ArgumentPromotion.cpp revision 757068f3bad425fb126fe16ab7b8a82a636e6bbd
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/Analysis/AliasAnalysis.h" 40#include "llvm/Analysis/CallGraph.h" 41#include "llvm/Target/TargetData.h" 42#include "llvm/Support/CallSite.h" 43#include "llvm/Support/CFG.h" 44#include "llvm/Support/Debug.h" 45#include "llvm/ADT/DepthFirstIterator.h" 46#include "llvm/ADT/Statistic.h" 47#include "llvm/ADT/StringExtras.h" 48#include "llvm/Support/Compiler.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 VISIBILITY_HIDDEN ArgPromotion : public CallGraphSCCPass { 61 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 62 AU.addRequired<AliasAnalysis>(); 63 AU.addRequired<TargetData>(); 64 CallGraphSCCPass::getAnalysisUsage(AU); 65 } 66 67 virtual bool runOnSCC(const std::vector<CallGraphNode *> &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 bool PromoteArguments(CallGraphNode *CGN); 77 bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const; 78 Function *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; 87static RegisterPass<ArgPromotion> 88X("argpromotion", "Promote 'by reference' arguments to scalars"); 89 90Pass *llvm::createArgumentPromotionPass(unsigned maxElements) { 91 return new ArgPromotion(maxElements); 92} 93 94bool ArgPromotion::runOnSCC(const std::vector<CallGraphNode *> &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 (unsigned i = 0, e = SCC.size(); i != e; ++i) 101 LocalChange |= PromoteArguments(SCC[i]); 102 Changed |= LocalChange; // Remember that we changed something. 103 } while (LocalChange); 104 105 return Changed; 106} 107 108/// PromoteArguments - This method checks the specified function to see if there 109/// are any promotable arguments and if it is safe to promote the function (for 110/// example, all callers are direct). If safe to promote some arguments, it 111/// calls the DoPromotion method. 112/// 113bool ArgPromotion::PromoteArguments(CallGraphNode *CGN) { 114 Function *F = CGN->getFunction(); 115 116 // Make sure that it is local to this module. 117 if (!F || !F->hasLocalLinkage()) return false; 118 119 // First check: see if there are any pointer arguments! If not, quick exit. 120 SmallVector<std::pair<Argument*, unsigned>, 16> PointerArgs; 121 unsigned ArgNo = 0; 122 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 123 I != E; ++I, ++ArgNo) 124 if (isa<PointerType>(I->getType())) 125 PointerArgs.push_back(std::pair<Argument*, unsigned>(I, ArgNo)); 126 if (PointerArgs.empty()) return false; 127 128 // Second check: make sure that all callers are direct callers. We can't 129 // transform functions that have indirect callers. 130 if (F->hasAddressTaken()) 131 return false; 132 133 // Check to see which arguments are promotable. If an argument is promotable, 134 // add it to ArgsToPromote. 135 SmallPtrSet<Argument*, 8> ArgsToPromote; 136 SmallPtrSet<Argument*, 8> ByValArgsToTransform; 137 for (unsigned i = 0; i != PointerArgs.size(); ++i) { 138 bool isByVal = F->paramHasAttr(PointerArgs[i].second+1, Attribute::ByVal); 139 140 // If this is a byval argument, and if the aggregate type is small, just 141 // pass the elements, which is always safe. 142 Argument *PtrArg = PointerArgs[i].first; 143 if (isByVal) { 144 const Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType(); 145 if (const StructType *STy = dyn_cast<StructType>(AgTy)) { 146 if (maxElements > 0 && STy->getNumElements() > maxElements) { 147 DOUT << "argpromotion disable promoting argument '" 148 << PtrArg->getName() << "' because it would require adding more " 149 << "than " << maxElements << " arguments to the function.\n"; 150 } else { 151 // If all the elements are single-value types, we can promote it. 152 bool AllSimple = true; 153 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 154 if (!STy->getElementType(i)->isSingleValueType()) { 155 AllSimple = false; 156 break; 157 } 158 159 // Safe to transform, don't even bother trying to "promote" it. 160 // Passing the elements as a scalar will allow scalarrepl to hack on 161 // the new alloca we introduce. 162 if (AllSimple) { 163 ByValArgsToTransform.insert(PtrArg); 164 continue; 165 } 166 } 167 } 168 } 169 170 // Otherwise, see if we can promote the pointer to its value. 171 if (isSafeToPromoteArgument(PtrArg, isByVal)) 172 ArgsToPromote.insert(PtrArg); 173 } 174 175 // No promotable pointer arguments. 