ArgumentPromotion.cpp revision 6d8eb156e6be727570b300bac7712f745a318c7d
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> 50using namespace llvm; 51 52STATISTIC(NumArgumentsPromoted , "Number of pointer arguments promoted"); 53STATISTIC(NumAggregatesPromoted, "Number of aggregate arguments promoted"); 54STATISTIC(NumByValArgsPromoted , "Number of byval arguments promoted"); 55STATISTIC(NumArgumentsDead , "Number of dead pointer args eliminated"); 56 57namespace { 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 87char ArgPromotion::ID = 0; 88INITIALIZE_PASS_BEGIN(ArgPromotion, "argpromotion", 89 "Promote 'by reference' arguments to scalars", false, false) 90INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 91INITIALIZE_AG_DEPENDENCY(CallGraph) 92INITIALIZE_PASS_END(ArgPromotion, "argpromotion", 93 "Promote 'by reference' arguments to scalars", false, false) 94 95Pass *llvm::createArgumentPromotionPass(unsigned maxElements) { 96 return new ArgPromotion(maxElements); 97} 98 99bool 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/// 122CallGraphNode *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. 139 if (F->hasAddressTaken()) 140 return 0; 141 142 // Check to see which arguments are promotable. If an argument is promotable, 143 // add it to ArgsToPromote. 144 SmallPtrSet<Argument*, 8> ArgsToPromote; 145 SmallPtrSet<Argument*, 8> ByValArgsToTransform; 146 for (unsigned i = 0; i != PointerArgs.size(); ++i) { 147 bool isByVal = F->paramHasAttr(PointerArgs[i].second+1, Attribute::ByVal); 148 149 // If this is a byval argument, and if the aggregate type is small, just 150 // pass the elements, which is always safe. 151 Argument *PtrArg = PointerArgs[i].first; 152 if (isByVal) { 153 const Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType(); 154 if (const StructType *STy = dyn_cast<StructType>(AgTy)) { 155 if (maxElements > 0 && STy->getNumElements() > maxElements) { 156 DEBUG(dbgs() << "argpromotion disable promoting argument '" 157 << PtrArg->getName() << "' because it would require adding more" 158 << " than " << maxElements << " arguments to the function.\n"); 159 } else { 160 // If all the elements are single-value types, we can promote it. 161 bool AllSimple = true; 162 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 163 if (!STy->getElementType(i)->isSingleValueType()) { 164 AllSimple = false; 165 break; 166 } 167 168 // Safe to transform, don't even bother trying to "promote" it. 169 // Passing the elements as a scalar will allow scalarrepl to hack on 170 // the new alloca we introduce. 171 if (AllSimple) { 172 ByValArgsToTransform.insert(PtrArg); 173 continue; 174 } 175 } 176 } 177 } 178 179 // Otherwise, see if we can promote the pointer to its value. 180 if (isSafeToPromoteArgument(PtrArg, isByVal)) 181 ArgsToPromote.insert(PtrArg); 182 } 183 184 // No promotable pointer arguments. 185 if (ArgsToPromote.empty() && ByValArgsToTransform.empty()) 186 return 0; 187 188 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform); 189} 190 191/// AllCalleesPassInValidPointerForArgument - Return true if we can prove that 192/// all callees pass in a valid pointer for the specified function argument. 193static bool AllCalleesPassInValidPointerForArgument(Argument *Arg) { 194 Function *Callee = Arg->getParent(); 195 196 unsigned ArgNo = std::distance(Callee->arg_begin(), 197 Function::arg_iterator(Arg)); 198 199 // Look at all call sites of the function. At this pointer we know we only 200 // have direct callees. 201 for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end(); 202 UI != E; ++UI) { 203 CallSite CS(*UI); 204 assert(CS && "Should only have direct calls!"); 205 206 if (!CS.getArgument(ArgNo)->isDereferenceablePointer()) 207 return false; 208 } 209 return true; 210} 211 212/// Returns true if Prefix is a prefix of longer. That means, Longer has a size 213/// that is greater than or equal to the size of prefix, and each of the 214/// elements in Prefix is the same as the corresponding elements in Longer. 215/// 216/// This means it also returns true when Prefix and Longer are equal! 