ScalarReplAggregates.cpp revision de6df88529e20541dcfab7824af2eb0776194f01
1//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This transformation implements the well known scalar replacement of 11// aggregates transformation. This xform breaks up alloca instructions of 12// aggregate type (structure or array) into individual alloca instructions for 13// each member (if possible). Then, if possible, it transforms the individual 14// alloca instructions into nice clean scalar SSA form. 15// 16// This combines a simple SRoA algorithm with the Mem2Reg algorithm because 17// often interact, especially for C++ programs. As such, iterating between 18// SRoA, then Mem2Reg until we run out of things to promote works well. 19// 20//===----------------------------------------------------------------------===// 21 22#include "llvm/Transforms/Scalar.h" 23#include "llvm/Constants.h" 24#include "llvm/DerivedTypes.h" 25#include "llvm/Function.h" 26#include "llvm/Pass.h" 27#include "llvm/Instructions.h" 28#include "llvm/Analysis/Dominators.h" 29#include "llvm/Target/TargetData.h" 30#include "llvm/Transforms/Utils/PromoteMemToReg.h" 31#include "llvm/Support/GetElementPtrTypeIterator.h" 32#include "llvm/Support/MathExtras.h" 33#include "llvm/Support/Debug.h" 34#include "llvm/ADT/Statistic.h" 35#include "llvm/ADT/StringExtras.h" 36#include <iostream> 37using namespace llvm; 38 39namespace { 40 Statistic<> NumReplaced("scalarrepl", "Number of allocas broken up"); 41 Statistic<> NumPromoted("scalarrepl", "Number of allocas promoted"); 42 Statistic<> NumConverted("scalarrepl", 43 "Number of aggregates converted to scalar"); 44 45 struct SROA : public FunctionPass { 46 bool runOnFunction(Function &F); 47 48 bool performScalarRepl(Function &F); 49 bool performPromotion(Function &F); 50 51 // getAnalysisUsage - This pass does not require any passes, but we know it 52 // will not alter the CFG, so say so. 53 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 54 AU.addRequired<DominatorTree>(); 55 AU.addRequired<DominanceFrontier>(); 56 AU.addRequired<TargetData>(); 57 AU.setPreservesCFG(); 58 } 59 60 private: 61 int isSafeElementUse(Value *Ptr); 62 int isSafeUseOfAllocation(Instruction *User); 63 int isSafeAllocaToScalarRepl(AllocationInst *AI); 64 void CanonicalizeAllocaUsers(AllocationInst *AI); 65 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 66 67 const Type *CanConvertToScalar(Value *V, bool &IsNotTrivial); 68 void ConvertToScalar(AllocationInst *AI, const Type *Ty); 69 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset); 70 }; 71 72 RegisterOpt<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 73} 74 75// Public interface to the ScalarReplAggregates pass 76FunctionPass *llvm::createScalarReplAggregatesPass() { return new SROA(); } 77 78 79bool SROA::runOnFunction(Function &F) { 80 bool Changed = performPromotion(F); 81 while (1) { 82 bool LocalChange = performScalarRepl(F); 83 if (!LocalChange) break; // No need to repromote if no scalarrepl 84 Changed = true; 85 LocalChange = performPromotion(F); 86 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 87 } 88 89 return Changed; 90} 91 92 93bool SROA::performPromotion(Function &F) { 94 std::vector<AllocaInst*> Allocas; 95 const TargetData &TD = getAnalysis<TargetData>(); 96 DominatorTree &DT = getAnalysis<DominatorTree>(); 97 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 98 99 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 100 101 bool Changed = false; 102 103 while (1) { 104 Allocas.clear(); 105 106 // Find allocas that are safe to promote, by looking at all instructions in 107 // the entry node 108 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 109 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 110 if (isAllocaPromotable(AI, TD)) 111 Allocas.push_back(AI); 112 113 if (Allocas.empty()) break; 114 115 PromoteMemToReg(Allocas, DT, DF, TD); 116 NumPromoted += Allocas.size(); 117 Changed = true; 118 } 119 120 return Changed; 121} 122 123// performScalarRepl - This algorithm is a simple worklist driven algorithm, 124// which runs on all of the malloc/alloca instructions in the function, removing 125// them if they are only used by getelementptr instructions. 