ArgumentPromotion.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
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/ADT/DepthFirstIterator.h" 35#include "llvm/ADT/Statistic.h" 36#include "llvm/ADT/StringExtras.h" 37#include "llvm/Analysis/AliasAnalysis.h" 38#include "llvm/Analysis/CallGraph.h" 39#include "llvm/Analysis/CallGraphSCCPass.h" 40#include "llvm/IR/CFG.h" 41#include "llvm/IR/CallSite.h" 42#include "llvm/IR/Constants.h" 43#include "llvm/IR/DerivedTypes.h" 44#include "llvm/IR/Instructions.h" 45#include "llvm/IR/LLVMContext.h" 46#include "llvm/IR/Module.h" 47#include "llvm/Support/Debug.h" 48#include "llvm/Support/raw_ostream.h" 49#include <set> 50using namespace llvm; 51 52STATISTIC(NumArgumentsPromoted , "Number of pointer arguments promoted"); 53STATISTIC(NumAggregatesPromoted, "Number of aggregate arguments promoted"); 54STATISTIC(NumByValArgsPromoted , "Number of byval arguments promoted"); 55STATISTIC(NumArgumentsDead , "Number of dead pointer args eliminated"); 56 57namespace { 58 /// ArgPromotion - The 'by reference' to 'by value' argument promotion pass. 59 /// 60 struct ArgPromotion : public CallGraphSCCPass { 61 void getAnalysisUsage(AnalysisUsage &AU) const override { 62 AU.addRequired<AliasAnalysis>(); 63 CallGraphSCCPass::getAnalysisUsage(AU); 64 } 65 66 bool runOnSCC(CallGraphSCC &SCC) override; 67 static char ID; // Pass identification, replacement for typeid 68 explicit ArgPromotion(unsigned maxElements = 3) 69 : CallGraphSCCPass(ID), maxElements(maxElements) { 70 initializeArgPromotionPass(*PassRegistry::getPassRegistry()); 71 } 72 73 /// A vector used to hold the indices of a single GEP instruction 74 typedef std::vector<uint64_t> IndicesVector; 75 76 private: 77 CallGraphNode *PromoteArguments(CallGraphNode *CGN); 78 bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const; 79 CallGraphNode *DoPromotion(Function *F, 80 SmallPtrSet<Argument*, 8> &ArgsToPromote, 81 SmallPtrSet<Argument*, 8> &ByValArgsToTransform); 82 /// The maximum number of elements to expand, or 0 for unlimited. 83 unsigned maxElements; 84 }; 85} 86 87char ArgPromotion::ID = 0; 88INITIALIZE_PASS_BEGIN(ArgPromotion, "argpromotion", 89 "Promote 'by reference' arguments to scalars", false, false) 90INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 91INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass) 92INITIALIZE_PASS_END(ArgPromotion, "argpromotion", 93 "Promote 'by reference' arguments to scalars", false, false) 94 95Pass *llvm::createArgumentPromotionPass(unsigned maxElements) { 96 return new ArgPromotion(maxElements); 97} 98 99bool ArgPromotion::runOnSCC(CallGraphSCC &SCC) { 100 bool Changed = false, LocalChange; 101 102 do { // Iterate until we stop promoting from this SCC. 103 LocalChange = false; 104 // Attempt to promote arguments from all functions in this SCC. 105 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) { 106 if (CallGraphNode *CGN = PromoteArguments(*I)) { 107 LocalChange = true; 108 SCC.ReplaceNode(*I, CGN); 109 } 110 } 111 Changed |= LocalChange; // Remember that we changed something. 112 } while (LocalChange); 113 114 return Changed; 115} 116 117/// PromoteArguments - This method checks the specified function to see if there 118/// are any promotable arguments and if it is safe to promote the function (for 119/// example, all callers are direct). If safe to promote some arguments, it 120/// calls the DoPromotion method. 121/// 122CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) { 123 Function *F = CGN->getFunction(); 124 125 // Make sure that it is local to this module. 126 if (!F || !F->hasLocalLinkage()) return 0; 127 128 // First check: see if there are any pointer arguments! If not, quick exit. 129 SmallVector<Argument*, 16> PointerArgs; 130 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I) 131 if (I->getType()->isPointerTy()) 132 PointerArgs.push_back(I); 133 if (PointerArgs.empty()) return 0; 134 135 // Second check: make sure that all callers are direct callers. We can't 136 // transform functions that have indirect callers. Also see if the function 137 // is self-recursive. 