217static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix, 218 const ArgPromotion::IndicesVector &Longer) { 219 if (Prefix.size() > Longer.size()) 220 return false; 221 for (unsigned i = 0, e = Prefix.size(); i != e; ++i) 222 if (Prefix[i] != Longer[i]) 223 return false; 224 return true; 225} 226 227 228/// Checks if Indices, or a prefix of Indices, is in Set. 229static bool PrefixIn(const ArgPromotion::IndicesVector &Indices, 230 std::set<ArgPromotion::IndicesVector> &Set) { 231 std::set<ArgPromotion::IndicesVector>::iterator Low; 232 Low = Set.upper_bound(Indices); 233 if (Low != Set.begin()) 234 Low--; 235 // Low is now the last element smaller than or equal to Indices. This means 236 // it points to a prefix of Indices (possibly Indices itself), if such 237 // prefix exists. 238 // 239 // This load is safe if any prefix of its operands is safe to load. 240 return Low != Set.end() && IsPrefix(*Low, Indices); 241} 242 243/// Mark the given indices (ToMark) as safe in the given set of indices 244/// (Safe). Marking safe usually means adding ToMark to Safe. However, if there 245/// is already a prefix of Indices in Safe, Indices are implicitely marked safe 246/// already. Furthermore, any indices that Indices is itself a prefix of, are 247/// removed from Safe (since they are implicitely safe because of Indices now). 248static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark, 249 std::set<ArgPromotion::IndicesVector> &Safe) { 250 std::set<ArgPromotion::IndicesVector>::iterator Low; 251 Low = Safe.upper_bound(ToMark); 252 // Guard against the case where Safe is empty 253 if (Low != Safe.begin()) 254 Low--; 255 // Low is now the last element smaller than or equal to Indices. This 256 // means it points to a prefix of Indices (possibly Indices itself), if 257 // such prefix exists. 258 if (Low != Safe.end()) { 259 if (IsPrefix(*Low, ToMark)) 260 // If there is already a prefix of these indices (or exactly these 261 // indices) marked a safe, don't bother adding these indices 262 return; 263 264 // Increment Low, so we can use it as a "insert before" hint 265 ++Low; 266 } 267 // Insert 268 Low = Safe.insert(Low, ToMark); 269 ++Low; 270 // If there we're a prefix of longer index list(s), remove those 271 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end(); 272 while (Low != End && IsPrefix(ToMark, *Low)) { 273 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low; 274 ++Low; 275 Safe.erase(Remove); 276 } 277} 278 279/// isSafeToPromoteArgument - As you might guess from the name of this method, 280/// it checks to see if it is both safe and useful to promote the argument. 281/// This method limits promotion of aggregates to only promote up to three 282/// elements of the aggregate in order to avoid exploding the number of 283/// arguments passed in. 284bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { 285 typedef std::set<IndicesVector> GEPIndicesSet; 286 287 // Quick exit for unused arguments 288 if (Arg->use_empty()) 289 return true; 290 291 // We can only promote this argument if all of the uses are loads, or are GEP 292 // instructions (with constant indices) that are subsequently loaded. 293 // 294 // Promoting the argument causes it to be loaded in the caller 295 // unconditionally. This is only safe if we can prove that either the load 296 // would have happened in the callee anyway (ie, there is a load in the entry 297 // block) or the pointer passed in at every call site is guaranteed to be 298 // valid. 299 // In the former case, invalid loads can happen, but would have happened 300 // anyway, in the latter case, invalid loads won't happen. This prevents us 301 // from introducing an invalid load that wouldn't have happened in the 302 // original code. 303 // 304 // This set will contain all sets of indices that are loaded in the entry 305 // block, and thus are safe to unconditionally load in the caller. 306 GEPIndicesSet SafeToUnconditionallyLoad; 307 308 // This set contains all the sets of indices that we are planning to promote. 309 // This makes it possible to limit the number of arguments added. 310 GEPIndicesSet ToPromote; 311 312 // If the pointer is always valid, any load with first index 0 is valid. 