126// 127bool SROA::performScalarRepl(Function &F) { 128 std::vector<AllocationInst*> WorkList; 129 130 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 131 BasicBlock &BB = F.getEntryBlock(); 132 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 133 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 134 WorkList.push_back(A); 135 136 // Process the worklist 137 bool Changed = false; 138 while (!WorkList.empty()) { 139 AllocationInst *AI = WorkList.back(); 140 WorkList.pop_back(); 141 142 // If we can turn this aggregate value (potentially with casts) into a 143 // simple scalar value that can be mem2reg'd into a register value. 144 bool IsNotTrivial = false; 145 if (const Type *ActualType = CanConvertToScalar(AI, IsNotTrivial)) 146 if (IsNotTrivial) { 147 ConvertToScalar(AI, ActualType); 148 Changed = true; 149 continue; 150 } 151 152 // We cannot transform the allocation instruction if it is an array 153 // allocation (allocations OF arrays are ok though), and an allocation of a 154 // scalar value cannot be decomposed at all. 155 // 156 if (AI->isArrayAllocation() || 157 (!isa<StructType>(AI->getAllocatedType()) && 158 !isa<ArrayType>(AI->getAllocatedType()))) continue; 159 160 // Check that all of the users of the allocation are capable of being 161 // transformed. 162 switch (isSafeAllocaToScalarRepl(AI)) { 163 default: assert(0 && "Unexpected value!"); 164 case 0: // Not safe to scalar replace. 165 continue; 166 case 1: // Safe, but requires cleanup/canonicalizations first 167 CanonicalizeAllocaUsers(AI); 168 case 3: // Safe to scalar replace. 169 break; 170 } 171 172 DEBUG(std::cerr << "Found inst to xform: " << *AI); 173 Changed = true; 174 175 std::vector<AllocaInst*> ElementAllocas; 176 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 177 ElementAllocas.reserve(ST->getNumContainedTypes()); 178 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 179 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 180 AI->getAlignment(), 181 AI->getName() + "." + utostr(i), AI); 182 ElementAllocas.push_back(NA); 183 WorkList.push_back(NA); // Add to worklist for recursive processing 184 } 185 } else { 186 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 187 ElementAllocas.reserve(AT->getNumElements()); 188 const Type *ElTy = AT->getElementType(); 189 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 190 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 191 AI->getName() + "." + utostr(i), AI); 192 ElementAllocas.push_back(NA); 193 WorkList.push_back(NA); // Add to worklist for recursive processing 194 } 195 } 196 197 // Now that we have created the alloca instructions that we want to use, 198 // expand the getelementptr instructions to use them. 199 // 200 while (!AI->use_empty()) { 201 Instruction *User = cast<Instruction>(AI->use_back()); 202 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 203 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 204 unsigned Idx = 205 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getRawValue(); 206 207 assert(Idx < ElementAllocas.size() && "Index out of range?"); 208 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 209 210 Value *RepValue; 211 if (GEPI->getNumOperands() == 3) { 212 // Do not insert a new getelementptr instruction with zero indices, only 213 // to have it optimized out later. 214 RepValue = AllocaToUse; 215 } else { 216 // We are indexing deeply into the structure, so we still need a 217 // getelement ptr instruction to finish the indexing. This may be 218 // expanded itself once the worklist is rerun. 