ArgumentPromotion.cpp revision 148c539e3dab2f2579e4601485d11c7272339b95
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 ArgPromotion(unsigned maxElements = 3) : CallGraphSCCPass(&ID), 70 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->hasInternalLinkage()) 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 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); 131 UI != E; ++UI) { 132 CallSite CS = CallSite::get(*UI); 133 if (!CS.getInstruction()) // "Taking the address" of the function 134 return false; 135 136 // Ensure that this call site is CALLING the function, not passing it as 137 // an argument. 138 if (UI.getOperandNo() != 0) 139 return false; 140 } 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, ParamAttr::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 DOUT << "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()) return false; 186 187 Function *NewF = DoPromotion(F, ArgsToPromote, ByValArgsToTransform); 188 189 // Update the call graph to know that the function has been transformed. 190 getAnalysis<CallGraph>().changeFunction(F, NewF); 191 return true; 192} 193 194/// IsAlwaysValidPointer - Return true if the specified pointer is always legal 195/// to load. 196static bool IsAlwaysValidPointer(Value *V) { 197 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true; 198 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) 199 return IsAlwaysValidPointer(GEP->getOperand(0)); 200 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 201 if (CE->getOpcode() == Instruction::GetElementPtr) 202 return IsAlwaysValidPointer(CE->getOperand(0)); 203 204 return false; 205} 206 207/// AllCalleesPassInValidPointerForArgument - Return true if we can prove that 208/// all callees pass in a valid pointer for the specified function argument. 209static bool AllCalleesPassInValidPointerForArgument(Argument *Arg) { 210 Function *Callee = Arg->getParent(); 211 212 unsigned ArgNo = std::distance(Callee->arg_begin(), 213 Function::arg_iterator(Arg)); 214 215 // Look at all call sites of the function. At this pointer we know we only 216 // have direct callees. 217 for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end(); 218 UI != E; ++UI) { 219 CallSite CS = CallSite::get(*UI); 220 assert(CS.getInstruction() && "Should only have direct calls!"); 221 222 if (!IsAlwaysValidPointer(CS.getArgument(ArgNo))) 223 return false; 224 } 225 return true; 226} 227 228/// Returns true if Prefix is a prefix of longer. That means, Longer has a size 229/// that is greater than or equal to the size of prefix, and each of the 230/// elements in Prefix is the same as the corresponding elements in Longer. 231/// 232/// This means it also returns true when Prefix and Longer are equal! 233static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix, 234 const ArgPromotion::IndicesVector &Longer) { 235 if (Prefix.size() > Longer.size()) 236 return false; 237 for (unsigned i = 0, e = Prefix.size(); i != e; ++i) 238 if (Prefix[i] != Longer[i]) 239 return false; 240 return true; 241} 242 243 244/// Checks if Indices, or a prefix of Indices, is in Set. 245static bool PrefixIn(const ArgPromotion::IndicesVector &Indices, 246 std::set<ArgPromotion::IndicesVector> &Set) { 247 std::set<ArgPromotion::IndicesVector>::iterator Low; 248 Low = Set.upper_bound(Indices); 249 if (Low != Set.begin()) 250 Low--; 251 // Low is now the last element smaller than or equal to Indices. This means 252 // it points to a prefix of Indices (possibly Indices itself), if such 253 // prefix exists. 254 // 255 // This load is safe if any prefix of its operands is safe to load. 256 return Low != Set.end() && IsPrefix(*Low, Indices); 257} 258 259/// Mark the given indices (ToMark) as safe in the the given set of indices 260/// (Safe). Marking safe usually means adding ToMark to Safe. However, if there 261/// is already a prefix of Indices in Safe, Indices are implicitely marked safe 262/// already. Furthermore, any indices that Indices is itself a prefix of, are 263/// removed from Safe (since they are implicitely safe because of Indices now). 264static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark, 265 std::set<ArgPromotion::IndicesVector> &Safe) { 266 std::set<ArgPromotion::IndicesVector>::iterator Low; 267 Low = Safe.upper_bound(ToMark); 268 // Guard against the case where Safe is empty 269 if (Low != Safe.begin()) 270 Low--; 271 // Low is now the last element smaller than or equal to Indices. This 272 // means it points to a prefix of Indices (possibly Indices itself), if 273 // such prefix exists. 