138 bool isSelfRecursive = false; 139 for (Use &U : F->uses()) { 140 CallSite CS(U.getUser()); 141 // Must be a direct call. 142 if (CS.getInstruction() == 0 || !CS.isCallee(&U)) return 0; 143 144 if (CS.getInstruction()->getParent()->getParent() == F) 145 isSelfRecursive = true; 146 } 147 148 // Check to see which arguments are promotable. If an argument is promotable, 149 // add it to ArgsToPromote. 150 SmallPtrSet<Argument*, 8> ArgsToPromote; 151 SmallPtrSet<Argument*, 8> ByValArgsToTransform; 152 for (unsigned i = 0, e = PointerArgs.size(); i != e; ++i) { 153 Argument *PtrArg = PointerArgs[i]; 154 Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType(); 155 156 // If this is a byval argument, and if the aggregate type is small, just 157 // pass the elements, which is always safe. This does not apply to 158 // inalloca. 159 if (PtrArg->hasByValAttr()) { 160 if (StructType *STy = dyn_cast<StructType>(AgTy)) { 161 if (maxElements > 0 && STy->getNumElements() > maxElements) { 162 DEBUG(dbgs() << "argpromotion disable promoting argument '" 163 << PtrArg->getName() << "' because it would require adding more" 164 << " than " << maxElements << " arguments to the function.\n"); 165 continue; 166 } 167 168 // If all the elements are single-value types, we can promote it. 169 bool AllSimple = true; 170 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 171 if (!STy->getElementType(i)->isSingleValueType()) { 172 AllSimple = false; 173 break; 174 } 175 } 176 177 // Safe to transform, don't even bother trying to "promote" it. 178 // Passing the elements as a scalar will allow scalarrepl to hack on 179 // the new alloca we introduce. 180 if (AllSimple) { 181 ByValArgsToTransform.insert(PtrArg); 182 continue; 183 } 184 } 185 } 186 187 // If the argument is a recursive type and we're in a recursive 188 // function, we could end up infinitely peeling the function argument. 189 if (isSelfRecursive) { 190 if (StructType *STy = dyn_cast<StructType>(AgTy)) { 191 bool RecursiveType = false; 192 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 193 if (STy->getElementType(i) == PtrArg->getType()) { 194 RecursiveType = true; 195 break; 196 } 197 } 198 if (RecursiveType) 199 continue; 200 } 201 } 202 203 // Otherwise, see if we can promote the pointer to its value. 204 if (isSafeToPromoteArgument(PtrArg, PtrArg->hasByValOrInAllocaAttr())) 205 ArgsToPromote.insert(PtrArg); 206 } 207 208 // No promotable pointer arguments. 209 if (ArgsToPromote.empty() && ByValArgsToTransform.empty()) 210 return 0; 211 212 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform); 213} 214 215/// AllCallersPassInValidPointerForArgument - Return true if we can prove that 216/// all callees pass in a valid pointer for the specified function argument. 217static bool AllCallersPassInValidPointerForArgument(Argument *Arg) { 218 Function *Callee = Arg->getParent(); 219 220 unsigned ArgNo = Arg->getArgNo(); 221 222 // Look at all call sites of the function. At this pointer we know we only 223 // have direct callees. 224 for (User *U : Callee->users()) { 225 CallSite CS(U); 226 assert(CS && "Should only have direct calls!"); 227 228 if (!CS.getArgument(ArgNo)->isDereferenceablePointer()) 229 return false; 230 } 231 return true; 232} 233 234/// Returns true if Prefix is a prefix of longer. That means, Longer has a size 235/// that is greater than or equal to the size of prefix, and each of the 236/// elements in Prefix is the same as the corresponding elements in Longer. 237/// 238/// This means it also returns true when Prefix and Longer are equal! 239static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix, 240 const ArgPromotion::IndicesVector &Longer) { 241 if (Prefix.size() > Longer.size()) 242 return false; 243 return std::equal(Prefix.begin(), Prefix.end(), Longer.begin()); 244} 245 246 247/// Checks if Indices, or a prefix of Indices, is in Set. 248static bool PrefixIn(const ArgPromotion::IndicesVector &Indices, 249 std::set<ArgPromotion::IndicesVector> &Set) { 250 std::set<ArgPromotion::IndicesVector>::iterator Low; 251 Low = Set.upper_bound(Indices); 252 if (Low != Set.begin()) 253 Low--; 254 // Low is now the last element smaller than or equal to Indices. This means 255 // it points to a prefix of Indices (possibly Indices itself), if such 256 // prefix exists. 