219 // 220 std::string OldName = GEPI->getName(); // Steal the old name. 221 std::vector<Value*> NewArgs; 222 NewArgs.push_back(Constant::getNullValue(Type::IntTy)); 223 NewArgs.insert(NewArgs.end(), GEPI->op_begin()+3, GEPI->op_end()); 224 GEPI->setName(""); 225 RepValue = new GetElementPtrInst(AllocaToUse, NewArgs, OldName, GEPI); 226 } 227 228 // Move all of the users over to the new GEP. 229 GEPI->replaceAllUsesWith(RepValue); 230 // Delete the old GEP 231 GEPI->eraseFromParent(); 232 } 233 234 // Finally, delete the Alloca instruction 235 AI->getParent()->getInstList().erase(AI); 236 NumReplaced++; 237 } 238 239 return Changed; 240} 241 242 243/// isSafeElementUse - Check to see if this use is an allowed use for a 244/// getelementptr instruction of an array aggregate allocation. 245/// 246int SROA::isSafeElementUse(Value *Ptr) { 247 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 248 I != E; ++I) { 249 Instruction *User = cast<Instruction>(*I); 250 switch (User->getOpcode()) { 251 case Instruction::Load: break; 252 case Instruction::Store: 253 // Store is ok if storing INTO the pointer, not storing the pointer 254 if (User->getOperand(0) == Ptr) return 0; 255 break; 256 case Instruction::GetElementPtr: { 257 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 258 if (GEP->getNumOperands() > 1) { 259 if (!isa<Constant>(GEP->getOperand(1)) || 260 !cast<Constant>(GEP->getOperand(1))->isNullValue()) 261 return 0; // Using pointer arithmetic to navigate the array... 262 } 263 if (!isSafeElementUse(GEP)) return 0; 264 break; 265 } 266 default: 267 DEBUG(std::cerr << " Transformation preventing inst: " << *User); 268 return 0; 269 } 270 } 271 return 3; // All users look ok :) 272} 273 274/// AllUsersAreLoads - Return true if all users of this value are loads. 275static bool AllUsersAreLoads(Value *Ptr) { 276 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 277 I != E; ++I) 278 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 279 return false; 280 return true; 281} 282 283/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 284/// aggregate allocation. 285/// 286int SROA::isSafeUseOfAllocation(Instruction *User) { 287 if (!isa<GetElementPtrInst>(User)) return 0; 288 289 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 290 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 291 292 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 293 if (I == E || 294 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) 295 return 0; 296 297 ++I; 298 if (I == E) return 0; // ran out of GEP indices?? 299 300 // If this is a use of an array allocation, do a bit more checking for sanity. 301 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 302 uint64_t NumElements = AT->getNumElements(); 303 304 if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) { 305 // Check to make sure that index falls within the array. If not, 306 // something funny is going on, so we won't do the optimization. 307 // 308 if (cast<ConstantInt>(GEPI->getOperand(2))->getRawValue() >= NumElements) 309 return 0; 310 311 // We cannot scalar repl this level of the array unless any array 312 // sub-indices are in-range constants. In particular, consider: 313 // A[0][i]. We cannot know that the user isn't doing invalid things like 314 // allowing i to index an out-of-range subscript that accesses A[1]. 315 // 316 // Scalar replacing *just* the outer index of the array is probably not 317 // going to be a win anyway, so just give up. 318 for (++I; I != E && isa<ArrayType>(*I); ++I) { 319 const ArrayType *SubArrayTy = cast<ArrayType>(*I); 320 uint64_t NumElements = SubArrayTy->getNumElements(); 321 if (!isa<ConstantInt>(I.getOperand())) return 0; 322 if (cast<ConstantInt>(I.getOperand())->getRawValue() >= NumElements) 323 return 0; 324 } 325 326 } else { 327 // If this is an array index and the index is not constant, we cannot 328 // promote... that is unless the array has exactly one or two elements in 329 // it, in which case we CAN promote it, but we have to canonicalize this 330 // out if this is the only problem. 