ArgumentPromotion.cpp revision 15676ac89f4f0c3e10108f267ddad19149ad9cb5
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/Compiler.h" 45#include "llvm/Support/CFG.h" 46#include "llvm/Support/Debug.h" 47#include "llvm/Support/raw_ostream.h" 48#include "llvm/ADT/DepthFirstIterator.h" 49#include "llvm/ADT/Statistic.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 VISIBILITY_HIDDEN ArgPromotion : public CallGraphSCCPass { 62 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 63 AU.addRequired<AliasAnalysis>(); 64 AU.addRequired<TargetData>(); 65 CallGraphSCCPass::getAnalysisUsage(AU); 66 } 67 68 virtual bool runOnSCC(const std::vector<CallGraphNode *> &SCC); 69 static char ID; // Pass identification, replacement for typeid 70 explicit ArgPromotion(unsigned maxElements = 3) 71 : CallGraphSCCPass(&ID), maxElements(maxElements) {} 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 bool PromoteArguments(CallGraphNode *CGN); 78 bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const; 79 Function *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; 88static RegisterPass<ArgPromotion> 89X("argpromotion", "Promote 'by reference' arguments to scalars"); 90 91Pass *llvm::createArgumentPromotionPass(unsigned maxElements) { 92 return new ArgPromotion(maxElements); 93} 94 95bool ArgPromotion::runOnSCC(const std::vector<CallGraphNode *> &SCC) { 96 bool Changed = false, LocalChange; 97 98 do { // Iterate until we stop promoting from this SCC. 99 LocalChange = false; 100 // Attempt to promote arguments from all functions in this SCC. 101 for (unsigned i = 0, e = SCC.size(); i != e; ++i) 102 LocalChange |= PromoteArguments(SCC[i]); 103 Changed |= LocalChange; // Remember that we changed something. 104 } while (LocalChange); 105 106 return Changed; 107} 108 109/// PromoteArguments - This method checks the specified function to see if there 110/// are any promotable arguments and if it is safe to promote the function (for 111/// example, all callers are direct). If safe to promote some arguments, it 112/// calls the DoPromotion method. 113/// 114bool ArgPromotion::PromoteArguments(CallGraphNode *CGN) { 115 Function *F = CGN->getFunction(); 116 117 // Make sure that it is local to this module. 118 if (!F || !F->hasLocalLinkage()) return false; 119 120 // First check: see if there are any pointer arguments! If not, quick exit. 121 SmallVector<std::pair<Argument*, unsigned>, 16> PointerArgs; 122 unsigned ArgNo = 0; 123 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 124 I != E; ++I, ++ArgNo) 125 if (isa<PointerType>(I->getType())) 126 PointerArgs.push_back(std::pair<Argument*, unsigned>(I, ArgNo)); 127 if (PointerArgs.empty()) return false; 128 129 // Second check: make sure that all callers are direct callers. We can't 130 // transform functions that have indirect callers. 131 if (F->hasAddressTaken()) 132 return false; 133 134 // Check to see which arguments are promotable. If an argument is promotable, 135 // add it to ArgsToPromote. 136 SmallPtrSet<Argument*, 8> ArgsToPromote; 137 SmallPtrSet<Argument*, 8> ByValArgsToTransform; 138 for (unsigned i = 0; i != PointerArgs.size(); ++i) { 139 bool isByVal = F->paramHasAttr(PointerArgs[i].second+1, Attribute::ByVal); 140 141 // If this is a byval argument, and if the aggregate type is small, just 142 // pass the elements, which is always safe. 143 Argument *PtrArg = PointerArgs[i].first; 144 if (isByVal) { 145 const Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType(); 146 if (const StructType *STy = dyn_cast<StructType>(AgTy)) { 147 if (maxElements > 0 && STy->getNumElements() > maxElements) { 148 DEBUG(errs() << "argpromotion disable promoting argument '" 149 << PtrArg->getName() << "' because it would require adding more" 150 << " than " << maxElements << " arguments to the function.\n"); 151 } else { 152 // If all the elements are single-value types, we can promote it. 153 bool AllSimple = true; 154 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 155 if (!STy->getElementType(i)->isSingleValueType()) { 156 AllSimple = false; 157 break; 158 } 159 160 // Safe to transform, don't even bother trying to "promote" it. 161 // Passing the elements as a scalar will allow scalarrepl to hack on 162 // the new alloca we introduce. 163 if (AllSimple) { 164 ByValArgsToTransform.insert(PtrArg); 165 continue; 166 } 167 } 168 } 169 } 170 171 // Otherwise, see if we can promote the pointer to its value. 172 if (isSafeToPromoteArgument(PtrArg, isByVal)) 173 ArgsToPromote.insert(PtrArg); 174 } 175 176 // No promotable pointer arguments. 