313 if (isByVal || AllCalleesPassInValidPointerForArgument(Arg)) 314 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0)); 315 316 // First, iterate the entry block and mark loads of (geps of) arguments as 317 // safe. 318 BasicBlock *EntryBlock = Arg->getParent()->begin(); 319 // Declare this here so we can reuse it 320 IndicesVector Indices; 321 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end(); 322 I != E; ++I) 323 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 324 Value *V = LI->getPointerOperand(); 325 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) { 326 V = GEP->getPointerOperand(); 327 if (V == Arg) { 328 // This load actually loads (part of) Arg? Check the indices then. 329 Indices.reserve(GEP->getNumIndices()); 330 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 331 II != IE; ++II) 332 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II)) 333 Indices.push_back(CI->getSExtValue()); 334 else 335 // We found a non-constant GEP index for this argument? Bail out 336 // right away, can't promote this argument at all. 337 return false; 338 339 // Indices checked out, mark them as safe 340 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad); 341 Indices.clear(); 342 } 343 } else if (V == Arg) { 344 // Direct loads are equivalent to a GEP with a single 0 index. 345 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad); 346 } 347 } 348 349 // Now, iterate all uses of the argument to see if there are any uses that are 350 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote. 351 SmallVector<LoadInst*, 16> Loads; 352 IndicesVector Operands; 353 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end(); 354 UI != E; ++UI) { 355 User *U = *UI; 356 Operands.clear(); 357 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 358 if (LI->isVolatile()) return false; // Don't hack volatile loads 359 Loads.push_back(LI); 360 // Direct loads are equivalent to a GEP with a zero index and then a load. 361 Operands.push_back(0); 362 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 363 if (GEP->use_empty()) { 364 // Dead GEP's cause trouble later. Just remove them if we run into 365 // them. 366 getAnalysis<AliasAnalysis>().deleteValue(GEP); 367 GEP->eraseFromParent(); 368 // TODO: This runs the above loop over and over again for dead GEPs 369 // Couldn't we just do increment the UI iterator earlier and erase the 370 // use? 371 return isSafeToPromoteArgument(Arg, isByVal); 372 } 373 374 // Ensure that all of the indices are constants. 375 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end(); 376 i != e; ++i) 377 if (ConstantInt *C = dyn_cast<ConstantInt>(*i)) 378 Operands.push_back(C->getSExtValue()); 379 else 380 return false; // Not a constant operand GEP! 381 382 // Ensure that the only users of the GEP are load instructions. 383 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end(); 384 UI != E; ++UI) 385 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 386 if (LI->isVolatile()) return false; // Don't hack volatile loads 387 Loads.push_back(LI); 388 } else { 389 // Other uses than load? 390 return false; 391 } 392 } else { 393 return false; // Not a load or a GEP. 394 } 395 396 // Now, see if it is safe to promote this load / loads of this GEP. Loading 397 // is safe if Operands, or a prefix of Operands, is marked as safe. 398 if (!PrefixIn(Operands, SafeToUnconditionallyLoad)) 399 return false; 400 401 // See if we are already promoting a load with these indices. If not, check 402 // to make sure that we aren't promoting too many elements. If so, nothing 403 // to do. 404 if (ToPromote.find(Operands) == ToPromote.end()) { 405 if (maxElements > 0 && ToPromote.size() == maxElements) { 406 DEBUG(dbgs() << "argpromotion not promoting argument '" 407 << Arg->getName() << "' because it would require adding more " 408 << "than " << maxElements << " arguments to the function.\n"); 409 // We limit aggregate promotion to only promoting up to a fixed number 410 // of elements of the aggregate. 411 return false; 412 } 413 ToPromote.insert(Operands); 414 } 415 } 416 417 if (Loads.empty()) return true; // No users, this is a dead argument. 418 419 // Okay, now we know that the argument is only used by load instructions and 420 // it is safe to unconditionally perform all of them. Use alias analysis to 421 // check to see if the pointer is guaranteed to not be modified from entry of 422 // the function to each of the load instructions. 