331 if ((NumElements == 1 || NumElements == 2) && 332 AllUsersAreLoads(GEPI)) 333 return 1; // Canonicalization required! 334 return 0; 335 } 336 } 337 338 // If there are any non-simple uses of this getelementptr, make sure to reject 339 // them. 340 return isSafeElementUse(GEPI); 341} 342 343/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 344/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 345/// or 1 if safe after canonicalization has been performed. 346/// 347int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 348 // Loop over the use list of the alloca. We can only transform it if all of 349 // the users are safe to transform. 350 // 351 int isSafe = 3; 352 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 353 I != E; ++I) { 354 isSafe &= isSafeUseOfAllocation(cast<Instruction>(*I)); 355 if (isSafe == 0) { 356 DEBUG(std::cerr << "Cannot transform: " << *AI << " due to user: " 357 << **I); 358 return 0; 359 } 360 } 361 // If we require cleanup, isSafe is now 1, otherwise it is 3. 362 return isSafe; 363} 364 365/// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified 366/// allocation, but only if cleaned up, perform the cleanups required. 367void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) { 368 // At this point, we know that the end result will be SROA'd and promoted, so 369 // we can insert ugly code if required so long as sroa+mem2reg will clean it 370 // up. 371 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 372 UI != E; ) { 373 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(*UI++); 374 gep_type_iterator I = gep_type_begin(GEPI); 375 ++I; 376 377 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 378 uint64_t NumElements = AT->getNumElements(); 379 380 if (!isa<ConstantInt>(I.getOperand())) { 381 if (NumElements == 1) { 382 GEPI->setOperand(2, Constant::getNullValue(Type::IntTy)); 383 } else { 384 assert(NumElements == 2 && "Unhandled case!"); 385 // All users of the GEP must be loads. At each use of the GEP, insert 386 // two loads of the appropriate indexed GEP and select between them. 387 Value *IsOne = BinaryOperator::createSetNE(I.getOperand(), 388 Constant::getNullValue(I.getOperand()->getType()), 389 "isone", GEPI); 390 // Insert the new GEP instructions, which are properly indexed. 391 std::vector<Value*> Indices(GEPI->op_begin()+1, GEPI->op_end()); 392 Indices[1] = Constant::getNullValue(Type::IntTy); 393 Value *ZeroIdx = new GetElementPtrInst(GEPI->getOperand(0), Indices, 394 GEPI->getName()+".0", GEPI); 395 Indices[1] = ConstantInt::get(Type::IntTy, 1); 396 Value *OneIdx = new GetElementPtrInst(GEPI->getOperand(0), Indices, 397 GEPI->getName()+".1", GEPI); 398 // Replace all loads of the variable index GEP with loads from both 399 // indexes and a select. 400 while (!GEPI->use_empty()) { 401 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 402 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 403 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 404 Value *R = new SelectInst(IsOne, One, Zero, LI->getName(), LI); 405 LI->replaceAllUsesWith(R); 406 LI->eraseFromParent(); 407 } 408 GEPI->eraseFromParent(); 409 } 410 } 411 } 412 } 413} 414 415/// MergeInType - Add the 'In' type to the accumulated type so far. If the 416/// types are incompatible, return true, otherwise update Accum and return 417/// false. 418/// 419/// There are two cases we handle here: 420/// 1) An effectively integer union, where the pieces are stored into as 421/// smaller integers (common with byte swap and other idioms). 422/// 2) A union of a vector and its elements. Here we turn element accesses 423/// into insert/extract element operations. 424static bool MergeInType(const Type *In, const Type *&Accum) { 425 // If this is our first type, just use it. 426 const PackedType *PTy; 427 if (Accum == Type::VoidTy || In == Accum) { 428 Accum = In; 429 } else if (In->isIntegral() && Accum->isIntegral()) { // integer union. 