257 // 258 // This load is safe if any prefix of its operands is safe to load. 259 return Low != Set.end() && IsPrefix(*Low, Indices); 260} 261 262/// Mark the given indices (ToMark) as safe in the given set of indices 263/// (Safe). Marking safe usually means adding ToMark to Safe. However, if there 264/// is already a prefix of Indices in Safe, Indices are implicitely marked safe 265/// already. Furthermore, any indices that Indices is itself a prefix of, are 266/// removed from Safe (since they are implicitely safe because of Indices now). 267static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark, 268 std::set<ArgPromotion::IndicesVector> &Safe) { 269 std::set<ArgPromotion::IndicesVector>::iterator Low; 270 Low = Safe.upper_bound(ToMark); 271 // Guard against the case where Safe is empty 272 if (Low != Safe.begin()) 273 Low--; 274 // Low is now the last element smaller than or equal to Indices. This 275 // means it points to a prefix of Indices (possibly Indices itself), if 276 // such prefix exists. 277 if (Low != Safe.end()) { 278 if (IsPrefix(*Low, ToMark)) 279 // If there is already a prefix of these indices (or exactly these 280 // indices) marked a safe, don't bother adding these indices 281 return; 282 283 // Increment Low, so we can use it as a "insert before" hint 284 ++Low; 285 } 286 // Insert 287 Low = Safe.insert(Low, ToMark); 288 ++Low; 289 // If there we're a prefix of longer index list(s), remove those 290 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end(); 291 while (Low != End && IsPrefix(ToMark, *Low)) { 292 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low; 293 ++Low; 294 Safe.erase(Remove); 295 } 296} 297 298/// isSafeToPromoteArgument - As you might guess from the name of this method, 299/// it checks to see if it is both safe and useful to promote the argument. 300/// This method limits promotion of aggregates to only promote up to three 301/// elements of the aggregate in order to avoid exploding the number of 302/// arguments passed in. 303bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, 304 bool isByValOrInAlloca) const { 305 typedef std::set<IndicesVector> GEPIndicesSet; 306 307 // Quick exit for unused arguments 308 if (Arg->use_empty()) 309 return true; 310 311 // We can only promote this argument if all of the uses are loads, or are GEP 312 // instructions (with constant indices) that are subsequently loaded. 313 // 314 // Promoting the argument causes it to be loaded in the caller 315 // unconditionally. This is only safe if we can prove that either the load 316 // would have happened in the callee anyway (ie, there is a load in the entry 317 // block) or the pointer passed in at every call site is guaranteed to be 318 // valid. 319 // In the former case, invalid loads can happen, but would have happened 320 // anyway, in the latter case, invalid loads won't happen. This prevents us 321 // from introducing an invalid load that wouldn't have happened in the 322 // original code. 323 // 324 // This set will contain all sets of indices that are loaded in the entry 325 // block, and thus are safe to unconditionally load in the caller. 326 // 327 // This optimization is also safe for InAlloca parameters, because it verifies 328 // that the address isn't captured. 329 GEPIndicesSet SafeToUnconditionallyLoad; 330 331 // This set contains all the sets of indices that we are planning to promote. 332 // This makes it possible to limit the number of arguments added. 333 GEPIndicesSet ToPromote; 334 335 // If the pointer is always valid, any load with first index 0 is valid. 336 if (isByValOrInAlloca || AllCallersPassInValidPointerForArgument(Arg)) 337 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0)); 338 339 // First, iterate the entry block and mark loads of (geps of) arguments as 340 // safe. 341 BasicBlock *EntryBlock = Arg->getParent()->begin(); 342 // Declare this here so we can reuse it 343 IndicesVector Indices; 344 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end(); 345 I != E; ++I) 346 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 347 Value *V = LI->getPointerOperand(); 348 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) { 349 V = GEP->getPointerOperand(); 350 if (V == Arg) { 351 // This load actually