176 if (ArgsToPromote.empty() && ByValArgsToTransform.empty()) return false; 177 178 Function *NewF = DoPromotion(F, ArgsToPromote, ByValArgsToTransform); 179 180 // Update the call graph to know that the function has been transformed. 181 getAnalysis<CallGraph>().changeFunction(F, NewF); 182 return true; 183} 184 185/// IsAlwaysValidPointer - Return true if the specified pointer is always legal 186/// to load. 187static bool IsAlwaysValidPointer(Value *V) { 188 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true; 189 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) 190 return IsAlwaysValidPointer(GEP->getOperand(0)); 191 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 192 if (CE->getOpcode() == Instruction::GetElementPtr) 193 return IsAlwaysValidPointer(CE->getOperand(0)); 194 195 return false; 196} 197 198/// AllCalleesPassInValidPointerForArgument - Return true if we can prove that 199/// all callees pass in a valid pointer for the specified function argument. 200static bool AllCalleesPassInValidPointerForArgument(Argument *Arg) { 201 Function *Callee = Arg->getParent(); 202 203 unsigned ArgNo = std::distance(Callee->arg_begin(), 204 Function::arg_iterator(Arg)); 205 206 // Look at all call sites of the function. At this pointer we know we only 207 // have direct callees. 208 for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end(); 209 UI != E; ++UI) { 210 CallSite CS = CallSite::get(*UI); 211 assert(CS.getInstruction() && "Should only have direct calls!"); 212 213 if (!IsAlwaysValidPointer(CS.getArgument(ArgNo))) 214 return false; 215 } 216 return true; 217} 218 219/// Returns true if Prefix is a prefix of longer. That means, Longer has a size 220/// that is greater than or equal to the size of prefix, and each of the 221/// elements in Prefix is the same as the corresponding elements in Longer. 222/// 223/// This means it also returns true when Prefix and Longer are equal! 224static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix, 225 const ArgPromotion::IndicesVector &Longer) { 226 if (Prefix.size() > Longer.size()) 227 return false; 228 for (unsigned i = 0, e = Prefix.size(); i != e; ++i) 229 if (Prefix[i] != Longer[i]) 230 return false; 231 return true; 232} 233 234 235/// Checks if Indices, or a prefix of Indices, is in Set. 236static bool PrefixIn(const ArgPromotion::IndicesVector &Indices, 237 std::set<ArgPromotion::IndicesVector> &Set) { 238 std::set<ArgPromotion::IndicesVector>::iterator Low; 239 Low = Set.upper_bound(Indices); 240 if (Low != Set.begin()) 241 Low--; 242 // Low is now the last element smaller than or equal to Indices. This means 243 // it points to a prefix of Indices (possibly Indices itself), if such 244 // prefix exists. 245 // 246 // This load is safe if any prefix of its operands is safe to load. 247 return Low != Set.end() && IsPrefix(*Low, Indices); 248} 249 250/// Mark the given indices (ToMark) as safe in the the given set of indices 251/// (Safe). Marking safe usually means adding ToMark to Safe. However, if there 252/// is already a prefix of Indices in Safe, Indices are implicitely marked safe 253/// already. Furthermore, any indices that Indices is itself a prefix of, are 254/// removed from Safe (since they are implicitely safe because of Indices now). 255static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark, 256 std::set<ArgPromotion::IndicesVector> &Safe) { 257 std::set<ArgPromotion::IndicesVector>::iterator Low; 258 Low = Safe.upper_bound(ToMark); 259 // Guard against the case where Safe is empty 260 if (Low != Safe.begin()) 261 Low--; 262 // Low is now the last element smaller than or equal to Indices. This 263 // means it points to a prefix of Indices (possibly Indices itself), if 264 // such prefix exists. 265 if (Low != Safe.end()) { 266 if (IsPrefix(*Low, ToMark)) 267 // If there is already a prefix of these indices (or exactly these 268 // indices) marked a safe, don't bother adding these indices 269 return; 270 271 // Increment Low, so we can use it as a "insert before" hint 272 ++Low; 273 } 274 // Insert 275 Low = Safe.insert(Low, ToMark); 276 ++Low; 277 // If there we're a prefix of longer index list(s), remove those 278 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end(); 279 while (Low != End && IsPrefix(ToMark, *Low)) { 280 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low; 281 ++Low; 282 Safe.erase(Remove); 283 } 284} 285 286/// isSafeToPromoteArgument - As you might guess from the name of this method, 287/// it checks to see if it is both safe and useful to promote the argument. 