423 424 // Because there could be several/many load instructions, remember which 425 // blocks we know to be transparent to the load. 426 SmallPtrSet<BasicBlock*, 16> TranspBlocks; 427 428 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 429 430 for (unsigned i = 0, e = Loads.size(); i != e; ++i) { 431 // Check to see if the load is invalidated from the start of the block to 432 // the load itself. 433 LoadInst *Load = Loads[i]; 434 BasicBlock *BB = Load->getParent(); 435 436 AliasAnalysis::Location Loc = AA.getLocation(Load); 437 if (AA.canInstructionRangeModify(BB->front(), *Load, Loc)) 438 return false; // Pointer is invalidated! 439 440 // Now check every path from the entry block to the load for transparency. 441 // To do this, we perform a depth first search on the inverse CFG from the 442 // loading block. 443 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 444 BasicBlock *P = *PI; 445 for (idf_ext_iterator<BasicBlock*, SmallPtrSet<BasicBlock*, 16> > 446 I = idf_ext_begin(P, TranspBlocks), 447 E = idf_ext_end(P, TranspBlocks); I != E; ++I) 448 if (AA.canBasicBlockModify(**I, Loc)) 449 return false; 450 } 451 } 452 453 // If the path from the entry of the function to each load is free of 454 // instructions that potentially invalidate the load, we can make the 455 // transformation! 456 return true; 457} 458 459/// DoPromotion - This method actually performs the promotion of the specified 460/// arguments, and returns the new function. At this point, we know that it's 461/// safe to do so. 462CallGraphNode *ArgPromotion::DoPromotion(Function *F, 463 SmallPtrSet<Argument*, 8> &ArgsToPromote, 464 SmallPtrSet<Argument*, 8> &ByValArgsToTransform) { 465 466 // Start by computing a new prototype for the function, which is the same as 467 // the old function, but has modified arguments. 468 const FunctionType *FTy = F->getFunctionType(); 469 std::vector<const Type*> Params; 470 471 typedef std::set<IndicesVector> ScalarizeTable; 472 473 // ScalarizedElements - If we are promoting a pointer that has elements 474 // accessed out of it, keep track of which elements are accessed so that we 475 // can add one argument for each. 476 // 477 // Arguments that are directly loaded will have a zero element value here, to 478 // handle cases where there are both a direct load and GEP accesses. 479 // 480 std::map<Argument*, ScalarizeTable> ScalarizedElements; 481 482 // OriginalLoads - Keep track of a representative load instruction from the 483 // original function so that we can tell the alias analysis implementation 484 // what the new GEP/Load instructions we are inserting look like. 485 std::map<IndicesVector, LoadInst*> OriginalLoads; 486 487 // Attributes - Keep track of the parameter attributes for the arguments 488 // that we are *not* promoting. For the ones that we do promote, the parameter 489 // attributes are lost 490 SmallVector<AttributeWithIndex, 8> AttributesVec; 491 const AttrListPtr &PAL = F->getAttributes(); 492 493 // Add any return attributes. 494 if (Attributes attrs = PAL.getRetAttributes()) 495 AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); 496 497 // First, determine the new argument list 498 unsigned ArgIndex = 1; 499 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; 500 ++I, ++ArgIndex) { 501 if (ByValArgsToTransform.count(I)) { 502 // Simple byval argument? Just add all the struct element types. 503 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 504 const StructType *STy = cast<StructType>(AgTy); 505 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 506 Params.push_back(STy->getElementType(i)); 507 ++NumByValArgsPromoted; 508 } else if (!ArgsToPromote.count(I)) { 509 // Unchanged argument 510 Params.push_back(I->getType()); 511 if (Attributes attrs = PAL.getParamAttributes(ArgIndex)) 512 AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs)); 513 } else if (I->use_empty()) { 514 // Dead argument (which are always marked as promotable) 515 ++NumArgumentsDead; 516 } else { 517 // Okay, this is being promoted. This means that the only uses are loads 518 // or GEPs which are only used by loads 519 520 // In this table, we will track which indices are loaded from the argument 521 // (where direct loads are tracked as no indices). 