ScalarReplAggregates.cpp revision d6bf5cf6415d1d12d66257e640471958920601b5
1//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 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 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#define DEBUG_TYPE "scalarrepl" 23#include "llvm/Transforms/Scalar.h" 24#include "llvm/Constants.h" 25#include "llvm/DerivedTypes.h" 26#include "llvm/Function.h" 27#include "llvm/GlobalVariable.h" 28#include "llvm/Instructions.h" 29#include "llvm/IntrinsicInst.h" 30#include "llvm/LLVMContext.h" 31#include "llvm/Pass.h" 32#include "llvm/Analysis/Dominators.h" 33#include "llvm/Target/TargetData.h" 34#include "llvm/Transforms/Utils/PromoteMemToReg.h" 35#include "llvm/Transforms/Utils/Local.h" 36#include "llvm/Support/Debug.h" 37#include "llvm/Support/ErrorHandling.h" 38#include "llvm/Support/GetElementPtrTypeIterator.h" 39#include "llvm/Support/IRBuilder.h" 40#include "llvm/Support/MathExtras.h" 41#include "llvm/Support/raw_ostream.h" 42#include "llvm/ADT/SmallVector.h" 43#include "llvm/ADT/Statistic.h" 44using namespace llvm; 45 46STATISTIC(NumReplaced, "Number of allocas broken up"); 47STATISTIC(NumPromoted, "Number of allocas promoted"); 48STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 49STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 50 51namespace { 52 struct SROA : public FunctionPass { 53 static char ID; // Pass identification, replacement for typeid 54 explicit SROA(signed T = -1) : FunctionPass(&ID) { 55 if (T == -1) 56 SRThreshold = 128; 57 else 58 SRThreshold = T; 59 } 60 61 bool runOnFunction(Function &F); 62 63 bool performScalarRepl(Function &F); 64 bool performPromotion(Function &F); 65 66 // getAnalysisUsage - This pass does not require any passes, but we know it 67 // will not alter the CFG, so say so. 68 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 69 AU.addRequired<DominatorTree>(); 70 AU.addRequired<DominanceFrontier>(); 71 AU.setPreservesCFG(); 72 } 73 74 private: 75 TargetData *TD; 76 77 /// DeadInsts - Keep track of instructions we have made dead, so that 78 /// we can remove them after we are done working. 79 SmallVector<Value*, 32> DeadInsts; 80 81 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 82 /// information about the uses. All these fields are initialized to false 83 /// and set to true when something is learned. 84 struct AllocaInfo { 85 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 86 bool isUnsafe : 1; 87 88 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 89 bool isMemCpySrc : 1; 90 91 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 92 bool isMemCpyDst : 1; 93 94 AllocaInfo() 95 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false) {} 96 }; 97 98 unsigned SRThreshold; 99 100 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 101 102 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 103 104 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 105 AllocaInfo &Info); 106 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset, 107 AllocaInfo &Info); 108 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 109 const Type *MemOpType, bool isStore, AllocaInfo &Info); 110 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 111 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 112 const Type *&IdxTy); 113 114 void DoScalarReplacement(AllocaInst *AI, 115 std::vector<AllocaInst*> &WorkList); 116 void DeleteDeadInstructions(); 117 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocaInst *Base); 118 119 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 120 SmallVector<AllocaInst*, 32> &NewElts); 121 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 122 SmallVector<AllocaInst*, 32> &NewElts); 123 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 124 SmallVector<AllocaInst*, 32> &NewElts); 125 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 126 AllocaInst *AI, 127 SmallVector<AllocaInst*, 32> &NewElts); 128 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 129 SmallVector<AllocaInst*, 32> &NewElts); 130 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 131 SmallVector<AllocaInst*, 32> &NewElts); 132 133 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 134 }; 135} 136 137char SROA::ID = 0; 138static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 139 140// Public interface to the ScalarReplAggregates pass 141FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 142 return new SROA(Threshold); 143} 144 145 146//===----------------------------------------------------------------------===// 147// Convert To Scalar Optimization. 148//===----------------------------------------------------------------------===// 149 150namespace { 151/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 152/// optimization, which scans the uses of an alloca and determines if it can 153/// rewrite it in terms of a single new alloca that can be mem2reg'd. 154class ConvertToScalarInfo { 155 /// AllocaSize - The size of the alloca being considered. 156 unsigned AllocaSize; 157 const TargetData &TD; 158 159 /// IsNotTrivial - This is set to true if there is some access to the object 160 /// which means that mem2reg can't promote it. 161 bool IsNotTrivial; 162 163 /// VectorTy - This tracks the type that we should promote the vector to if 164 /// it is possible to turn it into a vector. This starts out null, and if it 165 /// isn't possible to turn into a vector type, it gets set to VoidTy. 166 const Type *VectorTy; 167 168 /// HadAVector - True if there is at least one vector access to the alloca. 169 /// We don't want to turn random arrays into vectors and use vector element 170 /// insert/extract, but if there are element accesses to something that is 171 /// also declared as a vector, we do want to promote to a vector. 172 bool HadAVector; 173 174public: 175 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 176 : AllocaSize(Size), TD(td) { 177 IsNotTrivial = false; 178 VectorTy = 0; 179 HadAVector = false; 180 } 181 182 AllocaInst *TryConvert(AllocaInst *AI); 183 184private: 185 bool CanConvertToScalar(Value *V, uint64_t Offset); 186 void MergeInType(const Type *In, uint64_t Offset); 187 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 188 189 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 190 uint64_t Offset, IRBuilder<> &Builder); 191 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 192 uint64_t Offset, IRBuilder<> &Builder); 193}; 194} // end anonymous namespace. 195 196/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 197/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 198/// alloca if possible or null if not. 199AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 200 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 201 // out. 202 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 203 return 0; 204 205 // If we were able to find a vector type that can handle this with 206 // insert/extract elements, and if there was at least one use that had 207 // a vector type, promote this to a vector. We don't want to promote 208 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 209 // we just get a lot of insert/extracts. If at least one vector is 210 // involved, then we probably really do have a union of vector/array. 211 const Type *NewTy; 212 if (VectorTy && VectorTy->isVectorTy() && HadAVector) { 213 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 214 << *VectorTy << '\n'); 215 NewTy = VectorTy; // Use the vector type. 216 } else { 217 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 218 // Create and insert the integer alloca. 219 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 220 } 221 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 222 ConvertUsesToScalar(AI, NewAI, 0); 223 return NewAI; 224} 225 226/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy) 227/// so far at the offset specified by Offset (which is specified in bytes). 228/// 229/// There are two cases we handle here: 230/// 1) A union of vector types of the same size and potentially its elements. 231/// Here we turn element accesses into insert/extract element operations. 232/// This promotes a <4 x float> with a store of float to the third element 233/// into a <4 x float> that uses insert element. 234/// 2) A fully general blob of memory, which we turn into some (potentially 235/// large) integer type with extract and insert operations where the loads 236/// and stores would mutate the memory. We mark this by setting VectorTy 237/// to VoidTy. 