288/// This method limits promotion of aggregates to only promote up to three 289/// elements of the aggregate in order to avoid exploding the number of 290/// arguments passed in. 291bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { 292 typedef std::set<IndicesVector> GEPIndicesSet; 293 294 // Quick exit for unused arguments 295 if (Arg->use_empty()) 296 return true; 297 298 // We can only promote this argument if all of the uses are loads, or are GEP 299 // instructions (with constant indices) that are subsequently loaded. 300 // 301 // Promoting the argument causes it to be loaded in the caller 302 // unconditionally. This is only safe if we can prove that either the load 303 // would have happened in the callee anyway (ie, there is a load in the entry 304 // block) or the pointer passed in at every call site is guaranteed to be 305 // valid. 306 // In the former case, invalid loads can happen, but would have happened 307 // anyway, in the latter case, invalid loads won't happen. This prevents us 308 // from introducing an invalid load that wouldn't have happened in the 309 // original code. 310 // 311 // This set will contain all sets of indices that are loaded in the entry 312 // block, and thus are safe to unconditionally load in the caller. 313 GEPIndicesSet SafeToUnconditionallyLoad; 314 315 // This set contains all the sets of indices that we are planning to promote. 316 // This makes it possible to limit the number of arguments added. 317 GEPIndicesSet ToPromote; 318 319 // If the pointer is always valid, any load with first index 0 is valid. 320 if(isByVal || AllCalleesPassInValidPointerForArgument(Arg)) 321 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0)); 322 323 // First, iterate the entry block and mark loads of (geps of) arguments as 324 // safe. 325 BasicBlock *EntryBlock = Arg->getParent()->begin(); 326 // Declare this here so we can reuse it 327 IndicesVector Indices; 328 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end(); 329 I != E; ++I) 330 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 331 Value *V = LI->getPointerOperand(); 332 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) { 333 V = GEP->getPointerOperand(); 334 if (V == Arg) { 335 // This load actually loads (part of) Arg? Check the indices then. 336 Indices.reserve(GEP->getNumIndices()); 337 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 338 II != IE; ++II) 339 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II)) 340 Indices.push_back(CI->getSExtValue()); 341 else 342 // We found a non-constant GEP index for this argument? Bail out 343 // right away, can't promote this argument at all. 344 return false; 345 346 // Indices checked out, mark them as safe 347 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad); 348 Indices.clear(); 349 } 350 } else if (V == Arg) { 351 // Direct loads are equivalent to a GEP with a single 0 index. 352 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad); 353 } 354 } 355 356 // Now, iterate all uses of the argument to see if there are any uses that are 357 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote. 358 SmallVector<LoadInst*, 16> Loads; 359 IndicesVector Operands; 360 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end(); 361 UI != E; ++UI) { 362 Operands.clear(); 363 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 364 if (LI->isVolatile()) return false; // Don't hack volatile loads 365 Loads.push_back(LI); 366 // Direct loads are equivalent to a GEP with a zero index and then a load. 367 Operands.push_back(0); 368 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 369 if (GEP->use_empty()) { 370 // Dead GEP's cause trouble later. Just remove them if we run into 371 // them. 372 getAnalysis<AliasAnalysis>().deleteValue(GEP); 373 GEP->eraseFromParent(); 374 // TODO: This runs the above loop over and over again for dead GEPS 375 // Couldn't we just do increment the UI iterator earlier and erase the 376 // use? 377 return isSafeToPromoteArgument(Arg, isByVal); 378 } 379 380 // Ensure that all of the indices are constants. 381 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end(); 382 i != e; ++i) 383 if (ConstantInt *C = dyn_cast<ConstantInt>(*i)) 384 Operands.push_back(C->getSExtValue()); 385 else 386 return false; // Not a constant operand GEP! 387 388 // Ensure that the only users of the GEP are load instructions. 