522 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 523 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 524 ++UI) { 525 Instruction *User = cast<Instruction>(*UI); 526 assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User)); 527 IndicesVector Indices; 528 Indices.reserve(User->getNumOperands() - 1); 529 // Since loads will only have a single operand, and GEPs only a single 530 // non-index operand, this will record direct loads without any indices, 531 // and gep+loads with the GEP indices. 532 for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end(); 533 II != IE; ++II) 534 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue()); 535 // GEPs with a single 0 index can be merged with direct loads 536 if (Indices.size() == 1 && Indices.front() == 0) 537 Indices.clear(); 538 ArgIndices.insert(Indices); 539 LoadInst *OrigLoad; 540 if (LoadInst *L = dyn_cast<LoadInst>(User)) 541 OrigLoad = L; 542 else 543 // Take any load, we will use it only to update Alias Analysis 544 OrigLoad = cast<LoadInst>(User->use_back()); 545 OriginalLoads[Indices] = OrigLoad; 546 } 547 548 // Add a parameter to the function for each element passed in. 549 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 550 E = ArgIndices.end(); SI != E; ++SI) { 551 // not allowed to dereference ->begin() if size() is 0 552 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), 553 SI->begin(), 554 SI->end())); 555 assert(Params.back()); 556 } 557 558 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty()) 559 ++NumArgumentsPromoted; 560 else 561 ++NumAggregatesPromoted; 562 } 563 } 564 565 // Add any function attributes. 566 if (Attributes attrs = PAL.getFnAttributes()) 567 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); 568 569 const Type *RetTy = FTy->getReturnType(); 570 571 // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which 572 // have zero fixed arguments. 573 bool ExtraArgHack = false; 574 if (Params.empty() && FTy->isVarArg()) { 575 ExtraArgHack = true; 576 Params.push_back(Type::getInt32Ty(F->getContext())); 577 } 578 579 // Construct the new function type using the new arguments. 580 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg()); 581 582 // Create the new function body and insert it into the module. 583 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName()); 584 NF->copyAttributesFrom(F); 585 586 587 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n" 588 << "From: " << *F); 589 590 // Recompute the parameter attributes list based on the new arguments for 591 // the function. 592 NF->setAttributes(AttrListPtr::get(AttributesVec.begin(), 593 AttributesVec.end())); 594 AttributesVec.clear(); 595 596 F->getParent()->getFunctionList().insert(F, NF); 597 NF->takeName(F); 598 599 // Get the alias analysis information that we need to update to reflect our 600 // changes. 601 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 602 603 // Get the callgraph information that we need to update to reflect our 604 // changes. 605 CallGraph &CG = getAnalysis<CallGraph>(); 606 607 // Get a new callgraph node for NF. 608 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF); 609 610 // Loop over all of the callers of the function, transforming the call sites 611 // to pass in the loaded pointers. 612 // 613 SmallVector<Value*, 16> Args; 614 while (!F->use_empty()) { 615 CallSite CS(F->use_back()); 616 assert(CS.getCalledFunction() == F); 617 Instruction *Call = CS.getInstruction(); 618 const AttrListPtr &CallPAL = CS.getAttributes(); 619 620 // Add any return attributes. 621 if (Attributes attrs = CallPAL.getRetAttributes()) 622 AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); 623 624 // Loop over the operands, inserting GEP and loads in the caller as 625 // appropriate. 626 CallSite::arg_iterator AI = CS.arg_begin(); 627 ArgIndex = 1; 628 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 629 I != E; ++I, ++AI, ++ArgIndex) 630 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 631 Args.push_back(*AI); // Unmodified argument 632 633 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) 634 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 635 636 } else if (ByValArgsToTransform.count(I)) { 637 // Emit a GEP and load for each element of the struct. 638 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 639 const StructType *STy = cast<StructType>(AgTy); 640 Value *Idxs[2] = { 641 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; 642 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 643 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); 644 Value *Idx = GetElementPtrInst::Create(*AI, Idxs, Idxs+2, 645 (*AI)->getName()+"."