238void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) { 239 // If we already decided to turn this into a blob of integer memory, there is 240 // nothing to be done. 241 if (VectorTy && VectorTy->isVoidTy()) 242 return; 243 244 // If this could be contributing to a vector, analyze it. 245 246 // If the In type is a vector that is the same size as the alloca, see if it 247 // matches the existing VecTy. 248 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 249 // Remember if we saw a vector type. 250 HadAVector = true; 251 252 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 253 // If we're storing/loading a vector of the right size, allow it as a 254 // vector. If this the first vector we see, remember the type so that 255 // we know the element size. If this is a subsequent access, ignore it 256 // even if it is a differing type but the same size. Worst case we can 257 // bitcast the resultant vectors. 258 if (VectorTy == 0) 259 VectorTy = VInTy; 260 return; 261 } 262 } else if (In->isFloatTy() || In->isDoubleTy() || 263 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 264 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 265 // If we're accessing something that could be an element of a vector, see 266 // if the implied vector agrees with what we already have and if Offset is 267 // compatible with it. 268 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 269 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 270 (VectorTy == 0 || 271 cast<VectorType>(VectorTy)->getElementType() 272 ->getPrimitiveSizeInBits()/8 == EltSize)) { 273 if (VectorTy == 0) 274 VectorTy = VectorType::get(In, AllocaSize/EltSize); 275 return; 276 } 277 } 278 279 // Otherwise, we have a case that we can't handle with an optimized vector 280 // form. We can still turn this into a large integer. 281 VectorTy = Type::getVoidTy(In->getContext()); 282} 283 284/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 285/// its accesses to a single vector type, return true and set VecTy to 286/// the new type. If we could convert the alloca into a single promotable 287/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 288/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 289/// is the current offset from the base of the alloca being analyzed. 290/// 291/// If we see at least one access to the value that is as a vector type, set the 292/// SawVec flag. 293bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 294 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 295 Instruction *User = cast<Instruction>(*UI); 296 297 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 298 // Don't break volatile loads. 299 if (LI->isVolatile()) 300 return false; 301 MergeInType(LI->getType(), Offset); 302 continue; 303 } 304 305 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 306 // Storing the pointer, not into the value? 307 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 308 MergeInType(SI->getOperand(0)->getType(), Offset); 309 continue; 310 } 311 312 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 313 IsNotTrivial = true; // Can't be mem2reg'd. 314 if (!CanConvertToScalar(BCI, Offset)) 315 return false; 316 continue; 317 } 318 319 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 320 // If this is a GEP with a variable indices, we can't handle it. 321 if (!GEP->hasAllConstantIndices()) 322 return false; 323 324 // Compute the offset that this GEP adds to the pointer. 325 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 326 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 327 &Indices[0], Indices.size()); 328 // See if all uses can be converted. 329 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 330 return false; 331 IsNotTrivial = true; // Can't be mem2reg'd. 332 continue; 333 } 334 335 // If this is a constant sized memset of a constant value (e.g. 0) we can 336 // handle it. 337 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 338 // Store of constant value and constant size. 339 if (!isa<ConstantInt>(MSI->getValue()) || 340 !isa<ConstantInt>(MSI->getLength())) 341 return false; 342 IsNotTrivial = true; // Can't be mem2reg'd. 343 continue; 344 } 345 346 // If this is a memcpy or memmove into or out of the whole allocation, we 347 // can handle it like a load or store of the scalar type. 348 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 349 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 350 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 351 return false; 352 353 IsNotTrivial = true; // Can't be mem2reg'd. 354 continue; 355 } 356 357 // Otherwise, we cannot handle this! 358 return false; 359 } 360 361 return true; 362} 363 364/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 365/// directly. This happens when we are converting an "integer union" to a 366/// single integer scalar, or when we are converting a "vector union" to a 367/// vector with insert/extractelement instructions. 368/// 369/// Offset is an offset from the original alloca, in bits that need to be 370/// shifted to the right. By the end of this, there should be no uses of Ptr. 371void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 372 uint64_t Offset) { 373 while (!Ptr->use_empty()) { 374 Instruction *User = cast<Instruction>(Ptr->use_back()); 375 376 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 377 ConvertUsesToScalar(CI, NewAI, Offset); 378 CI->eraseFromParent(); 379 continue; 380 } 381 382 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 383 // Compute the offset that this GEP adds to the pointer. 384 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 385 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 386 &Indices[0], Indices.size()); 387 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 388 GEP->eraseFromParent(); 389 continue; 390 } 391 392 IRBuilder<> Builder(User->getParent(), User); 393 394 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 395 // The load is a bit extract from NewAI shifted right by Offset bits. 396 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 397 Value *NewLoadVal 398 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 399 LI->replaceAllUsesWith(NewLoadVal); 400 LI->eraseFromParent(); 401 continue; 402 } 403 404 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 405 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 406 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 407 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 408 Builder); 409 Builder.CreateStore(New, NewAI); 410 SI->eraseFromParent(); 411 412 // If the load we just inserted is now dead, then the inserted store 413 // overwrote the entire thing. 414 if (Old->use_empty()) 415 Old->eraseFromParent(); 416 continue; 417 } 418 419 // If this is a constant sized memset of a constant value (e.g. 0) we can 420 // transform it into a store of the expanded constant value. 421 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 422 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 423 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 424 if (NumBytes != 0) { 425 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 426 427 // Compute the value replicated the right number of times. 428 APInt APVal(NumBytes*8, Val); 429 430 // Splat the value if non-zero. 431 if (Val) 432 for (unsigned i = 1; i != NumBytes; ++i) 433 APVal |= APVal << 8; 434 435 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 436 Value *New = ConvertScalar_InsertValue( 437 ConstantInt::get(User->getContext(), APVal), 438 Old, Offset, Builder); 439 Builder.CreateStore(New, NewAI); 440 441 // If the load we just inserted is now dead, then the memset overwrote 442 // the entire thing. 443 if (Old->use_empty()) 444 Old->eraseFromParent(); 445 } 446 MSI->eraseFromParent(); 447 continue; 448 } 449 450 // If this is a memcpy or memmove into or out of the whole allocation, we 451 // can handle it like a load or store of the scalar type. 452 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 453 assert(Offset == 0 && "must be store to start of alloca"); 454 455 // If the source and destination are both to the same alloca, then this is 456 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 457 // as appropriate. 458 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject(0)); 459 460 if (MTI->getSource()->getUnderlyingObject(0) != OrigAI) { 461 // Dest must be OrigAI, change this to be a load from the original 462 // pointer (bitcasted), then a store to our new alloca. 463 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 464 Value *SrcPtr = MTI->getSource(); 465 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 466 467 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 468 SrcVal->setAlignment(MTI->getAlignment()); 469 Builder.CreateStore(SrcVal, NewAI); 470 } else if (MTI->getDest()->getUnderlyingObject(0) != OrigAI) { 471 // Src must be OrigAI, change this to be a load from NewAI then a store 472 // through the original dest pointer (bitcasted). 