274 if (Low != Safe.end()) { 275 if (IsPrefix(*Low, ToMark)) 276 // If there is already a prefix of these indices (or exactly these 277 // indices) marked a safe, don't bother adding these indices 278 return; 279 280 // Increment Low, so we can use it as a "insert before" hint 281 ++Low; 282 } 283 // Insert 284 Low = Safe.insert(Low, ToMark); 285 ++Low; 286 // If there we're a prefix of longer index list(s), remove those 287 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end(); 288 while (Low != End && IsPrefix(ToMark, *Low)) { 289 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low; 290 ++Low; 291 Safe.erase(Remove); 292 } 293} 294 295/// isSafeToPromoteArgument - As you might guess from the name of this method, 296/// it checks to see if it is both safe and useful to promote the argument. 297/// This method limits promotion of aggregates to only promote up to three 298/// elements of the aggregate in order to avoid exploding the number of 299/// arguments passed in. 300bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { 301 typedef std::set<IndicesVector> GEPIndicesSet; 302 303 // Quick exit for unused arguments 304 if (Arg->use_empty()) 305 return true; 306 307 // We can only promote this argument if all of the uses are loads, or are GEP 308 // instructions (with constant indices) that are subsequently loaded. 309 // 310 // Promoting the argument causes it to be loaded in the caller 311 // unconditionally. This is only safe if we can prove that either the load 312 // would have happened in the callee anyway (ie, there is a load in the entry 313 // block) or the pointer passed in at every call site is guaranteed to be 314 // valid. 315 // In the former case, invalid loads can happen, but would have happened 316 // anyway, in the latter case, invalid loads won't happen. This prevents us 317 // from introducing an invalid load that wouldn't have happened in the 318 // original code. 319 // 320 // This set will contain all sets of indices that are loaded in the entry 321 // block, and thus are safe to unconditionally load in the caller. 322 GEPIndicesSet SafeToUnconditionallyLoad; 323 324 // This set contains all the sets of indices that we are planning to promote. 325 // This makes it possible to limit the number of arguments added. 326 GEPIndicesSet ToPromote; 327 328 // If the pointer is always valid, any load with first index 0 is valid. 329 if(isByVal || AllCalleesPassInValidPointerForArgument(Arg)) 330 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0)); 331 332 // First, iterate the entry block and mark loads of (geps of) arguments as 333 // safe. 334 BasicBlock *EntryBlock = Arg->getParent()->begin(); 335 // Declare this here so we can reuse it 336 IndicesVector Indices; 337 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end(); 338 I != E; ++I) 339 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 340 Value *V = LI->getPointerOperand(); 341 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) { 342 V = GEP->getPointerOperand(); 343 if (V == Arg) { 344 // This load actually loads (part of) Arg? Check the indices then. 345 Indices.reserve(GEP->getNumIndices()); 346 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 347 II != IE; ++II) 348 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II)) 349 Indices.push_back(CI->getSExtValue()); 350 else 351 // We found a non-constant GEP index for this argument? Bail out 352 // right away, can't promote this argument at all. 353 return false; 354 355 // Indices checked out, mark them as safe 356 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad); 357 Indices.clear(); 358 } 359 } else if (V == Arg) { 360 // Direct loads are equivalent to a GEP with a single 0 index. 