loads (part of) Arg? Check the indices then. 352 Indices.reserve(GEP->getNumIndices()); 353 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 354 II != IE; ++II) 355 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II)) 356 Indices.push_back(CI->getSExtValue()); 357 else 358 // We found a non-constant GEP index for this argument? Bail out 359 // right away, can't promote this argument at all. 360 return false; 361 362 // Indices checked out, mark them as safe 363 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad); 364 Indices.clear(); 365 } 366 } else if (V == Arg) { 367 // Direct loads are equivalent to a GEP with a single 0 index. 368 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad); 369 } 370 } 371 372 // Now, iterate all uses of the argument to see if there are any uses that are 373 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote. 374 SmallVector<LoadInst*, 16> Loads; 375 IndicesVector Operands; 376 for (Use &U : Arg->uses()) { 377 User *UR = U.getUser(); 378 Operands.clear(); 379 if (LoadInst *LI = dyn_cast<LoadInst>(UR)) { 380 // Don't hack volatile/atomic loads 381 if (!LI->isSimple()) return false; 382 Loads.push_back(LI); 383 // Direct loads are equivalent to a GEP with a zero index and then a load. 384 Operands.push_back(0); 385 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UR)) { 386 if (GEP->use_empty()) { 387 // Dead GEP's cause trouble later. Just remove them if we run into 388 // them. 389 getAnalysis<AliasAnalysis>().deleteValue(GEP); 390 GEP->eraseFromParent(); 391 // TODO: This runs the above loop over and over again for dead GEPs 392 // Couldn't we just do increment the UI iterator earlier and erase the 393 // use? 394 return isSafeToPromoteArgument(Arg, isByValOrInAlloca); 395 } 396 397 // Ensure that all of the indices are constants. 398 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end(); 399 i != e; ++i) 400 if (ConstantInt *C = dyn_cast<ConstantInt>(*i)) 401 Operands.push_back(C->getSExtValue()); 402 else 403 return false; // Not a constant operand GEP! 404 405 // Ensure that the only users of the GEP are load instructions. 406 for (User *GEPU : GEP->users()) 407 if (LoadInst *LI = dyn_cast<LoadInst>(GEPU)) { 408 // Don't hack volatile/atomic loads 409 if (!LI->isSimple()) return false; 410 Loads.push_back(LI); 411 } else { 412 // Other uses than load? 413 return false; 414 } 415 } else { 416 return false; // Not a load or a GEP. 417 } 418 419 // Now, see if it is safe to promote this load / loads of this GEP. Loading 420 // is safe if Operands, or a prefix of Operands, is marked as safe. 421 if (!PrefixIn(Operands, SafeToUnconditionallyLoad)) 422 return false; 423 424 // See if we are already promoting a load with these indices. If not, check 425 // to make sure that we aren't promoting too many elements. If so, nothing 426 // to do. 427 if (ToPromote.find(Operands) == ToPromote.end()) { 428 if (maxElements > 0 && ToPromote.size() == maxElements) { 429 DEBUG(dbgs() << "argpromotion not promoting argument '" 430 << Arg->getName() << "' because it would require adding more " 431 << "than " << maxElements << " arguments to the function.\n"); 432 // We limit aggregate promotion to only promoting up to a fixed number 433 // of elements of the aggregate. 434 return false; 435 } 436 ToPromote.insert(Operands); 437 } 438 } 439 440 if (Loads.empty()) return true; // No users, this is a dead argument. 441 442 // Okay, now we know that the argument is only used by load instructions and 443 // it is safe to unconditionally perform all of them. Use alias analysis to 444 // check to see if the pointer is guaranteed to not be modified from entry of 445 // the function to each of the load instructions. 446 447 // Because there could be several/many load instructions, remember which 448 // blocks we know to be transparent to the load. 449 SmallPtrSet<BasicBlock*, 16> TranspBlocks; 450 451 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 452 453 for (unsigned i = 0, e = Loads.size(); i != e; ++i) { 454 // Check to see if the load is invalidated from the start of the block to 455 // the load itself. 