389 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end(); 390 UI != E; ++UI) 391 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 392 if (LI->isVolatile()) return false; // Don't hack volatile loads 393 Loads.push_back(LI); 394 } else { 395 // Other uses than load? 396 return false; 397 } 398 } else { 399 return false; // Not a load or a GEP. 400 } 401 402 // Now, see if it is safe to promote this load / loads of this GEP. Loading 403 // is safe if Operands, or a prefix of Operands, is marked as safe. 404 if (!PrefixIn(Operands, SafeToUnconditionallyLoad)) 405 return false; 406 407 // See if we are already promoting a load with these indices. If not, check 408 // to make sure that we aren't promoting too many elements. If so, nothing 409 // to do. 410 if (ToPromote.find(Operands) == ToPromote.end()) { 411 if (maxElements > 0 && ToPromote.size() == maxElements) { 412 DOUT << "argpromotion not promoting argument '" 413 << Arg->getName() << "' because it would require adding more " 414 << "than " << maxElements << " arguments to the function.\n"; 415 // We limit aggregate promotion to only promoting up to a fixed number 416 // of elements of the aggregate. 417 return false; 418 } 419 ToPromote.insert(Operands); 420 } 421 } 422 423 if (Loads.empty()) return true; // No users, this is a dead argument. 424 425 // Okay, now we know that the argument is only used by load instructions and 426 // it is safe to unconditionally perform all of them. Use alias analysis to 427 // check to see if the pointer is guaranteed to not be modified from entry of 428 // the function to each of the load instructions. 429 430 // Because there could be several/many load instructions, remember which 431 // blocks we know to be transparent to the load. 432 SmallPtrSet<BasicBlock*, 16> TranspBlocks; 433 434 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 435 TargetData &TD = getAnalysis<TargetData>(); 436 437 for (unsigned i = 0, e = Loads.size(); i != e; ++i) { 438 // Check to see if the load is invalidated from the start of the block to 439 // the load itself. 440 LoadInst *Load = Loads[i]; 441 BasicBlock *BB = Load->getParent(); 442 443 const PointerType *LoadTy = 444 cast<PointerType>(Load->getPointerOperand()->getType()); 445 unsigned LoadSize = (unsigned)TD.getTypeStoreSize(LoadTy->getElementType()); 446 447 if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize)) 448 return false; // Pointer is invalidated! 449 450 // Now check every path from the entry block to the load for transparency. 451 // To do this, we perform a depth first search on the inverse CFG from the 452 // loading block. 453 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 454 for (idf_ext_iterator<BasicBlock*, SmallPtrSet<BasicBlock*, 16> > 455 I = idf_ext_begin(*PI, TranspBlocks), 456 E = idf_ext_end(*PI, TranspBlocks); I != E; ++I) 457 if (AA.canBasicBlockModify(**I, Arg, LoadSize)) 458 return false; 459 } 460 461 // If the path from the entry of the function to each load is free of 462 // instructions that potentially invalidate the load, we can make the 463 // transformation! 464 return true; 465} 466 467/// DoPromotion - This method actually performs the promotion of the specified 468/// arguments, and returns the new function. At this point, we know that it's 469/// safe to do so. 470Function *ArgPromotion::DoPromotion(Function *F, 471 SmallPtrSet<Argument*, 8> &ArgsToPromote, 472 SmallPtrSet<Argument*, 8> &ByValArgsToTransform) { 473 474 // Start by computing a new prototype for the function, which is the same as 475 // the old function, but has modified arguments. 476 const FunctionType *FTy = F->getFunctionType(); 477 std::vector<const Type*> Params; 478 479 typedef std::set<IndicesVector> ScalarizeTable; 480 481 // ScalarizedElements - If we are promoting a pointer that has elements 482 // accessed out of it, keep track of which elements are accessed so that we 483 // can add one argument for each. 484 // 485 // Arguments that are directly loaded will have a zero element value here, to 486 // handle cases where there are both a direct load and GEP accesses. 487 // 488 std::map<Argument*, ScalarizeTable> ScalarizedElements; 489 490 // OriginalLoads - Keep track of a representative load instruction from the 491 // original function so that we can tell the alias analysis implementation 492 // what the new GEP/Load instructions we are inserting look like. 493 std::map<IndicesVector, LoadInst*> OriginalLoads; 494 495 // Attributes - Keep track of the parameter attributes for the arguments 496 // that we are *not* promoting. For the ones that we do promote, the parameter 497 // attributes are lost 498 SmallVector<AttributeWithIndex, 8> AttributesVec; 499 const AttrListPtr &PAL = F->getAttributes(); 500 501 // Add any return attributes. 