361 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad); 362 } 363 } 364 365 // Now, iterate all uses of the argument to see if there are any uses that are 366 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote. 367 SmallVector<LoadInst*, 16> Loads; 368 IndicesVector Operands; 369 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end(); 370 UI != E; ++UI) { 371 Operands.clear(); 372 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 373 if (LI->isVolatile()) return false; // Don't hack volatile loads 374 Loads.push_back(LI); 375 // Direct loads are equivalent to a GEP with a zero index and then a load. 376 Operands.push_back(0); 377 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 378 if (GEP->use_empty()) { 379 // Dead GEP's cause trouble later. Just remove them if we run into 380 // them. 381 getAnalysis<AliasAnalysis>().deleteValue(GEP); 382 GEP->eraseFromParent(); 383 // TODO: This runs the above loop over and over again for dead GEPS 384 // Couldn't we just do increment the UI iterator earlier and erase the 385 // use? 386 return isSafeToPromoteArgument(Arg, isByVal); 387 } 388 389 // Ensure that all of the indices are constants. 390 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end(); 391 i != e; ++i) 392 if (ConstantInt *C = dyn_cast<ConstantInt>(*i)) 393 Operands.push_back(C->getSExtValue()); 394 else 395 return false; // Not a constant operand GEP! 396 397 // Ensure that the only users of the GEP are load instructions. 398 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end(); 399 UI != E; ++UI) 400 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 401 if (LI->isVolatile()) return false; // Don't hack volatile loads 402 Loads.push_back(LI); 403 } else { 404 // Other uses than load? 405 return false; 406 } 407 } else { 408 return false; // Not a load or a GEP. 409 } 410 411 // Now, see if it is safe to promote this load / loads of this GEP. Loading 412 // is safe if Operands, or a prefix of Operands, is marked as safe. 413 if (!PrefixIn(Operands, SafeToUnconditionallyLoad)) 414 return false; 415 416 // See if we are already promoting a load with these indices. If not, check 417 // to make sure that we aren't promoting too many elements. If so, nothing 418 // to do. 419 if (ToPromote.find(Operands) == ToPromote.end()) { 420 if (maxElements > 0 && ToPromote.size() == maxElements) { 421 DOUT << "argpromotion not promoting argument '" 422 << Arg->getName() << "' because it would require adding more " 423 << "than " << maxElements << " arguments to the function.\n"; 424 // We limit aggregate promotion to only promoting up to a fixed number 425 // of elements of the aggregate. 426 return false; 427 } 428 ToPromote.insert(Operands); 429 } 430 } 431 432 if (Loads.empty()) return true; // No users, this is a dead argument. 433 434 // Okay, now we know that the argument is only used by load instructions and 435 // it is safe to unconditionally perform all of them. Use alias analysis to 436 // check to see if the pointer is guaranteed to not be modified from entry of 437 // the function to each of the load instructions. 438 439 // Because there could be several/many load instructions, remember which 440 // blocks we know to be transparent to the load. 441 SmallPtrSet<BasicBlock*, 16> TranspBlocks; 442 443 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 444 TargetData &TD = getAnalysis<TargetData>(); 445 446 for (unsigned i = 0, e = Loads.size(); i != e; ++i) { 447 // Check to see if the load is invalidated from the start of the block to 448 // the load itself. 449 LoadInst *Load = Loads[i]; 450 BasicBlock *BB = Load->getParent(); 451 452 const PointerType *LoadTy = 453 cast<PointerType>(Load->getPointerOperand()->getType()); 454 unsigned LoadSize = (unsigned)TD.getTypeStoreSize(LoadTy->getElementType()); 455 456 if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize)) 457 return false; // Pointer is invalidated! 458 459 // Now check every path from the entry block to the load for transparency. 460 // To do this, we perform a depth first search on the inverse CFG from the 461 // loading block. 462 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 463 for (idf_ext_iterator<BasicBlock*, SmallPtrSet<BasicBlock*, 16> > 464 I = idf_ext_begin(*PI, TranspBlocks), 465 E = idf_ext_end(*PI, TranspBlocks); I != E; ++I) 466 if (AA.canBasicBlockModify(**I, Arg, LoadSize)) 467 return false; 468 } 469 470 // If the path from the entry of the function to each load is free of 471 // instructions that potentially invalidate the load, we can make the 472 // transformation! 473 return true; 474} 475 476/// DoPromotion - This method actually performs the promotion of the specified 477/// arguments, and returns the new function. At this point, we know that it's 478/// safe to do so. 479Function *ArgPromotion::DoPromotion(Function *F, 480 SmallPtrSet<Argument*, 8> &ArgsToPromote, 481 SmallPtrSet<Argument*, 8> &ByValArgsToTransform) { 482 483 // Start by computing a new prototype for the function, which is the same as 484 // the old function, but has modified arguments. 