456 LoadInst *Load = Loads[i]; 457 BasicBlock *BB = Load->getParent(); 458 459 AliasAnalysis::Location Loc = AA.getLocation(Load); 460 if (AA.canInstructionRangeModify(BB->front(), *Load, Loc)) 461 return false; // Pointer is invalidated! 462 463 // Now check every path from the entry block to the load for transparency. 464 // To do this, we perform a depth first search on the inverse CFG from the 465 // loading block. 466 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 467 BasicBlock *P = *PI; 468 for (idf_ext_iterator<BasicBlock*, SmallPtrSet<BasicBlock*, 16> > 469 I = idf_ext_begin(P, TranspBlocks), 470 E = idf_ext_end(P, TranspBlocks); I != E; ++I) 471 if (AA.canBasicBlockModify(**I, Loc)) 472 return false; 473 } 474 } 475 476 // If the path from the entry of the function to each load is free of 477 // instructions that potentially invalidate the load, we can make the 478 // transformation! 479 return true; 480} 481 482/// DoPromotion - This method actually performs the promotion of the specified 483/// arguments, and returns the new function. At this point, we know that it's 484/// safe to do so. 485CallGraphNode *ArgPromotion::DoPromotion(Function *F, 486 SmallPtrSet<Argument*, 8> &ArgsToPromote, 487 SmallPtrSet<Argument*, 8> &ByValArgsToTransform) { 488 489 // Start by computing a new prototype for the function, which is the same as 490 // the old function, but has modified arguments. 491 FunctionType *FTy = F->getFunctionType(); 492 std::vector<Type*> Params; 493 494 typedef std::set<IndicesVector> ScalarizeTable; 495 496 // ScalarizedElements - If we are promoting a pointer that has elements 497 // accessed out of it, keep track of which elements are accessed so that we 498 // can add one argument for each. 499 // 500 // Arguments that are directly loaded will have a zero element value here, to 501 // handle cases where there are both a direct load and GEP accesses. 502 // 503 std::map<Argument*, ScalarizeTable> ScalarizedElements; 504 505 // OriginalLoads - Keep track of a representative load instruction from the 506 // original function so that we can tell the alias analysis implementation 507 // what the new GEP/Load instructions we are inserting look like. 508 // We need to keep the original loads for each argument and the elements 509 // of the argument that are accessed. 510 std::map<std::pair<Argument*, IndicesVector>, LoadInst*> OriginalLoads; 511 512 // Attribute - Keep track of the parameter attributes for the arguments 513 // that we are *not* promoting. For the ones that we do promote, the parameter 514 // attributes are lost 515 SmallVector<AttributeSet, 8> AttributesVec; 516 const AttributeSet &PAL = F->getAttributes(); 517 518 // Add any return attributes. 519 if (PAL.hasAttributes(AttributeSet::ReturnIndex)) 520 AttributesVec.push_back(AttributeSet::get(F->getContext(), 521 PAL.getRetAttributes())); 522 523 // First, determine the new argument list 524 unsigned ArgIndex = 1; 525 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; 526 ++I, ++ArgIndex) { 527 if (ByValArgsToTransform.count(I)) { 528 // Simple byval argument? Just add all the struct element types. 529 Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 530 StructType *STy = cast<StructType>(AgTy); 531 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) 532 Params.push_back(STy->getElementType(i)); 533 ++NumByValArgsPromoted; 534 } else if (!ArgsToPromote.count(I)) { 535 // Unchanged argument 536 Params.push_back(I->getType()); 537 AttributeSet attrs = PAL.getParamAttributes(ArgIndex); 538 if (attrs.hasAttributes(ArgIndex)) { 539 AttrBuilder B(attrs, ArgIndex); 540 AttributesVec. 541 push_back(AttributeSet::get(F->getContext(), Params.size(), B)); 542 } 543 } else if (I->use_empty()) { 544 // Dead argument (which are always marked as promotable) 545 ++NumArgumentsDead; 546 } else { 547 // Okay, this is being promoted. This means that the only uses are loads 548 // or GEPs which are only used by loads 549 550 // In this table, we will track which indices are loaded from the argument 551 // (where direct loads are tracked as no indices). 