177 if (ArgsToPromote.empty() && ByValArgsToTransform.empty()) return false; 178 179 Function *NewF = DoPromotion(F, ArgsToPromote, ByValArgsToTransform); 180 181 // Update the call graph to know that the function has been transformed. 182 getAnalysis<CallGraph>().changeFunction(F, NewF); 183 return true; 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 = CallSite::get(*UI); 212 assert(CS.getInstruction() && "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 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 Operands.clear(); 364 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 365 if (LI->isVolatile()) return false; // Don't hack volatile loads 366 Loads.push_back(LI); 367 // Direct loads are equivalent to a GEP with a zero index and then a load. 368 Operands.push_back(0); 369 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 370 if (GEP->use_empty()) { 371 // Dead GEP's cause trouble later. Just remove them if we run into 372 // them. 373 getAnalysis<AliasAnalysis>().deleteValue(GEP); 374 GEP->eraseFromParent(); 375 // TODO: This runs the above loop over and over again for dead GEPS 376 // Couldn't we just do increment the UI iterator earlier and erase the 377 // use? 378 return isSafeToPromoteArgument(Arg, isByVal); 379 } 380 381 // Ensure that all of the indices are constants. 382 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end(); 383 i != e; ++i) 384 if (ConstantInt *C = dyn_cast<ConstantInt>(*i)) 385 Operands.push_back(C->getSExtValue()); 386 else 387 return false; // Not a constant operand GEP! 388 389 // Ensure that the only users of the GEP are load instructions. 390 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end(); 391 UI != E; ++UI) 392 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 393 if (LI->isVolatile()) return false; // Don't hack volatile loads 394 Loads.push_back(LI); 395 } else { 396 // Other uses than load? 397 return false; 398 } 399 } else { 400 return false; // Not a load or a GEP. 401 } 402 403 // Now, see if it is safe to promote this load / loads of this GEP. Loading 404 // is safe if Operands, or a prefix of Operands, is marked as safe. 405 if (!PrefixIn(Operands, SafeToUnconditionallyLoad)) 406 return false; 407 408 // See if we are already promoting a load with these indices. If not, check 409 // to make sure that we aren't promoting too many elements. If so, nothing 410 // to do. 411 if (ToPromote.find(Operands) == ToPromote.end()) { 412 if (maxElements > 0 && ToPromote.size() == maxElements) { 413 DEBUG(errs() << "argpromotion not promoting argument '" 414 << Arg->getName() << "' because it would require adding more " 415 << "than " << maxElements << " arguments to the function.\n"); 416 // We limit aggregate promotion to only promoting up to a fixed number 417 // of elements of the aggregate. 418 return false; 419 } 420 ToPromote.insert(Operands); 421 } 422 } 423 424 if (Loads.empty()) return true; // No users, this is a dead argument. 425 426 // Okay, now we know that the argument is only used by load instructions and 427 // it is safe to unconditionally perform all of them. Use alias analysis to 428 // check to see if the pointer is guaranteed to not be modified from entry of 429 // the function to each of the load instructions. 430 431 // Because there could be several/many load instructions, remember which 432 // blocks we know to be transparent to the load. 433 SmallPtrSet<BasicBlock*, 16> TranspBlocks; 434 435 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 436 TargetData &TD = getAnalysis<TargetData>(); 437 438 for (unsigned i = 0, e = Loads.size(); i != e; ++i) { 439 // Check to see if the load is invalidated from the start of the block to 440 // the load itself. 441 LoadInst *Load = Loads[i]; 442 BasicBlock *BB = Load->getParent(); 443 444 const PointerType *LoadTy = 445 cast<PointerType>(Load->getPointerOperand()->getType()); 446 unsigned LoadSize = (unsigned)TD.getTypeStoreSize(LoadTy->getElementType()); 447 448 if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize)) 449 return false; // Pointer is invalidated! 450 451 // Now check every path from the entry block to the load for transparency. 452 // To do this, we perform a depth first search on the inverse CFG from the 453 // loading block. 454 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 455 for (idf_ext_iterator<BasicBlock*, SmallPtrSet<BasicBlock*, 16> > 456 I = idf_ext_begin(*PI, TranspBlocks), 457 E = idf_ext_end(*PI, TranspBlocks); I != E; ++I) 458 if (AA.canBasicBlockModify(**I, Arg, LoadSize)) 459 return false; 460 } 461 462 // If the path from the entry of the function to each load is free of 463 // instructions that potentially invalidate the load, we can make the 464 // transformation! 