473 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 474 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 475 476 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 477 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 478 NewStore->setAlignment(MTI->getAlignment()); 479 } else { 480 // Noop transfer. Src == Dst 481 } 482 483 MTI->eraseFromParent(); 484 continue; 485 } 486 487 llvm_unreachable("Unsupported operation!"); 488 } 489} 490 491/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 492/// or vector value FromVal, extracting the bits from the offset specified by 493/// Offset. This returns the value, which is of type ToType. 494/// 495/// This happens when we are converting an "integer union" to a single 496/// integer scalar, or when we are converting a "vector union" to a vector with 497/// insert/extractelement instructions. 498/// 499/// Offset is an offset from the original alloca, in bits that need to be 500/// shifted to the right. 501Value *ConvertToScalarInfo:: 502ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 503 uint64_t Offset, IRBuilder<> &Builder) { 504 // If the load is of the whole new alloca, no conversion is needed. 505 if (FromVal->getType() == ToType && Offset == 0) 506 return FromVal; 507 508 // If the result alloca is a vector type, this is either an element 509 // access or a bitcast to another vector type of the same size. 510 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 511 if (ToType->isVectorTy()) 512 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 513 514 // Otherwise it must be an element access. 515 unsigned Elt = 0; 516 if (Offset) { 517 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 518 Elt = Offset/EltSize; 519 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 520 } 521 // Return the element extracted out of it. 522 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 523 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 524 if (V->getType() != ToType) 525 V = Builder.CreateBitCast(V, ToType, "tmp"); 526 return V; 527 } 528 529 // If ToType is a first class aggregate, extract out each of the pieces and 530 // use insertvalue's to form the FCA. 531 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 532 const StructLayout &Layout = *TD.getStructLayout(ST); 533 Value *Res = UndefValue::get(ST); 534 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 535 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 536 Offset+Layout.getElementOffsetInBits(i), 537 Builder); 538 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 539 } 540 return Res; 541 } 542 543 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 544 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 545 Value *Res = UndefValue::get(AT); 546 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 547 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 548 Offset+i*EltSize, Builder); 549 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 550 } 551 return Res; 552 } 553 554 // Otherwise, this must be a union that was converted to an integer value. 555 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 556 557 // If this is a big-endian system and the load is narrower than the 558 // full alloca type, we need to do a shift to get the right bits. 559 int ShAmt = 0; 560 if (TD.isBigEndian()) { 561 // On big-endian machines, the lowest bit is stored at the bit offset 562 // from the pointer given by getTypeStoreSizeInBits. This matters for 563 // integers with a bitwidth that is not a multiple of 8. 564 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 565 TD.getTypeStoreSizeInBits(ToType) - Offset; 566 } else { 567 ShAmt = Offset; 568 } 569 570 // Note: we support negative bitwidths (with shl) which are not defined. 571 // We do this to support (f.e.) loads off the end of a structure where 572 // only some bits are used. 573 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 574 FromVal = Builder.CreateLShr(FromVal, 575 ConstantInt::get(FromVal->getType(), 576 ShAmt), "tmp"); 577 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 578 FromVal = Builder.CreateShl(FromVal, 579 ConstantInt::get(FromVal->getType(), 580 -ShAmt), "tmp"); 581 582 // Finally, unconditionally truncate the integer to the right width. 583 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 584 if (LIBitWidth < NTy->getBitWidth()) 585 FromVal = 586 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 587 LIBitWidth), "tmp"); 588 else if (LIBitWidth > NTy->getBitWidth()) 589 FromVal = 590 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 591 LIBitWidth), "tmp"); 592 593 // If the result is an integer, this is a trunc or bitcast. 594 if (ToType->isIntegerTy()) { 595 // Should be done. 596 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 597 // Just do a bitcast, we know the sizes match up. 598 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 599 } else { 600 // Otherwise must be a pointer. 601 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 602 } 603 assert(FromVal->getType() == ToType && "Didn't convert right?"); 604 return FromVal; 605} 606 607/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 608/// or vector value "Old" at the offset specified by Offset. 609/// 610/// This happens when we are converting an "integer union" to a 611/// single integer scalar, or when we are converting a "vector union" to a 612/// vector with insert/extractelement instructions. 613/// 614/// Offset is an offset from the original alloca, in bits that need to be 615/// shifted to the right. 616Value *ConvertToScalarInfo:: 617ConvertScalar_InsertValue(Value *SV, Value *Old, 618 uint64_t Offset, IRBuilder<> &Builder) { 619 // Convert the stored type to the actual type, shift it left to insert 620 // then 'or' into place. 621 const Type *AllocaType = Old->getType(); 622 LLVMContext &Context = Old->getContext(); 623 624 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 625 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 626 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 627 628 // Changing the whole vector with memset or with an access of a different 629 // vector type? 630 if (ValSize == VecSize) 631 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 632 633 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 634 635 // Must be an element insertion. 636 unsigned Elt = Offset/EltSize; 637 638 if (SV->getType() != VTy->getElementType()) 639 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 640 641 SV = Builder.CreateInsertElement(Old, SV, 642 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 643 "tmp"); 644 return SV; 645 } 646 647 // If SV is a first-class aggregate value, insert each value recursively. 648 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 649 const StructLayout &Layout = *TD.getStructLayout(ST); 650 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 651 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 652 Old = ConvertScalar_InsertValue(Elt, Old, 653 Offset+Layout.getElementOffsetInBits(i), 654 Builder); 655 } 656 return Old; 657 } 658 659 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 660 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 661 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 662 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 663 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 664 } 665 return Old; 666 } 667 668 // If SV is a float, convert it to the appropriate integer type. 669 // If it is a pointer, do the same. 670 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 671 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 672 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 673 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 674 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 675 SV = Builder.CreateBitCast(SV, 676 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 677 else if (SV->getType()->isPointerTy()) 678 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 679 680 // Zero extend or truncate the value if needed. 681 if (SV->getType() != AllocaType) { 682 if (SV->getType()->getPrimitiveSizeInBits() < 683 AllocaType->getPrimitiveSizeInBits()) 684 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 685 else { 686 // Truncation may be needed if storing more than the alloca can hold 687 // (undefined behavior). 