552 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 553 for (User *U : I->users()) { 554 Instruction *UI = cast<Instruction>(U); 555 assert(isa<LoadInst>(UI) || isa<GetElementPtrInst>(UI)); 556 IndicesVector Indices; 557 Indices.reserve(UI->getNumOperands() - 1); 558 // Since loads will only have a single operand, and GEPs only a single 559 // non-index operand, this will record direct loads without any indices, 560 // and gep+loads with the GEP indices. 561 for (User::op_iterator II = UI->op_begin() + 1, IE = UI->op_end(); 562 II != IE; ++II) 563 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue()); 564 // GEPs with a single 0 index can be merged with direct loads 565 if (Indices.size() == 1 && Indices.front() == 0) 566 Indices.clear(); 567 ArgIndices.insert(Indices); 568 LoadInst *OrigLoad; 569 if (LoadInst *L = dyn_cast<LoadInst>(UI)) 570 OrigLoad = L; 571 else 572 // Take any load, we will use it only to update Alias Analysis 573 OrigLoad = cast<LoadInst>(UI->user_back()); 574 OriginalLoads[std::make_pair(I, Indices)] = OrigLoad; 575 } 576 577 // Add a parameter to the function for each element passed in. 578 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 579 E = ArgIndices.end(); SI != E; ++SI) { 580 // not allowed to dereference ->begin() if size() is 0 581 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), *SI)); 582 assert(Params.back()); 583 } 584 585 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty()) 586 ++NumArgumentsPromoted; 587 else 588 ++NumAggregatesPromoted; 589 } 590 } 591 592 // Add any function attributes. 593 if (PAL.hasAttributes(AttributeSet::FunctionIndex)) 594 AttributesVec.push_back(AttributeSet::get(FTy->getContext(), 595 PAL.getFnAttributes())); 596 597 Type *RetTy = FTy->getReturnType(); 598 599 // Construct the new function type using the new arguments. 600 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg()); 601 602 // Create the new function body and insert it into the module. 603 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName()); 604 NF->copyAttributesFrom(F); 605 606 607 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n" 608 << "From: " << *F); 609 610 // Recompute the parameter attributes list based on the new arguments for 611 // the function. 612 NF->setAttributes(AttributeSet::get(F->getContext(), AttributesVec)); 613 AttributesVec.clear(); 614 615 F->getParent()->getFunctionList().insert(F, NF); 616 NF->takeName(F); 617 618 // Get the alias analysis information that we need to update to reflect our 619 // changes. 620 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 621 622 // Get the callgraph information that we need to update to reflect our 623 // changes. 624 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph(); 625 626 // Get a new callgraph node for NF. 627 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF); 628 629 // Loop over all of the callers of the function, transforming the call sites 630 // to pass in the loaded pointers. 631 // 632 SmallVector<Value*, 16> Args; 633 while (!F->use_empty()) { 634 CallSite CS(F->user_back()); 635 assert(CS.getCalledFunction() == F); 636 Instruction *Call = CS.getInstruction(); 637 const AttributeSet &CallPAL = CS.getAttributes(); 638 639 // Add any return attributes. 640 if (CallPAL.hasAttributes(AttributeSet::ReturnIndex)) 641 AttributesVec.push_back(AttributeSet::get(F->getContext(), 642 CallPAL.getRetAttributes())); 643 644 // Loop over the operands, inserting GEP and loads in the caller as 645 // appropriate. 646 CallSite::arg_iterator AI = CS.arg_begin(); 647 ArgIndex = 1; 648 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 649 I != E; ++I, ++AI, ++ArgIndex) 650 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 651 Args.push_back(*AI); // Unmodified argument 652 653 if (CallPAL.hasAttributes(ArgIndex)) { 654 AttrBuilder B(CallPAL, ArgIndex); 655 AttributesVec. 656 push_back(AttributeSet::get(F->getContext(), Args.size(), B)); 657 } 658 } else if (ByValArgsToTransform.count(I)) { 659 // Emit a GEP and load for each element of the struct. 660 Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 661 StructType *STy = cast<StructType>(AgTy); 662 Value *Idxs[2] = { 663 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; 664 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 665 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); 666 Value *Idx = GetElementPtrInst::Create(*AI, Idxs, 667 (*AI)->getName()+"."