465 return true; 466} 467 468/// DoPromotion - This method actually performs the promotion of the specified 469/// arguments, and returns the new function. At this point, we know that it's 470/// safe to do so. 471Function *ArgPromotion::DoPromotion(Function *F, 472 SmallPtrSet<Argument*, 8> &ArgsToPromote, 473 SmallPtrSet<Argument*, 8> &ByValArgsToTransform) { 474 475 // Start by computing a new prototype for the function, which is the same as 476 // the old function, but has modified arguments. 477 const FunctionType *FTy = F->getFunctionType(); 478 std::vector<const Type*> Params; 479 480 typedef std::set<IndicesVector> ScalarizeTable; 481 482 // ScalarizedElements - If we are promoting a pointer that has elements 483 // accessed out of it, keep track of which elements are accessed so that we 484 // can add one argument for each. 485 // 486 // Arguments that are directly loaded will have a zero element value here, to 487 // handle cases where there are both a direct load and GEP accesses. 488 // 489 std::map<Argument*, ScalarizeTable> ScalarizedElements; 490 491 // OriginalLoads - Keep track of a representative load instruction from the 492 // original function so that we can tell the alias analysis implementation 493 // what the new GEP/Load instructions we are inserting look like. 494 std::map<IndicesVector, LoadInst*> OriginalLoads; 495 496 // Attributes - Keep track of the parameter attributes for the arguments 497 // that we are *not* promoting. For the ones that we do promote, the parameter 498 // attributes are lost 499 SmallVector<AttributeWithIndex, 8> AttributesVec; 500 const AttrListPtr &PAL = F->getAttributes(); 501 502 // Add any return attributes. 503 if (Attributes attrs = PAL.getRetAttributes()) 504 AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); 505 506 // First, determine the new argument list 507 unsigned ArgIndex = 1; 508 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; 509 ++I, ++ArgIndex) { 510 if (ByValArgsToTransform.count(I)) { 511 // Simple byval argument? Just add all the struct element types. 512 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 513 const StructType *STy = cast<StructType>(AgTy); 514 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 515 Params.push_back(STy->getElementType(i)); 516 ++NumByValArgsPromoted; 517 } else if (!ArgsToPromote.count(I)) { 518 // Unchanged argument 519 Params.push_back(I->getType()); 520 if (Attributes attrs = PAL.getParamAttributes(ArgIndex)) 521 AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs)); 522 } else if (I->use_empty()) { 523 // Dead argument (which are always marked as promotable) 524 ++NumArgumentsDead; 525 } else { 526 // Okay, this is being promoted. This means that the only uses are loads 527 // or GEPs which are only used by loads 528 529 // In this table, we will track which indices are loaded from the argument 530 // (where direct loads are tracked as no indices). 531 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 532 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 533 ++UI) { 534 Instruction *User = cast<Instruction>(*UI); 535 assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User)); 536 IndicesVector Indices; 537 Indices.reserve(User->getNumOperands() - 1); 538 // Since loads will only have a single operand, and GEPs only a single 539 // non-index operand, this will record direct loads without any indices, 540 // and gep+loads with the GEP indices. 541 for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end(); 542 II != IE; ++II) 543 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue()); 544 // GEPs with a single 0 index can be merged with direct loads 545 if (Indices.size() == 1 && Indices.front() == 0) 546 Indices.clear(); 547 ArgIndices.insert(Indices); 548 LoadInst *OrigLoad; 549 if (LoadInst *L = dyn_cast<LoadInst>(User)) 550 OrigLoad = L; 551 else 552 // Take any load, we will use it only to update Alias Analysis 553 OrigLoad = cast<LoadInst>(User->use_back()); 554 OriginalLoads[Indices] = OrigLoad; 555 } 556 557 // Add a parameter to the function for each element passed in. 558 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 559 E = ArgIndices.end(); SI != E; ++SI) { 560 // not allowed to dereference ->begin() if size() is 0 561 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), 562 SI->begin(), 563 SI->end())); 564 assert(Params.back()); 565 } 566 567 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty()) 568 ++NumArgumentsPromoted; 569 else 570 ++NumAggregatesPromoted; 571 } 572 } 573 574 // Add any function attributes. 575 if (Attributes attrs = PAL.getFnAttributes()) 576 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); 577 578 const Type *RetTy = FTy->getReturnType(); 579 LLVMContext &Context = RetTy->getContext(); 580 581 // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which 582 // have zero fixed arguments. 