrigLoad->getAlignment()); 697 Args.push_back(newLoad); 698 AA.copyValue(OrigLoad, Args.back()); 699 } 700 } 701 702 if (ExtraArgHack) 703 Args.push_back(Constant::getNullValue(Type::getInt32Ty(F->getContext()))); 704 705 // Push any varargs arguments on the list. 706 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) { 707 Args.push_back(*AI); 708 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) 709 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 710 } 711 712 // Add any function attributes. 713 if (Attributes attrs = CallPAL.getFnAttributes()) 714 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); 715 716 Instruction *New; 717 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 718 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 719 Args.begin(), Args.end(), "", Call); 720 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 721 cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(), 722 AttributesVec.end())); 723 } else { 724 New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call); 725 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 726 cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(), 727 AttributesVec.end())); 728 if (cast<CallInst>(Call)->isTailCall()) 729 cast<CallInst>(New)->setTailCall(); 730 } 731 Args.clear(); 732 AttributesVec.clear(); 733 734 // Update the alias analysis implementation to know that we are replacing 735 // the old call with a new one. 736 AA.replaceWithNewValue(Call, New); 737 738 // Update the callgraph to know that the callsite has been transformed. 739 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()]; 740 CalleeNode->replaceCallEdge(Call, New, NF_CGN); 741 742 if (!Call->use_empty()) { 743 Call->replaceAllUsesWith(New); 744 New->takeName(Call); 745 } 746 747 // Finally, remove the old call from the program, reducing the use-count of 748 // F. 749 Call->eraseFromParent(); 750 } 751 752 // Since we have now created the new function, splice the body of the old 753 // function right into the new function, leaving the old rotting hulk of the 754 // function empty. 755 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 756 757 // Loop over the argument list, transfering uses of the old arguments over to 758 // the new arguments, also transfering over the names as well. 759 // 760 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), 761 I2 = NF->arg_begin(); I != E; ++I) { 762 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 763 // If this is an unmodified argument, move the name and users over to the 764 // new version. 765 I->replaceAllUsesWith(I2); 766 I2->takeName(I); 767 AA.replaceWithNewValue(I, I2); 768 ++I2; 769 continue; 770 } 771 772 if (ByValArgsToTransform.count(I)) { 773 // In the callee, we create an alloca, and store each of the new incoming 774 // arguments into the alloca. 775 Instruction *InsertPt = NF->begin()->begin(); 776 777 // Just add all the struct element types. 778 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 779 Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt); 780 const StructType *STy = cast<StructType>(AgTy); 781 Value *Idxs[2] = { 782 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; 783 784 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 785 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); 786 Value *Idx = 787 GetElementPtrInst::Create(TheAlloca, Idxs, Idxs+2, 788 TheAlloca->getName()+"."+Twine(i), 789 InsertPt); 790 I2->setName(I->getName()+"."+Twine(i)); 791 new StoreInst(I2++, Idx, InsertPt); 792 } 793 794 // Anything that used the arg should now use the alloca. 795 I->replaceAllUsesWith(TheAlloca); 796 TheAlloca->takeName(I); 797 AA.replaceWithNewValue(I, TheAlloca); 798 continue; 799 } 800 801 if (I->use_empty()) { 802 AA.deleteValue(I); 803 continue; 804 } 805 806 // Otherwise, if we promoted this argument, then all users are load 807 // instructions (or GEPs with only load users), and all loads should be 808 // using the new argument that we added. 809 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 810 811 while (!I->use_empty()) { 812 if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) { 813 assert(ArgIndices.begin()->empty() && 814 "Load element should sort to front!"); 815 I2->setName(I->getName()+".val"); 816 LI->replaceAllUsesWith(I2); 817 AA.replaceWithNewValue(LI, I2); 818 LI->eraseFromParent(); 819 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName() 820 << "' in function '" << F->getName() << "'\n"); 821 } else { 822 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back()); 823 IndicesVector Operands; 824 Operands.reserve(GEP->getNumIndices()); 825 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 826 II != IE; ++II) 827 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue()); 828 829 // GEPs with a single 0 index can be merged with direct loads 830 if (Operands.size() == 1 && Operands.front() == 0) 831 Operands.clear(); 832 833 Function::arg_iterator TheArg = I2; 834 for (ScalarizeTable::iterator It = ArgIndices.begin(); 835 *It != Operands; ++It, ++TheArg) { 836 assert(It != ArgIndices.end() && "GEP not handled??"); 837 } 838 839 std::string NewName = I->getName(); 840 for (unsigned i = 0, e = Operands.size(); i != e; ++i) { 841 NewName += "." + utostr(Operands[i]); 842 } 843 NewName += ".val"; 844 TheArg->setName(NewName); 845 846 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName() 847 << "' of function '" << NF->getName() << "'\n"); 848 849 // All of the uses must be load instructions. Replace them all with 850 // the argument specified by ArgNo. 851 while (!GEP->use_empty()) { 852 LoadInst *L = cast<LoadInst>(GEP->use_back()); 853 L->replaceAllUsesWith(TheArg); 854 AA.replaceWithNewValue(L, TheArg); 855 L->eraseFromParent(); 856 } 857 AA.deleteValue(GEP); 858 GEP->eraseFromParent(); 859 } 860 } 861 862 // Increment I2 past all of the arguments added for this promoted pointer. 863 for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i) 864 ++I2; 865 } 866 867 // Notify the alias analysis implementation that we inserted a new argument. 868 if (ExtraArgHack) 869 AA.copyValue(Constant::getNullValue(Type::getInt32Ty(F->getContext())), 870 NF->arg_begin()); 871 872 873 // Tell the alias analysis that the old function is about to disappear. 874 AA.replaceWithNewValue(F, NF); 875 876 877 NF_CGN->stealCalledFunctionsFrom(CG[F]); 878 879 // Now that the old function is dead, delete it. If there is a dangling 880 // reference to the CallgraphNode, just leave the dead function around for 881 // someone else to nuke. 882 CallGraphNode *CGN = CG[F]; 883 if (CGN->getNumReferences() == 0) 884 delete CG.removeFunctionFromModule(CGN); 885 else 886 F->setLinkage(Function::ExternalLinkage); 887 888 return NF_CGN; 889} 890