688 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 689 SrcWidth = DestWidth; 690 SrcStoreWidth = DestStoreWidth; 691 } 692 } 693 694 // If this is a big-endian system and the store is narrower than the 695 // full alloca type, we need to do a shift to get the right bits. 696 int ShAmt = 0; 697 if (TD.isBigEndian()) { 698 // On big-endian machines, the lowest bit is stored at the bit offset 699 // from the pointer given by getTypeStoreSizeInBits. This matters for 700 // integers with a bitwidth that is not a multiple of 8. 701 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 702 } else { 703 ShAmt = Offset; 704 } 705 706 // Note: we support negative bitwidths (with shr) which are not defined. 707 // We do this to support (f.e.) stores off the end of a structure where 708 // only some bits in the structure are set. 709 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 710 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 711 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 712 ShAmt), "tmp"); 713 Mask <<= ShAmt; 714 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 715 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 716 -ShAmt), "tmp"); 717 Mask = Mask.lshr(-ShAmt); 718 } 719 720 // Mask out the bits we are about to insert from the old value, and or 721 // in the new bits. 722 if (SrcWidth != DestWidth) { 723 assert(DestWidth > SrcWidth); 724 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 725 SV = Builder.CreateOr(Old, SV, "ins"); 726 } 727 return SV; 728} 729 730 731//===----------------------------------------------------------------------===// 732// SRoA Driver 733//===----------------------------------------------------------------------===// 734 735 736bool SROA::runOnFunction(Function &F) { 737 TD = getAnalysisIfAvailable<TargetData>(); 738 739 bool Changed = performPromotion(F); 740 741 // FIXME: ScalarRepl currently depends on TargetData more than it 742 // theoretically needs to. It should be refactored in order to support 743 // target-independent IR. Until this is done, just skip the actual 744 // scalar-replacement portion of this pass. 745 if (!TD) return Changed; 746 747 while (1) { 748 bool LocalChange = performScalarRepl(F); 749 if (!LocalChange) break; // No need to repromote if no scalarrepl 750 Changed = true; 751 LocalChange = performPromotion(F); 752 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 753 } 754 755 return Changed; 756} 757 758 759bool SROA::performPromotion(Function &F) { 760 std::vector<AllocaInst*> Allocas; 761 DominatorTree &DT = getAnalysis<DominatorTree>(); 762 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 763 764 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 765 766 bool Changed = false; 767 768 while (1) { 769 Allocas.clear(); 770 771 // Find allocas that are safe to promote, by looking at all instructions in 772 // the entry node 773 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 774 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 775 if (isAllocaPromotable(AI)) 776 Allocas.push_back(AI); 777 778 if (Allocas.empty()) break; 779 780 PromoteMemToReg(Allocas, DT, DF); 781 NumPromoted += Allocas.size(); 782 Changed = true; 783 } 784 785 return Changed; 786} 787 788 789/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 790/// SROA. It must be a struct or array type with a small number of elements. 791static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 792 const Type *T = AI->getAllocatedType(); 793 // Do not promote any struct into more than 32 separate vars. 794 if (const StructType *ST = dyn_cast<StructType>(T)) 795 return ST->getNumElements() <= 32; 796 // Arrays are much less likely to be safe for SROA; only consider 797 // them if they are very small. 798 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 799 return AT->getNumElements() <= 8; 800 return false; 801} 802 803 804// performScalarRepl - This algorithm is a simple worklist driven algorithm, 805// which runs on all of the malloc/alloca instructions in the function, removing 806// them if they are only used by getelementptr instructions. 807// 808bool SROA::performScalarRepl(Function &F) { 809 std::vector<AllocaInst*> WorkList; 810 811 // Scan the entry basic block, adding allocas to the worklist. 812 BasicBlock &BB = F.getEntryBlock(); 813 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 814 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 815 WorkList.push_back(A); 816 817 // Process the worklist 818 bool Changed = false; 819 while (!WorkList.empty()) { 820 AllocaInst *AI = WorkList.back(); 821 WorkList.pop_back(); 822 823 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 824 // with unused elements. 825 if (AI->use_empty()) { 826 AI->eraseFromParent(); 827 Changed = true; 828 continue; 829 } 830 831 // If this alloca is impossible for us to promote, reject it early. 832 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 833 continue; 834 835 // Check to see if this allocation is only modified by a memcpy/memmove from 836 // a constant global. If this is the case, we can change all users to use 837 // the constant global instead. This is commonly produced by the CFE by 838 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 839 // is only subsequently read. 840 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 841 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 842 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 843 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 844 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 845 TheCopy->eraseFromParent(); // Don't mutate the global. 846 AI->eraseFromParent(); 847 ++NumGlobals; 848 Changed = true; 849 continue; 850 } 851 852 // Check to see if we can perform the core SROA transformation. We cannot 853 // transform the allocation instruction if it is an array allocation 854 // (allocations OF arrays are ok though), and an allocation of a scalar 855 // value cannot be decomposed at all. 856 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 857 858 // Do not promote [0 x %struct]. 859 if (AllocaSize == 0) continue; 860 861 // Do not promote any struct whose size is too big. 862 if (AllocaSize > SRThreshold) continue; 863 864 // If the alloca looks like a good candidate for scalar replacement, and if 865 // all its users can be transformed, then split up the aggregate into its 866 // separate elements. 867 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 868 DoScalarReplacement(AI, WorkList); 869 Changed = true; 870 continue; 871 } 872 873 // If we can turn this aggregate value (potentially with casts) into a 874 // simple scalar value that can be mem2reg'd into a register value. 875 // IsNotTrivial tracks whether this is something that mem2reg could have 876 // promoted itself. If so, we don't want to transform it needlessly. Note 877 // that we can't just check based on the type: the alloca may be of an i32 878 // but that has pointer arithmetic to set byte 3 of it or something. 879 if (AllocaInst *NewAI = 880 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 881 NewAI->takeName(AI); 882 AI->eraseFromParent(); 883 ++NumConverted; 884 Changed = true; 885 continue; 886 } 887 888 // Otherwise, couldn't process this alloca. 889 } 890 891 return Changed; 892} 893 894/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 895/// predicate, do SROA now. 896void SROA::DoScalarReplacement(AllocaInst *AI, 897 std::vector<AllocaInst*> &WorkList) { 898 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 899 SmallVector<AllocaInst*, 32> ElementAllocas; 900 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 901 ElementAllocas.reserve(ST->getNumContainedTypes()); 902 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 903 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 904 AI->getAlignment(), 905 AI->getName() + "." + Twine(i), AI); 906 ElementAllocas.push_back(NA); 907 WorkList.push_back(NA); // Add to worklist for recursive processing 908 } 909 } else { 910 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 911 ElementAllocas.reserve(AT->getNumElements()); 912 const Type *ElTy = AT->getElementType(); 913 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 914 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 915 AI->getName() + "." + Twine(i), AI); 916 ElementAllocas.push_back(NA); 917 WorkList.push_back(NA); // Add to worklist for recursive processing 918 } 919 } 920 921 // Now that we have created the new alloca instructions, rewrite all the 922 // uses of the old alloca. 923 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 924 925 // Now erase any instructions that were made dead while rewriting the alloca. 