583 bool ExtraArgHack = false; 584 if (Params.empty() && FTy->isVarArg()) { 585 ExtraArgHack = true; 586 Params.push_back(Type::Int32Ty); 587 } 588 589 // Construct the new function type using the new arguments. 590 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg()); 591 592 // Create the new function body and insert it into the module... 593 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName()); 594 NF->copyAttributesFrom(F); 595 596 // Recompute the parameter attributes list based on the new arguments for 597 // the function. 598 NF->setAttributes(AttrListPtr::get(AttributesVec.begin(), AttributesVec.end())); 599 AttributesVec.clear(); 600 601 F->getParent()->getFunctionList().insert(F, NF); 602 NF->takeName(F); 603 604 // Get the alias analysis information that we need to update to reflect our 605 // changes. 606 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 607 608 // Get the callgraph information that we need to update to reflect our 609 // changes. 610 CallGraph &CG = getAnalysis<CallGraph>(); 611 612 // Loop over all of the callers of the function, transforming the call sites 613 // to pass in the loaded pointers. 614 // 615 SmallVector<Value*, 16> Args; 616 while (!F->use_empty()) { 617 CallSite CS = CallSite::get(F->use_back()); 618 Instruction *Call = CS.getInstruction(); 619 const AttrListPtr &CallPAL = CS.getAttributes(); 620 621 // Add any return attributes. 622 if (Attributes attrs = CallPAL.getRetAttributes()) 623 AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); 624 625 // Loop over the operands, inserting GEP and loads in the caller as 626 // appropriate. 627 CallSite::arg_iterator AI = CS.arg_begin(); 628 ArgIndex = 1; 629 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 630 I != E; ++I, ++AI, ++ArgIndex) 631 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 632 Args.push_back(*AI); // Unmodified argument 633 634 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) 635 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 636 637 } else if (ByValArgsToTransform.count(I)) { 638 // Emit a GEP and load for each element of the struct. 639 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 640 const StructType *STy = cast<StructType>(AgTy); 641 Value *Idxs[2] = { ConstantInt::get(Type::Int32Ty, 0), 0 }; 642 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 643 Idxs[1] = ConstantInt::get(Type::Int32Ty, 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 = (isa<StructType>(ElTy) ? Type::Int32Ty : Type::Int64Ty); 668 Ops.push_back(ConstantInt::get(IdxTy, *II)); 669 // Keep track of the type we're currently indexing 670 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II); 671 } 672 // And create a GEP to extract those indices 673 V = GetElementPtrInst::Create(V, Ops.begin(), Ops.end(), 674 V->getName()+".idx", Call); 675 Ops.clear(); 676 AA.copyValue(OrigLoad->getOperand(0), V); 677 } 678 Args.push_back(new LoadInst(V, V->getName()+".val", Call)); 679 AA.copyValue(OrigLoad, Args.back()); 680 } 681 } 682 683 if (ExtraArgHack) 684 Args.push_back(Context.getNullValue(Type::Int32Ty)); 685 686 // Push any varargs arguments on the list 687 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) { 688 Args.push_back(*AI); 689 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) 690 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); 691 } 692 693 // Add any function attributes. 694 if (Attributes attrs = CallPAL.getFnAttributes()) 695 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); 696 697 Instruction *New; 698 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 699 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 700 Args.begin(), Args.end(), "", Call); 701 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 702 cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(), 703 AttributesVec.end())); 704 } else { 705 New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call); 706 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 707 cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(), 708 AttributesVec.end())); 709 if (cast<CallInst>(Call)->isTailCall()) 710 cast<CallInst>(New)->setTailCall(); 711 } 712 Args.clear(); 713 AttributesVec.clear(); 714 715 // Update the alias analysis implementation to know that we are replacing 716 // the old call with a new one. 717 AA.replaceWithNewValue(Call, New); 718 719 // Update the callgraph to know that the callsite has been transformed. 