502 if (Attributes attrs = PAL.getRetAttributes()) 503 AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); 504 505 // First, determine the new argument list 506 unsigned ArgIndex = 1; 507 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; 508 ++I, ++ArgIndex) { 509 if (ByValArgsToTransform.count(I)) { 510 // Simple byval argument? Just add all the struct element types. 511 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 512 const StructType *STy = cast<StructType>(AgTy); 513 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 514 Params.push_back(STy->getElementType(i)); 515 ++NumByValArgsPromoted; 516 } else if (!ArgsToPromote.count(I)) { 517 // Unchanged argument 518 Params.push_back(I->getType()); 519 if (Attributes attrs = PAL.getParamAttributes(ArgIndex)) 520 AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs)); 521 } else if (I->use_empty()) { 522 // Dead argument (which are always marked as promotable) 523 ++NumArgumentsDead; 524 } else { 525 // Okay, this is being promoted. This means that the only uses are loads 526 // or GEPs which are only used by loads 527 528 // In this table, we will track which indices are loaded from the argument 529 // (where direct loads are tracked as no indices). 530 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 531 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 532 ++UI) { 533 Instruction *User = cast<Instruction>(*UI); 534 assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User)); 535 IndicesVector Indices; 536 Indices.reserve(User->getNumOperands() - 1); 537 // Since loads will only have a single operand, and GEPs only a single 538 // non-index operand, this will record direct loads without any indices, 539 // and gep+loads with the GEP indices. 540 for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end(); 541 II != IE; ++II) 542 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue()); 543 // GEPs with a single 0 index can be merged with direct loads 544 if (Indices.size() == 1 && Indices.front() == 0) 545 Indices.clear(); 546 ArgIndices.insert(Indices); 547 LoadInst *OrigLoad; 548 if (LoadInst *L = dyn_cast<LoadInst>(User)) 549 OrigLoad = L; 550 else 551 // Take any load, we will use it only to update Alias Analysis 552 OrigLoad = cast<LoadInst>(User->use_back()); 553 OriginalLoads[Indices] = OrigLoad; 554 } 555 556 // Add a parameter to the function for each element passed in. 557 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 558 E = ArgIndices.end(); SI != E; ++SI) { 559 // not allowed to dereference ->begin() if size() is 0 560 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), 561 SI->begin(), 562 SI->end())); 563 assert(Params.back()); 564 } 565 566 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty()) 567 ++NumArgumentsPromoted; 568 else 569 ++NumAggregatesPromoted; 570 } 571 } 572 573 // Add any function attributes. 574 if (Attributes attrs = PAL.getFnAttributes()) 575 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); 576 577 const Type *RetTy = FTy->getReturnType(); 578 579 // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which 580 // have zero fixed arguments. 581 bool ExtraArgHack = false; 582 if (Params.empty() && FTy->isVarArg()) { 583 ExtraArgHack = true; 584 Params.push_back(Type::Int32Ty); 585 } 586 587 // Construct the new function type using the new arguments. 588 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg()); 589 590 // Create the new function body and insert it into the module... 591 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName()); 592 NF->copyAttributesFrom(F); 593 594 // Recompute the parameter attributes list based on the new arguments for 595 // the function. 596 NF->setAttributes(AttrListPtr::get(AttributesVec.begin(), AttributesVec.end())); 597 AttributesVec.clear(); 598 599 F->getParent()->getFunctionList().insert(F, NF); 600 NF->takeName(F); 601 602 // Get the alias analysis information that we need to update to reflect our 603 // changes. 604 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 605 606 // Get the callgraph information that we need to update to reflect our 607 // changes. 