430 // Otherwise pick whichever type is larger. 431 if (In->getTypeID() > Accum->getTypeID()) 432 Accum = In; 433 } else if ((PTy = dyn_cast<PackedType>(Accum)) && 434 PTy->getElementType() == In) { 435 // Accum is a vector, and we are accessing an element: ok. 436 } else if ((PTy = dyn_cast<PackedType>(In)) && 437 PTy->getElementType() == Accum) { 438 // In is a vector, and accum is an element: ok, remember In. 439 Accum = In; 440 } else { 441 return true; 442 } 443 return false; 444} 445 446/// getUIntAtLeastAsBitAs - Return an unsigned integer type that is at least 447/// as big as the specified type. If there is no suitable type, this returns 448/// null. 449const Type *getUIntAtLeastAsBitAs(unsigned NumBits) { 450 if (NumBits > 64) return 0; 451 if (NumBits > 32) return Type::ULongTy; 452 if (NumBits > 16) return Type::UIntTy; 453 if (NumBits > 8) return Type::UShortTy; 454 return Type::UByteTy; 455} 456 457/// CanConvertToScalar - V is a pointer. If we can convert the pointee to a 458/// single scalar integer type, return that type. Further, if the use is not 459/// a completely trivial use that mem2reg could promote, set IsNotTrivial. If 460/// there are no uses of this pointer, return Type::VoidTy to differentiate from 461/// failure. 462/// 463const Type *SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial) { 464 const Type *UsedType = Type::VoidTy; // No uses, no forced type. 465 const TargetData &TD = getAnalysis<TargetData>(); 466 const PointerType *PTy = cast<PointerType>(V->getType()); 467 468 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 469 Instruction *User = cast<Instruction>(*UI); 470 471 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 472 if (MergeInType(LI->getType(), UsedType)) 473 return 0; 474 475 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 476 // Storing the pointer, not the into the value? 477 if (SI->getOperand(0) == V) return 0; 478 479 // NOTE: We could handle storing of FP imms into integers here! 480 481 if (MergeInType(SI->getOperand(0)->getType(), UsedType)) 482 return 0; 483 } else if (CastInst *CI = dyn_cast<CastInst>(User)) { 484 if (!isa<PointerType>(CI->getType())) return 0; 485 IsNotTrivial = true; 486 const Type *SubTy = CanConvertToScalar(CI, IsNotTrivial); 487 if (!SubTy || MergeInType(SubTy, UsedType)) return 0; 488 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 489 // Check to see if this is stepping over an element: GEP Ptr, int C 490 if (GEP->getNumOperands() == 2 && isa<ConstantInt>(GEP->getOperand(1))) { 491 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getRawValue(); 492 unsigned ElSize = TD.getTypeSize(PTy->getElementType()); 493 unsigned BitOffset = Idx*ElSize*8; 494 if (BitOffset > 64 || !isPowerOf2_32(ElSize)) return 0; 495 496 IsNotTrivial = true; 497 const Type *SubElt = CanConvertToScalar(GEP, IsNotTrivial); 498 if (SubElt == 0) return 0; 499 if (SubElt != Type::VoidTy && SubElt->isInteger()) { 500 const Type *NewTy = 501 getUIntAtLeastAsBitAs(SubElt->getPrimitiveSizeInBits()+BitOffset); 502 if (NewTy == 0 || MergeInType(NewTy, UsedType)) return 0; 503 continue; 504 } 505 } else if (GEP->getNumOperands() == 3 && 506 isa<ConstantInt>(GEP->getOperand(1)) && 507 isa<ConstantInt>(GEP->getOperand(2)) && 508 cast<Constant>(GEP->getOperand(1))->isNullValue()) { 509 // We are stepping into an element, e.g. a structure or an array: 510 // GEP Ptr, int 0, uint C 511 const Type *AggTy = PTy->getElementType(); 512 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getRawValue(); 513 514 if (const ArrayType *ATy = dyn_cast<ArrayType>(AggTy)) { 515 if (Idx >= ATy->getNumElements()) return 0; // Out of range. 516 } else if (const PackedType *PackedTy = dyn_cast<PackedType>(AggTy)) { 517 // Getting an element of the packed vector. 518 if (Idx >= PackedTy->getNumElements()) return 0; // Out of range. 