720 CG[Call->getParent()->getParent()]->replaceCallSite(Call, New); 721 722 if (!Call->use_empty()) { 723 Call->replaceAllUsesWith(New); 724 New->takeName(Call); 725 } 726 727 // Finally, remove the old call from the program, reducing the use-count of 728 // F. 729 Call->eraseFromParent(); 730 } 731 732 // Since we have now created the new function, splice the body of the old 733 // function right into the new function, leaving the old rotting hulk of the 734 // function empty. 735 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 736 737 // Loop over the argument list, transfering uses of the old arguments over to 738 // the new arguments, also transfering over the names as well. 739 // 740 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), 741 I2 = NF->arg_begin(); I != E; ++I) { 742 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 743 // If this is an unmodified argument, move the name and users over to the 744 // new version. 745 I->replaceAllUsesWith(I2); 746 I2->takeName(I); 747 AA.replaceWithNewValue(I, I2); 748 ++I2; 749 continue; 750 } 751 752 if (ByValArgsToTransform.count(I)) { 753 // In the callee, we create an alloca, and store each of the new incoming 754 // arguments into the alloca. 755 Instruction *InsertPt = NF->begin()->begin(); 756 757 // Just add all the struct element types. 758 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 759 Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt); 760 const StructType *STy = cast<StructType>(AgTy); 761 Value *Idxs[2] = { ConstantInt::get(Type::Int32Ty, 0), 0 }; 762 763 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 764 Idxs[1] = ConstantInt::get(Type::Int32Ty, i); 765 Value *Idx = 766 GetElementPtrInst::Create(TheAlloca, Idxs, Idxs+2, 767 TheAlloca->getName()+"."+Twine(i), 768 InsertPt); 769 I2->setName(I->getName()+"."+Twine(i)); 770 new StoreInst(I2++, Idx, InsertPt); 771 } 772 773 // Anything that used the arg should now use the alloca. 774 I->replaceAllUsesWith(TheAlloca); 775 TheAlloca->takeName(I); 776 AA.replaceWithNewValue(I, TheAlloca); 777 continue; 778 } 779 780 if (I->use_empty()) { 781 AA.deleteValue(I); 782 continue; 783 } 784 785 // Otherwise, if we promoted this argument, then all users are load 786 // instructions (or GEPs with only load users), and all loads should be 787 // using the new argument that we added. 788 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 789 790 while (!I->use_empty()) { 791 if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) { 792 assert(ArgIndices.begin()->empty() && 793 "Load element should sort to front!"); 794 I2->setName(I->getName()+".val"); 795 LI->replaceAllUsesWith(I2); 796 AA.replaceWithNewValue(LI, I2); 797 LI->eraseFromParent(); 798 DEBUG(errs() << "*** Promoted load of argument '" << I->getName() 799 << "' in function '" << F->getName() << "'\n"); 800 } else { 801 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back()); 802 IndicesVector Operands; 803 Operands.reserve(GEP->getNumIndices()); 804 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 805 II != IE; ++II) 806 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue()); 807 808 // GEPs with a single 0 index can be merged with direct loads 809 if (Operands.size() == 1 && Operands.front() == 0) 810 Operands.clear(); 811 812 Function::arg_iterator TheArg = I2; 813 for (ScalarizeTable::iterator It = ArgIndices.begin(); 814 *It != Operands; ++It, ++TheArg) { 815 assert(It != ArgIndices.end() && "GEP not handled??"); 816 } 817 818 std::string NewName = I->getName(); 819 for (unsigned i = 0, e = Operands.size(); i != e; ++i) { 820 NewName += "." + utostr(Operands[i]); 821 } 822 NewName += ".val"; 823 TheArg->setName(NewName); 824 825 DEBUG(errs() << "*** Promoted agg argument '" << TheArg->getName() 826 << "' of function '" << NF->getName() << "'\n"); 827 828 // All of the uses must be load instructions. Replace them all with 829 // the argument specified by ArgNo. 830 while (!GEP->use_empty()) { 831 LoadInst *L = cast<LoadInst>(GEP->use_back()); 832 L->replaceAllUsesWith(TheArg); 833 AA.replaceWithNewValue(L, TheArg); 834 L->eraseFromParent(); 835 } 836 AA.deleteValue(GEP); 837 GEP->eraseFromParent(); 838 } 839 } 840 841 // Increment I2 past all of the arguments added for this promoted pointer. 842 for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i) 843 ++I2; 844 } 845 846 // Notify the alias analysis implementation that we inserted a new argument. 847 if (ExtraArgHack) 848 AA.copyValue(Context.getNullValue(Type::Int32Ty), NF->arg_begin()); 849 850 851 // Tell the alias analysis that the old function is about to disappear. 852 AA.replaceWithNewValue(F, NF); 853 854 // Now that the old function is dead, delete it. 855 F->eraseFromParent(); 856 return NF; 857} 858