608 CallGraph &CG = getAnalysis<CallGraph>(); 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 = CallSite::get(F->use_back()); 616 Instruction *Call = CS.getInstruction(); 617 const AttrListPtr &CallPAL = CS.getAttributes(); 618 619 // Add any return attributes. 620 if (Attributes attrs = CallPAL.getRetAttributes()) 621 AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); 622 623 // Loop over the operands, inserting GEP and loads in the caller as 624 // appropriate. 625 CallSite::arg_iterator AI = CS.arg_begin(); 626 ArgIndex = 1; 627 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 628 I != E; ++I, ++AI, ++ArgIndex) 629 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 630 Args.push_back(*AI); // Unmodified argument 631 632 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) 633 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 634 635 } else if (ByValArgsToTransform.count(I)) { 636 // Emit a GEP and load for each element of the struct. 637 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 638 const StructType *STy = cast<StructType>(AgTy); 639 Value *Idxs[2] = { ConstantInt::get(Type::Int32Ty, 0), 0 }; 640 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 641 Idxs[1] = ConstantInt::get(Type::Int32Ty, i); 642 Value *Idx = GetElementPtrInst::Create(*AI, Idxs, Idxs+2, 643 (*AI)->getName()+"."+utostr(i), 644 Call); 645 // TODO: Tell AA about the new values? 646 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call)); 647 } 648 } else if (!I->use_empty()) { 649 // Non-dead argument: insert GEPs and loads as appropriate. 650 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 651 // Store the Value* version of the indices in here, but declare it now 652 // for reuse 653 std::vector<Value*> Ops; 654 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 655 E = ArgIndices.end(); SI != E; ++SI) { 656 Value *V = *AI; 657 LoadInst *OrigLoad = OriginalLoads[*SI]; 658 if (!SI->empty()) { 659 Ops.reserve(SI->size()); 660 const Type *ElTy = V->getType(); 661 for (IndicesVector::const_iterator II = SI->begin(), 662 IE = SI->end(); II != IE; ++II) { 663 // Use i32 to index structs, and i64 for others (pointers/arrays). 664 // This satisfies GEP constraints. 665 const Type *IdxTy = (isa<StructType>(ElTy) ? Type::Int32Ty : Type::Int64Ty); 666 Ops.push_back(ConstantInt::get(IdxTy, *II)); 667 // Keep track of the type we're currently indexing 668 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II); 669 } 670 // And create a GEP to extract those indices 671 V = GetElementPtrInst::Create(V, Ops.begin(), Ops.end(), 672 V->getName()+".idx", Call); 673 Ops.clear(); 674 AA.copyValue(OrigLoad->getOperand(0), V); 675 } 676 Args.push_back(new LoadInst(V, V->getName()+".val", Call)); 677 AA.copyValue(OrigLoad, Args.back()); 678 } 679 } 680 681 if (ExtraArgHack) 682 Args.push_back(Constant::getNullValue(Type::Int32Ty)); 683 684 // Push any varargs arguments on the list 685 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) { 686 Args.push_back(*AI); 687 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) 688 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 689 } 690 691 // Add any function attributes. 692 if (Attributes attrs = CallPAL.getFnAttributes()) 693 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); 694 695 Instruction *New; 696 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 697 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 698 Args.begin(), Args.end(), "", Call); 699 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 700 cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(), 701 AttributesVec.end())); 702 } else { 703 New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call); 704 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 705 cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(), 706 AttributesVec.end())); 707 if (cast<CallInst>(Call)->isTailCall()) 708 cast<CallInst>(New)->setTailCall(); 709 } 710 Args.clear(); 711 AttributesVec.clear(); 712 713 // Update the alias analysis implementation to know that we are replacing 714 // the old call with a new one. 715 AA.replaceWithNewValue(Call, New); 716 717 // Update the callgraph to know that the callsite has been transformed. 718 CG[Call->getParent()->getParent()]->replaceCallSite(Call, New); 719 720 if (!Call->use_empty()) { 721 Call->replaceAllUsesWith(New); 722 New->takeName(Call); 723 } 724 725 // Finally, remove the old call from the program, reducing the use-count of 726 // F. 727 Call->eraseFromParent(); 728 } 729 730 // Since we have now created the new function, splice the body of the old 731 // function right into the new function, leaving the old rotting hulk of the 732 // function empty. 