+utostr(i), 646 Call); 647 // TODO: Tell AA about the new values? 648 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call)); 649 } 650 } else if (!I->use_empty()) { 651 // Non-dead argument: insert GEPs and loads as appropriate. 652 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 653 // Store the Value* version of the indices in here, but declare it now 654 // for reuse. 655 std::vector<Value*> Ops; 656 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 657 E = ArgIndices.end(); SI != E; ++SI) { 658 Value *V = *AI; 659 LoadInst *OrigLoad = OriginalLoads[*SI]; 660 if (!SI->empty()) { 661 Ops.reserve(SI->size()); 662 const Type *ElTy = V->getType(); 663 for (IndicesVector::const_iterator II = SI->begin(), 664 IE = SI->end(); II != IE; ++II) { 665 // Use i32 to index structs, and i64 for others (pointers/arrays). 666 // This satisfies GEP constraints. 667 const Type *IdxTy = (ElTy->isStructTy() ? 668 Type::getInt32Ty(F->getContext()) : 669 Type::getInt64Ty(F->getContext())); 670 Ops.push_back(ConstantInt::get(IdxTy, *II)); 671 // Keep track of the type we're currently indexing. 672 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II); 673 } 674 // And create a GEP to extract those indices. 675 V = GetElementPtrInst::Create(V, Ops.begin(), Ops.end(), 676 V->getName()+".idx", Call); 677 Ops.clear(); 678 AA.copyValue(OrigLoad->getOperand(0), V); 679 } 680 // Since we're replacing a load make sure we take the alignment 681 // of the previous load. 682 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call); 683 newLoad->setAlignment(OrigLoad->getAlignment()); 684 // Transfer the TBAA info too. 685 newLoad->setMetadata(LLVMContext::MD_tbaa, 686 OrigLoad->getMetadata(LLVMContext::MD_tbaa)); 687 Args.push_back(newLoad); 688 AA.copyValue(OrigLoad, Args.back()); 689 } 690 } 691 692 if (ExtraArgHack) 693 Args.push_back(Constant::getNullValue(Type::getInt32Ty(F->getContext()))); 694 695 // Push any varargs arguments on the list. 696 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) { 697 Args.push_back(*AI); 698 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) 699 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 700 } 701 702 // Add any function attributes. 703 if (Attributes attrs = CallPAL.getFnAttributes()) 704 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); 705 706 Instruction *New; 707 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 708 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 709 Args.begin(), Args.end(), "", Call); 710 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 711 cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(), 712 AttributesVec.end())); 713 } else { 714 New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call); 715 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 716 cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(), 717 AttributesVec.end())); 718 if (cast<CallInst>(Call)->isTailCall()) 719 cast<CallInst>(New)->setTailCall(); 720 } 721 Args.clear(); 722 AttributesVec.clear(); 723 724 // Update the alias analysis implementation to know that we are replacing 725 // the old call with a new one. 726 AA.replaceWithNewValue(Call, New); 727 728 // Update the callgraph to know that the callsite has been transformed. 729 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()]; 730 CalleeNode->replaceCallEdge(Call, New, NF_CGN); 731 732 if (!Call->use_empty()) { 733 Call->replaceAllUsesWith(New); 734 New->takeName(Call); 735 } 736 737 // Finally, remove the old call from the program, reducing the use-count of 738 // F. 739 Call->eraseFromParent(); 740 } 741 742 // Since we have now created the new function, splice the body of the old 743 // function right into the new function, leaving the old rotting hulk of the 744 // function empty. 745 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 746 747 // Loop over the argument list, transfering uses of the old arguments over to 748 // the new arguments, also transfering over the names as well. 749 // 750 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), 751 I2 = NF->arg_begin(); I != E; ++I) { 752 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 753 // If this is an unmodified argument, move the name and users over to the 754 // new version. 