519 520 // Merge in the packed type. 521 if (MergeInType(PackedTy, UsedType)) return 0; 522 523 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial); 524 if (SubTy == 0) return 0; 525 526 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType)) 527 return 0; 528 529 // We'll need to change this to an insert/extract element operation. 530 IsNotTrivial = true; 531 continue; // Everything looks ok 532 533 } else if (isa<StructType>(AggTy)) { 534 // Structs are always ok. 535 } else { 536 return 0; 537 } 538 const Type *NTy = getUIntAtLeastAsBitAs(TD.getTypeSize(AggTy)*8); 539 if (NTy == 0 || MergeInType(NTy, UsedType)) return 0; 540 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial); 541 if (SubTy == 0) return 0; 542 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType)) 543 return 0; 544 continue; // Everything looks ok 545 } 546 return 0; 547 } else { 548 // Cannot handle this! 549 return 0; 550 } 551 } 552 553 return UsedType; 554} 555 556/// ConvertToScalar - The specified alloca passes the CanConvertToScalar 557/// predicate and is non-trivial. Convert it to something that can be trivially 558/// promoted into a register by mem2reg. 559void SROA::ConvertToScalar(AllocationInst *AI, const Type *ActualTy) { 560 DEBUG(std::cerr << "CONVERT TO SCALAR: " << *AI << " TYPE = " 561 << *ActualTy << "\n"); 562 ++NumConverted; 563 564 BasicBlock *EntryBlock = AI->getParent(); 565 assert(EntryBlock == &EntryBlock->getParent()->front() && 566 "Not in the entry block!"); 567 EntryBlock->getInstList().remove(AI); // Take the alloca out of the program. 568 569 if (ActualTy->isInteger()) 570 ActualTy = ActualTy->getUnsignedVersion(); 571 572 // Create and insert the alloca. 573 AllocaInst *NewAI = new AllocaInst(ActualTy, 0, AI->getName(), 574 EntryBlock->begin()); 575 ConvertUsesToScalar(AI, NewAI, 0); 576 delete AI; 577} 578 579 580/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 581/// directly. This happens when we are converting an "integer union" to a 582/// single integer scalar, or when we are converting a "vector union" to a 583/// vector with insert/extractelement instructions. 584/// 585/// Offset is an offset from the original alloca, in bits that need to be 586/// shifted to the right. By the end of this, there should be no uses of Ptr. 587void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset) { 588 bool isVectorInsert = isa<PackedType>(NewAI->getType()->getElementType()); 589 while (!Ptr->use_empty()) { 590 Instruction *User = cast<Instruction>(Ptr->use_back()); 591 592 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 593 // The load is a bit extract from NewAI shifted right by Offset bits. 594 Value *NV = new LoadInst(NewAI, LI->getName(), LI); 595 if (NV->getType() != LI->getType()) { 596 if (const PackedType *PTy = dyn_cast<PackedType>(NV->getType())) { 597 // Must be an element access. 598 unsigned Elt = Offset/PTy->getElementType()->getPrimitiveSizeInBits(); 599 NV = new ExtractElementInst(NV, ConstantUInt::get(Type::UIntTy, Elt), 600 "tmp", LI); 601 } else { 602 assert(NV->getType()->isInteger() && "Unknown promotion!"); 603 if (Offset && Offset < NV->getType()->getPrimitiveSizeInBits()) 604 NV = new ShiftInst(Instruction::Shr, NV, 605 ConstantUInt::get(Type::UByteTy, Offset), 606 LI->getName(), LI); 607 NV = new CastInst(NV, LI->getType(), LI->getName(), LI); 608 } 609 } 610 LI->replaceAllUsesWith(NV); 611 LI->eraseFromParent(); 612 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 613 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 614 615 // Convert the stored type to the actual type, shift it left to insert 616 // then 'or' into place. 617 Value *SV = SI->getOperand(0); 618 const Type *AllocaType = NewAI->getType()->getElementType(); 619 if (SV->getType() != AllocaType) { 620 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI); 621 622 if (const PackedType *PTy = dyn_cast<PackedType>(AllocaType)) { 623 // Must be an element insertion. 