485 const FunctionType *FTy = F->getFunctionType(); 486 std::vector<const Type*> Params; 487 488 typedef std::set<IndicesVector> ScalarizeTable; 489 490 // ScalarizedElements - If we are promoting a pointer that has elements 491 // accessed out of it, keep track of which elements are accessed so that we 492 // can add one argument for each. 493 // 494 // Arguments that are directly loaded will have a zero element value here, to 495 // handle cases where there are both a direct load and GEP accesses. 496 // 497 std::map<Argument*, ScalarizeTable> ScalarizedElements; 498 499 // OriginalLoads - Keep track of a representative load instruction from the 500 // original function so that we can tell the alias analysis implementation 501 // what the new GEP/Load instructions we are inserting look like. 502 std::map<IndicesVector, LoadInst*> OriginalLoads; 503 504 // ParamAttrs - Keep track of the parameter attributes for the arguments 505 // that we are *not* promoting. For the ones that we do promote, the parameter 506 // attributes are lost 507 SmallVector<ParamAttrsWithIndex, 8> ParamAttrsVec; 508 const PAListPtr &PAL = F->getParamAttrs(); 509 510 // Add any return attributes. 511 if (ParameterAttributes attrs = PAL.getParamAttrs(0)) 512 ParamAttrsVec.push_back(ParamAttrsWithIndex::get(0, attrs)); 513 514 // First, determine the new argument list 515 unsigned ArgIndex = 1; 516 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; 517 ++I, ++ArgIndex) { 518 if (ByValArgsToTransform.count(I)) { 519 // Simple byval argument? Just add all the struct element types. 520 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 521 const StructType *STy = cast<StructType>(AgTy); 522 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 523 Params.push_back(STy->getElementType(i)); 524 ++NumByValArgsPromoted; 525 } else if (!ArgsToPromote.count(I)) { 526 // Unchanged argument 527 Params.push_back(I->getType()); 528 if (ParameterAttributes attrs = PAL.getParamAttrs(ArgIndex)) 529 ParamAttrsVec.push_back(ParamAttrsWithIndex::get(Params.size(), attrs)); 530 } else if (I->use_empty()) { 531 // Dead argument (which are always marked as promotable) 532 ++NumArgumentsDead; 533 } else { 534 // Okay, this is being promoted. This means that the only uses are loads 535 // or GEPs which are only used by loads 536 537 // In this table, we will track which indices are loaded from the argument 538 // (where direct loads are tracked as no indices). 539 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 540 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 541 ++UI) { 542 Instruction *User = cast<Instruction>(*UI); 543 assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User)); 544 IndicesVector Indices; 545 Indices.reserve(User->getNumOperands() - 1); 546 // Since loads will only have a single operand, and GEPs only a single 547 // non-index operand, this will record direct loads without any indices, 548 // and gep+loads with the GEP indices. 549 for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end(); 550 II != IE; ++II) 551 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue()); 552 // GEPs with a single 0 index can be merged with direct loads 553 if (Indices.size() == 1 && Indices.front() == 0) 554 Indices.clear(); 555 ArgIndices.insert(Indices); 556 LoadInst *OrigLoad; 557 if (LoadInst *L = dyn_cast<LoadInst>(User)) 558 OrigLoad = L; 559 else 560 // Take any load, we will use it only to update Alias Analysis 561 OrigLoad = cast<LoadInst>(User->use_back()); 562 OriginalLoads[Indices] = OrigLoad; 563 } 564 565 // Add a parameter to the function for each element passed in. 566 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 567 E = ArgIndices.end(); SI != E; ++SI) { 568 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), 569 &*SI->begin(), 570 SI->size())); 571 assert(Params.back()); 572 } 573 574 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty()) 575 ++NumArgumentsPromoted; 576 else 577 ++NumAggregatesPromoted; 578 } 579 } 580 581 const Type *RetTy = FTy->getReturnType(); 582 583 // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which 584 // have zero fixed arguments. 585 bool ExtraArgHack = false; 586 if (Params.empty() && FTy->isVarArg()) { 587 ExtraArgHack = true; 588 Params.push_back(Type::Int32Ty); 589 } 590 591 // Construct the new function type using the new arguments. 