755 I->replaceAllUsesWith(I2); 756 I2->takeName(I); 757 AA.replaceWithNewValue(I, I2); 758 ++I2; 759 continue; 760 } 761 762 if (ByValArgsToTransform.count(I)) { 763 // In the callee, we create an alloca, and store each of the new incoming 764 // arguments into the alloca. 765 Instruction *InsertPt = NF->begin()->begin(); 766 767 // Just add all the struct element types. 768 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 769 Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt); 770 const StructType *STy = cast<StructType>(AgTy); 771 Value *Idxs[2] = { 772 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; 773 774 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 775 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); 776 Value *Idx = 777 GetElementPtrInst::Create(TheAlloca, Idxs, Idxs+2, 778 TheAlloca->getName()+"."+Twine(i), 779 InsertPt); 780 I2->setName(I->getName()+"."+Twine(i)); 781 new StoreInst(I2++, Idx, InsertPt); 782 } 783 784 // Anything that used the arg should now use the alloca. 785 I->replaceAllUsesWith(TheAlloca); 786 TheAlloca->takeName(I); 787 AA.replaceWithNewValue(I, TheAlloca); 788 continue; 789 } 790 791 if (I->use_empty()) { 792 AA.deleteValue(I); 793 continue; 794 } 795 796 // Otherwise, if we promoted this argument, then all users are load 797 // instructions (or GEPs with only load users), and all loads should be 798 // using the new argument that we added. 799 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 800 801 while (!I->use_empty()) { 802 if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) { 803 assert(ArgIndices.begin()->empty() && 804 "Load element should sort to front!"); 805 I2->setName(I->getName()+".val"); 806 LI->replaceAllUsesWith(I2); 807 AA.replaceWithNewValue(LI, I2); 808 LI->eraseFromParent(); 809 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName() 810 << "' in function '" << F->getName() << "'\n"); 811 } else { 812 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back()); 813 IndicesVector Operands; 814 Operands.reserve(GEP->getNumIndices()); 815 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 816 II != IE; ++II) 817 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue()); 818 819 // GEPs with a single 0 index can be merged with direct loads 820 if (Operands.size() == 1 && Operands.front() == 0) 821 Operands.clear(); 822 823 Function::arg_iterator TheArg = I2; 824 for (ScalarizeTable::iterator It = ArgIndices.begin(); 825 *It != Operands; ++It, ++TheArg) { 826 assert(It != ArgIndices.end() && "GEP not handled??"); 827 } 828 829 std::string NewName = I->getName(); 830 for (unsigned i = 0, e = Operands.size(); i != e; ++i) { 831 NewName += "." + utostr(Operands[i]); 832 } 833 NewName += ".val"; 834 TheArg->setName(NewName); 835 836 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName() 837 << "' of function '" << NF->getName() << "'\n"); 838 839 // All of the uses must be load instructions. Replace them all with 840 // the argument specified by ArgNo. 841 while (!GEP->use_empty()) { 842 LoadInst *L = cast<LoadInst>(GEP->use_back()); 843 L->replaceAllUsesWith(TheArg); 844 AA.replaceWithNewValue(L, TheArg); 845 L->eraseFromParent(); 846 } 847 AA.deleteValue(GEP); 848 GEP->eraseFromParent(); 849 } 850 } 851 852 // Increment I2 past all of the arguments added for this promoted pointer. 853 for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i) 854 ++I2; 855 } 856 857 // Notify the alias analysis implementation that we inserted a new argument. 858 if (ExtraArgHack) 859 AA.copyValue(Constant::getNullValue(Type::getInt32Ty(F->getContext())), 860 NF->arg_begin()); 861 862 863 // Tell the alias analysis that the old function is about to disappear. 864 AA.replaceWithNewValue(F, NF); 865 866 867 NF_CGN->stealCalledFunctionsFrom(CG[F]); 868 869 // Now that the old function is dead, delete it. If there is a dangling 870 // reference to the CallgraphNode, just leave the dead function around for 871 // someone else to nuke. 872 CallGraphNode *CGN = CG[F]; 873 if (CGN->getNumReferences() == 0) 874 delete CG.removeFunctionFromModule(CGN); 875 else 876 F->setLinkage(Function::ExternalLinkage); 877 878 return NF_CGN; 879} 880