624 unsigned Elt = Offset/PTy->getElementType()->getPrimitiveSizeInBits(); 625 SV = new InsertElementInst(Old, SV, 626 ConstantUInt::get(Type::UIntTy, Elt), 627 "tmp", SI); 628 } else { 629 // If SV is signed, convert it to unsigned, so that the next cast zero 630 // extends the value. 631 if (SV->getType()->isSigned()) 632 SV = new CastInst(SV, SV->getType()->getUnsignedVersion(), 633 SV->getName(), SI); 634 SV = new CastInst(SV, Old->getType(), SV->getName(), SI); 635 if (Offset && Offset < SV->getType()->getPrimitiveSizeInBits()) 636 SV = new ShiftInst(Instruction::Shl, SV, 637 ConstantUInt::get(Type::UByteTy, Offset), 638 SV->getName()+".adj", SI); 639 // Mask out the bits we are about to insert from the old value. 640 unsigned TotalBits = SV->getType()->getPrimitiveSizeInBits(); 641 unsigned InsertBits = 642 SI->getOperand(0)->getType()->getPrimitiveSizeInBits(); 643 if (TotalBits != InsertBits) { 644 assert(TotalBits > InsertBits); 645 uint64_t Mask = ~(((1ULL << InsertBits)-1) << Offset); 646 if (TotalBits != 64) 647 Mask = Mask & ((1ULL << TotalBits)-1); 648 Old = BinaryOperator::createAnd(Old, 649 ConstantUInt::get(Old->getType(), Mask), 650 Old->getName()+".mask", SI); 651 SV = BinaryOperator::createOr(Old, SV, SV->getName()+".ins", SI); 652 } 653 } 654 } 655 new StoreInst(SV, NewAI, SI); 656 SI->eraseFromParent(); 657 658 } else if (CastInst *CI = dyn_cast<CastInst>(User)) { 659 unsigned NewOff = Offset; 660 const TargetData &TD = getAnalysis<TargetData>(); 661 if (TD.isBigEndian() && !isVectorInsert) { 662 // Adjust the pointer. For example, storing 16-bits into a 32-bit 663 // alloca with just a cast makes it modify the top 16-bits. 664 const Type *SrcTy = cast<PointerType>(Ptr->getType())->getElementType(); 665 const Type *DstTy = cast<PointerType>(CI->getType())->getElementType(); 666 int PtrDiffBits = TD.getTypeSize(SrcTy)*8-TD.getTypeSize(DstTy)*8; 667 NewOff += PtrDiffBits; 668 } 669 ConvertUsesToScalar(CI, NewAI, NewOff); 670 CI->eraseFromParent(); 671 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 672 const PointerType *AggPtrTy = 673 cast<PointerType>(GEP->getOperand(0)->getType()); 674 const TargetData &TD = getAnalysis<TargetData>(); 675 unsigned AggSizeInBits = TD.getTypeSize(AggPtrTy->getElementType())*8; 676 677 // Check to see if this is stepping over an element: GEP Ptr, int C 678 unsigned NewOffset = Offset; 679 if (GEP->getNumOperands() == 2) { 680 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getRawValue(); 681 unsigned BitOffset = Idx*AggSizeInBits; 682 683 if (TD.isLittleEndian() || isVectorInsert) 684 NewOffset += BitOffset; 685 else 686 NewOffset -= BitOffset; 687 688 } else if (GEP->getNumOperands() == 3) { 689 // We know that operand #2 is zero. 690 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getRawValue(); 691 const Type *AggTy = AggPtrTy->getElementType(); 692 if (const SequentialType *SeqTy = dyn_cast<SequentialType>(AggTy)) { 693 unsigned ElSizeBits = TD.getTypeSize(SeqTy->getElementType())*8; 694 695 if (TD.isLittleEndian() || isVectorInsert) 696 NewOffset += ElSizeBits*Idx; 697 else 698 NewOffset += AggSizeInBits-ElSizeBits*(Idx+1); 699 } else if (const StructType *STy = dyn_cast<StructType>(AggTy)) { 700 unsigned EltBitOffset = TD.getStructLayout(STy)->MemberOffsets[Idx]*8; 701 702 if (TD.isLittleEndian() || isVectorInsert) 703 NewOffset += EltBitOffset; 704 else { 705 const PointerType *ElPtrTy = cast<PointerType>(GEP->getType()); 706 unsigned ElSizeBits = TD.getTypeSize(ElPtrTy->getElementType())*8; 707 NewOffset += AggSizeInBits-(EltBitOffset+ElSizeBits); 708 } 709 710 } else { 711 assert(0 && "Unsupported operation!"); 712 abort(); 713 } 714 } else { 715 assert(0 && "Unsupported operation!"); 716 abort(); 717 } 718 ConvertUsesToScalar(GEP, NewAI, NewOffset); 719 GEP->eraseFromParent(); 720 } else { 721 assert(0 && "Unsupported operation!"); 722 abort(); 723 } 724 } 725} 726