+utostr(i), 668 Call); 669 // TODO: Tell AA about the new values? 670 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call)); 671 } 672 } else if (!I->use_empty()) { 673 // Non-dead argument: insert GEPs and loads as appropriate. 674 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 675 // Store the Value* version of the indices in here, but declare it now 676 // for reuse. 677 std::vector<Value*> Ops; 678 for (ScalarizeTable::iterator SI = ArgIndices.begin(), 679 E = ArgIndices.end(); SI != E; ++SI) { 680 Value *V = *AI; 681 LoadInst *OrigLoad = OriginalLoads[std::make_pair(I, *SI)]; 682 if (!SI->empty()) { 683 Ops.reserve(SI->size()); 684 Type *ElTy = V->getType(); 685 for (IndicesVector::const_iterator II = SI->begin(), 686 IE = SI->end(); II != IE; ++II) { 687 // Use i32 to index structs, and i64 for others (pointers/arrays). 688 // This satisfies GEP constraints. 689 Type *IdxTy = (ElTy->isStructTy() ? 690 Type::getInt32Ty(F->getContext()) : 691 Type::getInt64Ty(F->getContext())); 692 Ops.push_back(ConstantInt::get(IdxTy, *II)); 693 // Keep track of the type we're currently indexing. 694 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II); 695 } 696 // And create a GEP to extract those indices. 697 V = GetElementPtrInst::Create(V, Ops, V->getName()+".idx", Call); 698 Ops.clear(); 699 AA.copyValue(OrigLoad->getOperand(0), V); 700 } 701 // Since we're replacing a load make sure we take the alignment 702 // of the previous load. 703 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call); 704 newLoad->setAlignment(OrigLoad->getAlignment()); 705 // Transfer the TBAA info too. 706 newLoad->setMetadata(LLVMContext::MD_tbaa, 707 OrigLoad->getMetadata(LLVMContext::MD_tbaa)); 708 Args.push_back(newLoad); 709 AA.copyValue(OrigLoad, Args.back()); 710 } 711 } 712 713 // Push any varargs arguments on the list. 714 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) { 715 Args.push_back(*AI); 716 if (CallPAL.hasAttributes(ArgIndex)) { 717 AttrBuilder B(CallPAL, ArgIndex); 718 AttributesVec. 719 push_back(AttributeSet::get(F->getContext(), Args.size(), B)); 720 } 721 } 722 723 // Add any function attributes. 724 if (CallPAL.hasAttributes(AttributeSet::FunctionIndex)) 725 AttributesVec.push_back(AttributeSet::get(Call->getContext(), 726 CallPAL.getFnAttributes())); 727 728 Instruction *New; 729 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) { 730 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), 731 Args, "", Call); 732 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv()); 733 cast<InvokeInst>(New)->setAttributes(AttributeSet::get(II->getContext(), 734 AttributesVec)); 735 } else { 736 New = CallInst::Create(NF, Args, "", Call); 737 cast<CallInst>(New)->setCallingConv(CS.getCallingConv()); 738 cast<CallInst>(New)->setAttributes(AttributeSet::get(New->getContext(), 739 AttributesVec)); 740 if (cast<CallInst>(Call)->isTailCall()) 741 cast<CallInst>(New)->setTailCall(); 742 } 743 Args.clear(); 744 AttributesVec.clear(); 745 746 // Update the alias analysis implementation to know that we are replacing 747 // the old call with a new one. 748 AA.replaceWithNewValue(Call, New); 749 750 // Update the callgraph to know that the callsite has been transformed. 751 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()]; 752 CalleeNode->replaceCallEdge(Call, New, NF_CGN); 753 754 if (!Call->use_empty()) { 755 Call->replaceAllUsesWith(New); 756 New->takeName(Call); 757 } 758 759 // Finally, remove the old call from the program, reducing the use-count of 760 // F. 761 Call->eraseFromParent(); 762 } 763 764 // Since we have now created the new function, splice the body of the old 765 // function right into the new function, leaving the old rotting hulk of the 766 // function empty. 767 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); 768 769 // Loop over the argument list, transferring uses of the old arguments over to 770 // the new arguments, also transferring over the names as well. 771 // 772 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), 773 I2 = NF->arg_begin(); I != E; ++I) { 774 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { 775 // If this is an unmodified argument, move the name and users over to the 776 // new version. 