926 DeleteDeadInstructions(); 927 AI->eraseFromParent(); 928 929 ++NumReplaced; 930} 931 932/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 933/// recursively including all their operands that become trivially dead. 934void SROA::DeleteDeadInstructions() { 935 while (!DeadInsts.empty()) { 936 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 937 938 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 939 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 940 // Zero out the operand and see if it becomes trivially dead. 941 // (But, don't add allocas to the dead instruction list -- they are 942 // already on the worklist and will be deleted separately.) 943 *OI = 0; 944 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 945 DeadInsts.push_back(U); 946 } 947 948 I->eraseFromParent(); 949 } 950} 951 952/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 953/// performing scalar replacement of alloca AI. The results are flagged in 954/// the Info parameter. Offset indicates the position within AI that is 955/// referenced by this instruction. 956void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 957 AllocaInfo &Info) { 958 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 959 Instruction *User = cast<Instruction>(*UI); 960 961 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 962 isSafeForScalarRepl(BC, AI, Offset, Info); 963 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 964 uint64_t GEPOffset = Offset; 965 isSafeGEP(GEPI, AI, GEPOffset, Info); 966 if (!Info.isUnsafe) 967 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 968 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 969 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 970 if (Length) 971 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 972 UI.getOperandNo() == CallInst::ArgOffset, Info); 973 else 974 MarkUnsafe(Info); 975 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 976 if (!LI->isVolatile()) { 977 const Type *LIType = LI->getType(); 978 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 979 LIType, false, Info); 980 } else 981 MarkUnsafe(Info); 982 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 983 // Store is ok if storing INTO the pointer, not storing the pointer 984 if (!SI->isVolatile() && SI->getOperand(0) != I) { 985 const Type *SIType = SI->getOperand(0)->getType(); 986 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 987 SIType, true, Info); 988 } else 989 MarkUnsafe(Info); 990 } else { 991 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 992 MarkUnsafe(Info); 993 } 994 if (Info.isUnsafe) return; 995 } 996} 997 998/// isSafeGEP - Check if a GEP instruction can be handled for scalar 999/// replacement. It is safe when all the indices are constant, in-bounds 1000/// references, and when the resulting offset corresponds to an element within 1001/// the alloca type. The results are flagged in the Info parameter. Upon 1002/// return, Offset is adjusted as specified by the GEP indices. 1003void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 1004 uint64_t &Offset, AllocaInfo &Info) { 1005 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1006 if (GEPIt == E) 1007 return; 1008 1009 // Walk through the GEP type indices, checking the types that this indexes 1010 // into. 1011 for (; GEPIt != E; ++GEPIt) { 1012 // Ignore struct elements, no extra checking needed for these. 1013 if ((*GEPIt)->isStructTy()) 1014 continue; 1015 1016 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1017 if (!IdxVal) 1018 return MarkUnsafe(Info); 1019 } 1020 1021 // Compute the offset due to this GEP and check if the alloca has a 1022 // component element at that offset. 1023 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1024 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1025 &Indices[0], Indices.size()); 1026 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 1027 MarkUnsafe(Info); 1028} 1029 1030/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1031/// alloca or has an offset and size that corresponds to a component element 1032/// within it. The offset checked here may have been formed from a GEP with a 1033/// pointer bitcasted to a different type. 1034void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 1035 const Type *MemOpType, bool isStore, 1036 AllocaInfo &Info) { 1037 // Check if this is a load/store of the entire alloca. 1038 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 1039 bool UsesAggregateType = (MemOpType == AI->getAllocatedType()); 1040 // This is safe for MemIntrinsics (where MemOpType is 0), integer types 1041 // (which are essentially the same as the MemIntrinsics, especially with 1042 // regard to copying padding between elements), or references using the 1043 // aggregate type of the alloca. 1044 if (!MemOpType || MemOpType->isIntegerTy() || UsesAggregateType) { 1045 if (!UsesAggregateType) { 1046 if (isStore) 1047 Info.isMemCpyDst = true; 1048 else 1049 Info.isMemCpySrc = true; 1050 } 1051 return; 1052 } 1053 } 1054 // Check if the offset/size correspond to a component within the alloca type. 1055 const Type *T = AI->getAllocatedType(); 1056 if (TypeHasComponent(T, Offset, MemSize)) 1057 return; 1058 1059 return MarkUnsafe(Info); 1060} 1061 1062/// TypeHasComponent - Return true if T has a component type with the 1063/// specified offset and size. If Size is zero, do not check the size. 1064bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1065 const Type *EltTy; 1066 uint64_t EltSize; 1067 if (const StructType *ST = dyn_cast<StructType>(T)) { 1068 const StructLayout *Layout = TD->getStructLayout(ST); 1069 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1070 EltTy = ST->getContainedType(EltIdx); 1071 EltSize = TD->getTypeAllocSize(EltTy); 1072 Offset -= Layout->getElementOffset(EltIdx); 1073 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1074 EltTy = AT->getElementType(); 1075 EltSize = TD->getTypeAllocSize(EltTy); 1076 if (Offset >= AT->getNumElements() * EltSize) 1077 return false; 1078 Offset %= EltSize; 1079 } else { 1080 return false; 1081 } 1082 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1083 return true; 1084 // Check if the component spans multiple elements. 1085 if (Offset + Size > EltSize) 1086 return false; 1087 return TypeHasComponent(EltTy, Offset, Size); 1088} 1089 1090/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1091/// the instruction I, which references it, to use the separate elements. 1092/// Offset indicates the position within AI that is referenced by this 1093/// instruction. 1094void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1095 SmallVector<AllocaInst*, 32> &NewElts) { 1096 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1097 Instruction *User = cast<Instruction>(*UI); 1098 1099 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1100 RewriteBitCast(BC, AI, Offset, NewElts); 1101 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1102 RewriteGEP(GEPI, AI, Offset, NewElts); 1103 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1104 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1105 uint64_t MemSize = Length->getZExtValue(); 1106 if (Offset == 0 && 1107 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1108 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1109 // Otherwise the intrinsic can only touch a single element and the 1110 // address operand will be updated, so nothing else needs to be done. 1111 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1112 const Type *LIType = LI->getType(); 1113 if (LIType == AI->getAllocatedType()) { 1114 // Replace: 1115 // %res = load { i32, i32 }* %alloc 1116 // with: 1117 // %load.0 = load i32* %alloc.0 1118 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1119 // %load.1 = load i32* %alloc.1 1120 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1121 // (Also works for arrays instead of structs) 1122 Value *Insert = UndefValue::get(LIType); 1123 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1124 Value *Load = new LoadInst(NewElts[i], "load", LI); 1125 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1126 } 1127 LI->replaceAllUsesWith(Insert); 1128 DeadInsts.push_back(LI); 1129 } else if (LIType->isIntegerTy() && 1130 TD->getTypeAllocSize(LIType) == 1131 TD->getTypeAllocSize(AI->getAllocatedType())) { 1132 // If this is a load of the entire alloca to an integer, rewrite it. 1133 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1134 } 1135 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1136 Value *Val = SI->getOperand(0); 1137 const Type *SIType = Val->getType(); 1138 if (SIType == AI->getAllocatedType()) { 1139 // Replace: 1140 // store { i32, i32 } %val, { i32, i32 }* %alloc 1141 // with: 1142 // %val.0 = extractvalue { i32, i32 } %val, 0 1143 // store i32 %val.0, i32* %alloc.0 1144 // %val.1 = extractvalue { i32, i32 } %val, 1 1145 // store i32 %val.1, i32* %alloc.1 1146 // (Also works for arrays instead of structs) 1147 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1148 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1149 new StoreInst(Extract, NewElts[i], SI); 1150 } 1151 DeadInsts.push_back(SI); 1152 } else if (SIType->isIntegerTy() && 1153 TD->getTypeAllocSize(SIType) == 1154 TD->getTypeAllocSize(AI->getAllocatedType())) { 1155 // If this is a store of the entire alloca from an integer, rewrite it. 1156 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1157 } 1158 } 1159 } 1160} 1161 1162/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1163/// and recursively continue updating all of its uses. 1164void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1165 SmallVector<AllocaInst*, 32> &NewElts) { 1166 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1167 if (BC->getOperand(0) != AI) 1168 return; 1169 1170 // The bitcast references the original alloca. Replace its uses with 1171 // references to the first new element alloca. 