733 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 734 735 // Loop over the argument list, transfering uses of the old arguments over to 736 // the new arguments, also transfering over the names as well. 737 // 738 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), 739 I2 = NF->arg_begin(); I != E; ++I) { 740 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 741 // If this is an unmodified argument, move the name and users over to the 742 // new version. 743 I->replaceAllUsesWith(I2); 744 I2->takeName(I); 745 AA.replaceWithNewValue(I, I2); 746 ++I2; 747 continue; 748 } 749 750 if (ByValArgsToTransform.count(I)) { 751 // In the callee, we create an alloca, and store each of the new incoming 752 // arguments into the alloca. 753 Instruction *InsertPt = NF->begin()->begin(); 754 755 // Just add all the struct element types. 756 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 757 Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt); 758 const StructType *STy = cast<StructType>(AgTy); 759 Value *Idxs[2] = { ConstantInt::get(Type::Int32Ty, 0), 0 }; 760 761 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 762 Idxs[1] = ConstantInt::get(Type::Int32Ty, i); 763 std::string Name = TheAlloca->getName()+"."+utostr(i); 764 Value *Idx = GetElementPtrInst::Create(TheAlloca, Idxs, Idxs+2, 765 Name, InsertPt); 766 I2->setName(I->getName()+"."+utostr(i)); 767 new StoreInst(I2++, Idx, InsertPt); 768 } 769 770 // Anything that used the arg should now use the alloca. 771 I->replaceAllUsesWith(TheAlloca); 772 TheAlloca->takeName(I); 773 AA.replaceWithNewValue(I, TheAlloca); 774 continue; 775 } 776 777 if (I->use_empty()) { 778 AA.deleteValue(I); 779 continue; 780 } 781 782 // Otherwise, if we promoted this argument, then all users are load 783 // instructions (or GEPs with only load users), and all loads should be 784 // using the new argument that we added. 785 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 786 787 while (!I->use_empty()) { 788 if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) { 789 assert(ArgIndices.begin()->empty() && 790 "Load element should sort to front!"); 791 I2->setName(I->getName()+".val"); 792 LI->replaceAllUsesWith(I2); 793 AA.replaceWithNewValue(LI, I2); 794 LI->eraseFromParent(); 795 DOUT << "*** Promoted load of argument '" << I->getName() 796 << "' in function '" << F->getName() << "'\n"; 797 } else { 798 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back()); 799 IndicesVector Operands; 800 Operands.reserve(GEP->getNumIndices()); 801 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 802 II != IE; ++II) 803 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue()); 804 805 // GEPs with a single 0 index can be merged with direct loads 806 if (Operands.size() == 1 && Operands.front() == 0) 807 Operands.clear(); 808 809 Function::arg_iterator TheArg = I2; 810 for (ScalarizeTable::iterator It = ArgIndices.begin(); 811 *It != Operands; ++It, ++TheArg) { 812 assert(It != ArgIndices.end() && "GEP not handled??"); 813 } 814 815 std::string NewName = I->getName(); 816 for (unsigned i = 0, e = Operands.size(); i != e; ++i) { 817 NewName += "." + utostr(Operands[i]); 818 } 819 NewName += ".val"; 820 TheArg->setName(NewName); 821 822 DOUT << "*** Promoted agg argument '" << TheArg->getName() 823 << "' of function '" << NF->getName() << "'\n"; 824 825 // All of the uses must be load instructions. Replace them all with 826 // the argument specified by ArgNo. 827 while (!GEP->use_empty()) { 828 LoadInst *L = cast<LoadInst>(GEP->use_back()); 829 L->replaceAllUsesWith(TheArg); 830 AA.replaceWithNewValue(L, TheArg); 831 L->eraseFromParent(); 832 } 833 AA.deleteValue(GEP); 834 GEP->eraseFromParent(); 835 } 836 } 837 838 // Increment I2 past all of the arguments added for this promoted pointer. 839 for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i) 840 ++I2; 841 } 842 843 // Notify the alias analysis implementation that we inserted a new argument. 844 if (ExtraArgHack) 845 AA.copyValue(Constant::getNullValue(Type::Int32Ty), NF->arg_begin()); 846 847 848 // Tell the alias analysis that the old function is about to disappear. 849 AA.replaceWithNewValue(F, NF); 850 851 // Now that the old function is dead, delete it. 852 F->eraseFromParent(); 853 return NF; 854} 855