592 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg()); 593 594 // Create the new function body and insert it into the module... 595 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName()); 596 NF->copyAttributesFrom(F); 597 598 // Recompute the parameter attributes list based on the new arguments for 599 // the function. 600 NF->setParamAttrs(PAListPtr::get(ParamAttrsVec.begin(), ParamAttrsVec.end())); 601 ParamAttrsVec.clear(); 602 603 F->getParent()->getFunctionList().insert(F, NF); 604 NF->takeName(F); 605 606 // Get the alias analysis information that we need to update to reflect our 607 // changes. 608 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 609 610 // Get the callgraph information that we need to update to reflect our 611 // changes. 612 CallGraph &CG = getAnalysis<CallGraph>(); 613 614 // Loop over all of the callers of the function, transforming the call sites 615 // to pass in the loaded pointers. 616 // 617 SmallVector<Value*, 16> Args; 618 while (!F->use_empty()) { 619 CallSite CS = CallSite::get(F->use_back()); 620 Instruction *Call = CS.getInstruction(); 621 const PAListPtr &CallPAL = CS.getParamAttrs(); 622 623 // Add any return attributes. 624 if (ParameterAttributes attrs = CallPAL.getParamAttrs(0)) 625 ParamAttrsVec.push_back(ParamAttrsWithIndex::get(0, attrs)); 626 627 // Loop over the operands, inserting GEP and loads in the caller as 628 // appropriate. 629 CallSite::arg_iterator AI = CS.arg_begin(); 630 ArgIndex = 1; 631 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 632 I != E; ++I, ++AI, ++ArgIndex) 633 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 634 Args.push_back(*AI); // Unmodified argument 635 636 if (ParameterAttributes Attrs = CallPAL.getParamAttrs(ArgIndex)) 637 ParamAttrsVec.push_back(ParamAttrsWithIndex::get(Args.size(), Attrs)); 638 639 } else if (ByValArgsToTransform.count(I)) { 640 // Emit a GEP and load for each element of the struct. 641 const Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 642 const StructType *STy = cast<StructType>(AgTy); 643 Value *Idxs[2] = { ConstantInt::get(Type::Int32Ty, 0), 0 }; 644 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 645 Idxs[1] = ConstantInt::get(Type::Int32Ty, i); 646 Value *Idx = GetElementPtrInst::Create(*AI, Idxs, Idxs+2, 647 (*AI)->getName()+"."+utostr(i), 648 Call); 649 // TODO: Tell AA about the new values? 650 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call)); 651 } 652 } else if (!I->use_empty()) { 653 // Non-dead argument: insert GEPs and loads as appropriate. 654 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 655 // Store the Value* version of the indices in here, but declare it now 656 // for reuse 657 std::vector<Value*> Ops; 658 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 659 E = ArgIndices.end(); SI != E; ++SI) { 660 Value *V = *AI; 661 LoadInst *OrigLoad = OriginalLoads[*SI]; 662 if (!SI->empty()) { 663 Ops.reserve(SI->size()); 664 const Type *ElTy = V->getType(); 665 for (IndicesVector::const_iterator II = SI->begin(), 666 IE = SI->end(); II != IE; ++II) { 667 // Use i32 to index structs, and i64 for others (pointers/arrays). 668 // This satisfies GEP constraints. 669 const Type *IdxTy = (isa<StructType>(ElTy) ? Type::Int32Ty : Type::Int64Ty); 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 Args.push_back(new LoadInst(V, V->getName()+".val", Call)); 681 AA.copyValue(OrigLoad, Args.back()); 682 } 683 } 684 685 if (ExtraArgHack) 686 Args.push_back(Constant::getNullValue(Type::Int32Ty)); 687 688 // Push any varargs arguments on the list 689 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) { 690 Args.push_back(*AI); 691 if (ParameterAttributes Attrs = CallPAL.getParamAttrs(ArgIndex)) 692 ParamAttrsVec.push_back(ParamAttrsWithIndex::get(Args.size(), Attrs)); 693 } 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)->setParamAttrs(PAListPtr::get(ParamAttrsVec.begin(), 701 ParamAttrsVec.end())); 702 } else { 703 New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call); 704 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 705 cast<CallInst>(New)->setParamAttrs(PAListPtr::get(ParamAttrsVec.begin(), 706 ParamAttrsVec.end())); 707 if (cast<CallInst>(Call)->isTailCall()) 708 cast<CallInst>(New)->setTailCall(); 709 } 710 Args.clear(); 711 ParamAttrsVec.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