1172 Instruction *Val = NewElts[0]; 1173 if (Val->getType() != BC->getDestTy()) { 1174 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1175 Val->takeName(BC); 1176 } 1177 BC->replaceAllUsesWith(Val); 1178 DeadInsts.push_back(BC); 1179} 1180 1181/// FindElementAndOffset - Return the index of the element containing Offset 1182/// within the specified type, which must be either a struct or an array. 1183/// Sets T to the type of the element and Offset to the offset within that 1184/// element. IdxTy is set to the type of the index result to be used in a 1185/// GEP instruction. 1186uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1187 const Type *&IdxTy) { 1188 uint64_t Idx = 0; 1189 if (const StructType *ST = dyn_cast<StructType>(T)) { 1190 const StructLayout *Layout = TD->getStructLayout(ST); 1191 Idx = Layout->getElementContainingOffset(Offset); 1192 T = ST->getContainedType(Idx); 1193 Offset -= Layout->getElementOffset(Idx); 1194 IdxTy = Type::getInt32Ty(T->getContext()); 1195 return Idx; 1196 } 1197 const ArrayType *AT = cast<ArrayType>(T); 1198 T = AT->getElementType(); 1199 uint64_t EltSize = TD->getTypeAllocSize(T); 1200 Idx = Offset / EltSize; 1201 Offset -= Idx * EltSize; 1202 IdxTy = Type::getInt64Ty(T->getContext()); 1203 return Idx; 1204} 1205 1206/// RewriteGEP - Check if this GEP instruction moves the pointer across 1207/// elements of the alloca that are being split apart, and if so, rewrite 1208/// the GEP to be relative to the new element. 1209void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1210 SmallVector<AllocaInst*, 32> &NewElts) { 1211 uint64_t OldOffset = Offset; 1212 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1213 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1214 &Indices[0], Indices.size()); 1215 1216 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1217 1218 const Type *T = AI->getAllocatedType(); 1219 const Type *IdxTy; 1220 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1221 if (GEPI->getOperand(0) == AI) 1222 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1223 1224 T = AI->getAllocatedType(); 1225 uint64_t EltOffset = Offset; 1226 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1227 1228 // If this GEP does not move the pointer across elements of the alloca 1229 // being split, then it does not needs to be rewritten. 1230 if (Idx == OldIdx) 1231 return; 1232 1233 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1234 SmallVector<Value*, 8> NewArgs; 1235 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1236 while (EltOffset != 0) { 1237 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1238 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1239 } 1240 Instruction *Val = NewElts[Idx]; 1241 if (NewArgs.size() > 1) { 1242 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1243 NewArgs.end(), "", GEPI); 1244 Val->takeName(GEPI); 1245 } 1246 if (Val->getType() != GEPI->getType()) 1247 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1248 GEPI->replaceAllUsesWith(Val); 1249 DeadInsts.push_back(GEPI); 1250} 1251 1252/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1253/// Rewrite it to copy or set the elements of the scalarized memory. 1254void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1255 AllocaInst *AI, 1256 SmallVector<AllocaInst*, 32> &NewElts) { 1257 // If this is a memcpy/memmove, construct the other pointer as the 1258 // appropriate type. The "Other" pointer is the pointer that goes to memory 1259 // that doesn't have anything to do with the alloca that we are promoting. For 1260 // memset, this Value* stays null. 1261 Value *OtherPtr = 0; 1262 unsigned MemAlignment = MI->getAlignment(); 1263 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1264 if (Inst == MTI->getRawDest()) 1265 OtherPtr = MTI->getRawSource(); 1266 else { 1267 assert(Inst == MTI->getRawSource()); 1268 OtherPtr = MTI->getRawDest(); 1269 } 1270 } 1271 1272 // If there is an other pointer, we want to convert it to the same pointer 1273 // type as AI has, so we can GEP through it safely. 1274 if (OtherPtr) { 1275 1276 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1277 // optimization, but it's also required to detect the corner case where 1278 // both pointer operands are referencing the same memory, and where 1279 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1280 // function is only called for mem intrinsics that access the whole 1281 // aggregate, so non-zero GEPs are not an issue here.) 1282 while (1) { 1283 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) { 1284 OtherPtr = BC->getOperand(0); 1285 continue; 1286 } 1287 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) { 1288 // All zero GEPs are effectively bitcasts. 1289 if (GEP->hasAllZeroIndices()) { 1290 OtherPtr = GEP->getOperand(0); 1291 continue; 1292 } 1293 } 1294 break; 1295 } 1296 // Copying the alloca to itself is a no-op: just delete it. 1297 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1298 // This code will run twice for a no-op memcpy -- once for each operand. 1299 // Put only one reference to MI on the DeadInsts list. 1300 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1301 E = DeadInsts.end(); I != E; ++I) 1302 if (*I == MI) return; 1303 DeadInsts.push_back(MI); 1304 return; 1305 } 1306 1307 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 1308 if (BCE->getOpcode() == Instruction::BitCast) 1309 OtherPtr = BCE->getOperand(0); 1310 1311 // If the pointer is not the right type, insert a bitcast to the right 1312 // type. 1313 if (OtherPtr->getType() != AI->getType()) 1314 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 1315 MI); 1316 } 1317 1318 // Process each element of the aggregate. 1319 Value *TheFn = MI->getCalledValue(); 1320 const Type *BytePtrTy = MI->getRawDest()->getType(); 1321 bool SROADest = MI->getRawDest() == Inst; 1322 1323 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1324 1325 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1326 // If this is a memcpy/memmove, emit a GEP of the other element address. 1327 Value *OtherElt = 0; 1328 unsigned OtherEltAlign = MemAlignment; 1329 1330 if (OtherPtr) { 1331 Value *Idx[2] = { Zero, 1332 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1333 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1334 OtherPtr->getName()+"."+Twine(i), 1335 MI); 1336 uint64_t EltOffset; 1337 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1338 const Type *OtherTy = OtherPtrTy->getElementType(); 1339 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1340 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1341 } else { 1342 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 1343 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1344 } 1345 1346 // The alignment of the other pointer is the guaranteed alignment of the 1347 // element, which is affected by both the known alignment of the whole 1348 // mem intrinsic and the alignment of the element. If the alignment of 1349 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 1350 // known alignment is just 4 bytes. 1351 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 1352 } 1353 1354 Value *EltPtr = NewElts[i]; 1355 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 1356 1357 // If we got down to a scalar, insert a load or store as appropriate. 1358 if (EltTy->isSingleValueType()) { 1359 if (isa<MemTransferInst>(MI)) { 1360 if (SROADest) { 1361 // From Other to Alloca. 