777 I->replaceAllUsesWith(I2); 778 I2->takeName(I); 779 AA.replaceWithNewValue(I, I2); 780 ++I2; 781 continue; 782 } 783 784 if (ByValArgsToTransform.count(I)) { 785 // In the callee, we create an alloca, and store each of the new incoming 786 // arguments into the alloca. 787 Instruction *InsertPt = NF->begin()->begin(); 788 789 // Just add all the struct element types. 790 Type *AgTy = cast<PointerType>(I->getType())->getElementType(); 791 Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt); 792 StructType *STy = cast<StructType>(AgTy); 793 Value *Idxs[2] = { 794 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; 795 796 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 797 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); 798 Value *Idx = 799 GetElementPtrInst::Create(TheAlloca, Idxs, 800 TheAlloca->getName()+"."+Twine(i), 801 InsertPt); 802 I2->setName(I->getName()+"."+Twine(i)); 803 new StoreInst(I2++, Idx, InsertPt); 804 } 805 806 // Anything that used the arg should now use the alloca. 807 I->replaceAllUsesWith(TheAlloca); 808 TheAlloca->takeName(I); 809 AA.replaceWithNewValue(I, TheAlloca); 810 811 // If the alloca is used in a call, we must clear the tail flag since 812 // the callee now uses an alloca from the caller. 813 for (User *U : TheAlloca->users()) { 814 CallInst *Call = dyn_cast<CallInst>(U); 815 if (!Call) 816 continue; 817 Call->setTailCall(false); 818 } 819 continue; 820 } 821 822 if (I->use_empty()) { 823 AA.deleteValue(I); 824 continue; 825 } 826 827 // Otherwise, if we promoted this argument, then all users are load 828 // instructions (or GEPs with only load users), and all loads should be 829 // using the new argument that we added. 830 ScalarizeTable &ArgIndices = ScalarizedElements[I]; 831 832 while (!I->use_empty()) { 833 if (LoadInst *LI = dyn_cast<LoadInst>(I->user_back())) { 834 assert(ArgIndices.begin()->empty() && 835 "Load element should sort to front!"); 836 I2->setName(I->getName()+".val"); 837 LI->replaceAllUsesWith(I2); 838 AA.replaceWithNewValue(LI, I2); 839 LI->eraseFromParent(); 840 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName() 841 << "' in function '" << F->getName() << "'\n"); 842 } else { 843 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->user_back()); 844 IndicesVector Operands; 845 Operands.reserve(GEP->getNumIndices()); 846 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); 847 II != IE; ++II) 848 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue()); 849 850 // GEPs with a single 0 index can be merged with direct loads 851 if (Operands.size() == 1 && Operands.front() == 0) 852 Operands.clear(); 853 854 Function::arg_iterator TheArg = I2; 855 for (ScalarizeTable::iterator It = ArgIndices.begin(); 856 *It != Operands; ++It, ++TheArg) { 857 assert(It != ArgIndices.end() && "GEP not handled??"); 858 } 859 860 std::string NewName = I->getName(); 861 for (unsigned i = 0, e = Operands.size(); i != e; ++i) { 862 NewName += "." + utostr(Operands[i]); 863 } 864 NewName += ".val"; 865 TheArg->setName(NewName); 866 867 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName() 868 << "' of function '" << NF->getName() << "'\n"); 869 870 // All of the uses must be load instructions. Replace them all with 871 // the argument specified by ArgNo. 872 while (!GEP->use_empty()) { 873 LoadInst *L = cast<LoadInst>(GEP->user_back()); 874 L->replaceAllUsesWith(TheArg); 875 AA.replaceWithNewValue(L, TheArg); 876 L->eraseFromParent(); 877 } 878 AA.deleteValue(GEP); 879 GEP->eraseFromParent(); 880 } 881 } 882 883 // Increment I2 past all of the arguments added for this promoted pointer. 884 std::advance(I2, ArgIndices.size()); 885 } 886 887 // Tell the alias analysis that the old function is about to disappear. 888 AA.replaceWithNewValue(F, NF); 889 890 891 NF_CGN->stealCalledFunctionsFrom(CG[F]); 892 893 // Now that the old function is dead, delete it. If there is a dangling 894 // reference to the CallgraphNode, just leave the dead function around for 895 // someone else to nuke. 896 CallGraphNode *CGN = CG[F]; 897 if (CGN->getNumReferences() == 0) 898 delete CG.removeFunctionFromModule(CGN); 899 else 900 F->setLinkage(Function::ExternalLinkage); 901 902 return NF_CGN; 903} 904