1362 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 1363 new StoreInst(Elt, EltPtr, MI); 1364 } else { 1365 // From Alloca to Other. 1366 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 1367 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 1368 } 1369 continue; 1370 } 1371 assert(isa<MemSetInst>(MI)); 1372 1373 // If the stored element is zero (common case), just store a null 1374 // constant. 1375 Constant *StoreVal; 1376 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 1377 if (CI->isZero()) { 1378 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 1379 } else { 1380 // If EltTy is a vector type, get the element type. 1381 const Type *ValTy = EltTy->getScalarType(); 1382 1383 // Construct an integer with the right value. 1384 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 1385 APInt OneVal(EltSize, CI->getZExtValue()); 1386 APInt TotalVal(OneVal); 1387 // Set each byte. 1388 for (unsigned i = 0; 8*i < EltSize; ++i) { 1389 TotalVal = TotalVal.shl(8); 1390 TotalVal |= OneVal; 1391 } 1392 1393 // Convert the integer value to the appropriate type. 1394 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 1395 if (ValTy->isPointerTy()) 1396 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 1397 else if (ValTy->isFloatingPointTy()) 1398 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 1399 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 1400 1401 // If the requested value was a vector constant, create it. 1402 if (EltTy != ValTy) { 1403 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 1404 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 1405 StoreVal = ConstantVector::get(&Elts[0], NumElts); 1406 } 1407 } 1408 new StoreInst(StoreVal, EltPtr, MI); 1409 continue; 1410 } 1411 // Otherwise, if we're storing a byte variable, use a memset call for 1412 // this element. 1413 } 1414 1415 // Cast the element pointer to BytePtrTy. 1416 if (EltPtr->getType() != BytePtrTy) 1417 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI); 1418 1419 // Cast the other pointer (if we have one) to BytePtrTy. 1420 if (OtherElt && OtherElt->getType() != BytePtrTy) { 1421 // Preserve address space of OtherElt 1422 const PointerType* OtherPTy = cast<PointerType>(OtherElt->getType()); 1423 const PointerType* PTy = cast<PointerType>(BytePtrTy); 1424 if (OtherPTy->getElementType() != PTy->getElementType()) { 1425 Type *NewOtherPTy = PointerType::get(PTy->getElementType(), 1426 OtherPTy->getAddressSpace()); 1427 OtherElt = new BitCastInst(OtherElt, NewOtherPTy, 1428 OtherElt->getNameStr(), MI); 1429 } 1430 } 1431 1432 unsigned EltSize = TD->getTypeAllocSize(EltTy); 1433 1434 // Finally, insert the meminst for this element. 1435 if (isa<MemTransferInst>(MI)) { 1436 Value *Ops[] = { 1437 SROADest ? EltPtr : OtherElt, // Dest ptr 1438 SROADest ? OtherElt : EltPtr, // Src ptr 1439 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 1440 // Align 1441 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign), 1442 MI->getVolatileCst() 1443 }; 1444 // In case we fold the address space overloaded memcpy of A to B 1445 // with memcpy of B to C, change the function to be a memcpy of A to C. 1446 const Type *Tys[] = { Ops[0]->getType(), Ops[1]->getType(), 1447 Ops[2]->getType() }; 1448 Module *M = MI->getParent()->getParent()->getParent(); 1449 TheFn = Intrinsic::getDeclaration(M, MI->getIntrinsicID(), Tys, 3); 1450 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1451 } else { 1452 assert(isa<MemSetInst>(MI)); 1453 Value *Ops[] = { 1454 EltPtr, MI->getOperand(2), // Dest, Value, 1455 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 1456 Zero, // Align 1457 ConstantInt::get(Type::getInt1Ty(MI->getContext()), 0) // isVolatile 1458 }; 1459 const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() }; 1460 Module *M = MI->getParent()->getParent()->getParent(); 1461 TheFn = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2); 1462 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1463 } 1464 } 1465 DeadInsts.push_back(MI); 1466} 1467 1468/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 1469/// overwrites the entire allocation. Extract out the pieces of the stored 1470/// integer and store them individually. 1471void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 1472 SmallVector<AllocaInst*, 32> &NewElts){ 1473 // Extract each element out of the integer according to its structure offset 1474 // and store the element value to the individual alloca. 1475 Value *SrcVal = SI->getOperand(0); 1476 const Type *AllocaEltTy = AI->getAllocatedType(); 1477 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1478 1479 // Handle tail padding by extending the operand 1480 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 1481 SrcVal = new ZExtInst(SrcVal, 1482 IntegerType::get(SI->getContext(), AllocaSizeBits), 1483 "", SI); 1484 1485 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 1486 << '\n'); 1487 1488 // There are two forms here: AI could be an array or struct. Both cases 1489 // have different ways to compute the element offset. 1490 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1491 const StructLayout *Layout = TD->getStructLayout(EltSTy); 1492 1493 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1494 // Get the number of bits to shift SrcVal to get the value. 1495 const Type *FieldTy = EltSTy->getElementType(i); 1496 uint64_t Shift = Layout->getElementOffsetInBits(i); 1497 1498 if (TD->isBigEndian()) 1499 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 1500 1501 Value *EltVal = SrcVal; 1502 if (Shift) { 1503 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1504 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1505 "sroa.store.elt", SI); 1506 } 1507 1508 // Truncate down to an integer of the right size. 1509 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1510 1511 // Ignore zero sized fields like {}, they obviously contain no data. 1512 if (FieldSizeBits == 0) continue; 1513 1514 if (FieldSizeBits != AllocaSizeBits) 1515 EltVal = new TruncInst(EltVal, 1516 IntegerType::get(SI->getContext(), FieldSizeBits), 1517 "", SI); 1518 Value *DestField = NewElts[i]; 1519 if (EltVal->getType() == FieldTy) { 1520 // Storing to an integer field of this size, just do it. 1521 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 1522 // Bitcast to the right element type (for fp/vector values). 1523 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 1524 } else { 1525 // Otherwise, bitcast the dest pointer (for aggregates). 1526 DestField = new BitCastInst(DestField, 1527 PointerType::getUnqual(EltVal->getType()), 1528 "", SI); 1529 } 1530 new StoreInst(EltVal, DestField, SI); 1531 } 1532 1533 } else { 1534 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1535 const Type *ArrayEltTy = ATy->getElementType(); 1536 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1537 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1538 1539 uint64_t Shift; 1540 1541 if (TD->isBigEndian()) 1542 Shift = AllocaSizeBits-ElementOffset; 1543 else 1544 Shift = 0; 1545 1546 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1547 // Ignore zero sized fields like {}, they obviously contain no data. 1548 if (ElementSizeBits == 0) continue; 1549 1550 Value *EltVal = SrcVal; 1551 if (Shift) { 1552 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1553 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1554 "sroa.store.elt", SI); 1555 } 1556 1557 // Truncate down to an integer of the right size. 1558 if (ElementSizeBits != AllocaSizeBits) 1559 EltVal = new TruncInst(EltVal, 1560 IntegerType::get(SI->getContext(), 1561 ElementSizeBits),"",SI); 1562 Value *DestField = NewElts[i]; 1563 if (EltVal->getType() == ArrayEltTy) { 1564 // Storing to an integer field of this size, just do it. 1565 } else if (ArrayEltTy->isFloatingPointTy() || 1566 ArrayEltTy->isVectorTy()) { 1567 // Bitcast to the right element type (for fp/vector values). 1568 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1569 } else { 1570 // Otherwise, bitcast the dest pointer (for aggregates). 1571 DestField = new BitCastInst(DestField, 1572 PointerType::getUnqual(EltVal->getType()), 1573 "", SI); 1574 } 1575 new StoreInst(EltVal, DestField, SI); 1576 1577 if (TD->isBigEndian()) 1578 Shift -= ElementOffset; 1579 else 1580 Shift += ElementOffset; 1581 } 1582 } 1583 1584 DeadInsts.push_back(SI); 1585} 1586 1587/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1588/// an integer. Load the individual pieces to form the aggregate value. 1589void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1590 SmallVector<AllocaInst*, 32> &NewElts) { 1591 // Extract each element out of the NewElts according to its structure offset 1592 // and form the result value. 1593 const Type *AllocaEltTy = AI->getAllocatedType(); 1594 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1595 1596 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1597 << '\n'); 1598 1599 // There are two forms here: AI could be an array or struct. Both cases 1600 // have different ways to compute the element offset. 1601 const StructLayout *Layout = 0; 1602 uint64_t ArrayEltBitOffset = 0; 1603 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1604 Layout = TD->getStructLayout(EltSTy); 1605 } else { 1606 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1607 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1608 } 1609 1610 Value *ResultVal = 1611 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1612 1613 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1614 // Load the value from the alloca. If the NewElt is an aggregate, cast 1615 // the pointer to an integer of the same size before doing the load. 1616 Value *SrcField = NewElts[i]; 1617 const Type *FieldTy = 1618 cast<PointerType>(SrcField->getType())->getElementType(); 1619 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1620 1621 // Ignore zero sized fields like {}, they obviously contain no data. 1622 if (FieldSizeBits == 0) continue; 1623 1624 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1625 FieldSizeBits); 1626 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1627 !FieldTy->isVectorTy()) 1628 SrcField = new BitCastInst(SrcField, 1629 PointerType::getUnqual(FieldIntTy), 1630 "", LI); 1631 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1632 1633 // If SrcField is a fp or vector of the right size but that isn't an 1634 // integer type, bitcast to an integer so we can shift it. 1635 if (SrcField->getType() != FieldIntTy) 1636 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1637 1638 // Zero extend the field to be the same size as the final alloca so that 1639 // we can shift and insert it. 1640 if (SrcField->getType() != ResultVal->getType()) 1641 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1642 1643 // Determine the number of bits to shift SrcField. 1644 uint64_t Shift; 1645 if (Layout) // Struct case. 1646 Shift = Layout->getElementOffsetInBits(i); 1647 else // Array case. 1648 Shift = i*ArrayEltBitOffset; 1649 1650 if (TD->isBigEndian()) 1651 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1652 1653 if (Shift) { 1654 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1655 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1656 } 1657 1658 // Don't create an 'or x, 0' on the first iteration. 1659 if (!isa<Constant>(ResultVal) || 1660 !cast<Constant>(ResultVal)->isNullValue()) 1661 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1662 else 1663 ResultVal = SrcField; 1664 } 1665 1666 // Handle tail padding by truncating the result 1667 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1668 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1669 1670 LI->replaceAllUsesWith(ResultVal); 1671 DeadInsts.push_back(LI); 1672} 1673 1674/// HasPadding - Return true if the specified type has any structure or 1675/// alignment padding, false otherwise. 1676static bool HasPadding(const Type *Ty, const TargetData &TD) { 1677 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1678 const StructLayout *SL = TD.getStructLayout(STy); 1679 unsigned PrevFieldBitOffset = 0; 1680 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1681 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1682 1683 // Padding in sub-elements? 1684 if (HasPadding(STy->getElementType(i), TD)) 1685 return true; 1686 1687 // Check to see if there is any padding between this element and the 1688 // previous one. 1689 if (i) { 1690 unsigned PrevFieldEnd = 1691 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1692 if (PrevFieldEnd < FieldBitOffset) 1693 return true; 1694 } 1695 1696 PrevFieldBitOffset = FieldBitOffset; 1697 } 1698 1699 // Check for tail padding. 1700 if (unsigned EltCount = STy->getNumElements()) { 1701 unsigned PrevFieldEnd = PrevFieldBitOffset + 1702 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1703 if (PrevFieldEnd < SL->getSizeInBits()) 1704 return true; 1705 } 1706 1707 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1708 return HasPadding(ATy->getElementType(), TD); 1709 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1710 return HasPadding(VTy->getElementType(), TD); 1711 } 1712 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1713} 1714 1715/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1716/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1717/// or 1 if safe after canonicalization has been performed. 1718bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1719 // Loop over the use list of the alloca. We can only transform it if all of 1720 // the users are safe to transform. 1721 AllocaInfo Info; 1722 1723 isSafeForScalarRepl(AI, AI, 0, Info); 1724 if (Info.isUnsafe) { 1725 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1726 return false; 1727 } 1728 1729 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1730 // source and destination, we have to be careful. In particular, the memcpy 1731 // could be moving around elements that live in structure padding of the LLVM 1732 // types, but may actually be used. In these cases, we refuse to promote the 1733 // struct. 1734 if (Info.isMemCpySrc && Info.isMemCpyDst && 1735 HasPadding(AI->getAllocatedType(), *TD)) 1736 return false; 1737 1738 return true; 1739} 1740 1741 1742 1743/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1744/// some part of a constant global variable. This intentionally only accepts 1745/// constant expressions because we don't can't rewrite arbitrary instructions. 1746static bool PointsToConstantGlobal(Value *V) { 1747 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1748 return GV->isConstant(); 1749 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1750 if (CE->getOpcode() == Instruction::BitCast || 1751 CE->getOpcode() == Instruction::GetElementPtr) 1752 return PointsToConstantGlobal(CE->getOperand(0)); 1753 return false; 1754} 1755 1756/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1757/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1758/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1759/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1760/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1761/// the alloca, and if the source pointer is a pointer to a constant global, we 1762/// can optimize this. 1763static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 1764 bool isOffset) { 1765 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1766 User *U = cast<Instruction>(*UI); 1767 1768 if (LoadInst *LI = dyn_cast<LoadInst>(U)) 1769 // Ignore non-volatile loads, they are always ok. 1770 if (!LI->isVolatile()) 1771 continue; 1772 1773 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1774 // If uses of the bitcast are ok, we are ok. 1775 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1776 return false; 1777 continue; 1778 } 1779 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 1780 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1781 // doesn't, it does. 1782 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1783 isOffset || !GEP->hasAllZeroIndices())) 1784 return false; 1785 continue; 1786 } 1787 1788 // If this is isn't our memcpy/memmove, reject it as something we can't 1789 // handle. 1790 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 1791 if (MI == 0) 1792 return false; 1793 1794 // If we already have seen a copy, reject the second one. 1795 if (TheCopy) return false; 1796 1797 // If the pointer has been offset from the start of the alloca, we can't 1798 // safely handle this. 1799 if (isOffset) return false; 1800 1801 // If the memintrinsic isn't using the alloca as the dest, reject it. 1802 if (UI.getOperandNo() != CallInst::ArgOffset) return false; 1803 1804 // If the source of the memcpy/move is not a constant global, reject it. 1805 if (!PointsToConstantGlobal(MI->getSource())) 1806 return false; 1807 1808 // Otherwise, the transform is safe. Remember the copy instruction. 1809 TheCopy = MI; 1810 } 1811 return true; 1812} 1813 1814/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1815/// modified by a copy from a constant global. If we can prove this, we can 1816/// replace any uses of the alloca with uses of the global directly. 1817MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1818 MemTransferInst *TheCopy = 0; 1819 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1820 return TheCopy; 1821 return 0; 1822} 1823