ScalarReplAggregates.cpp revision 72eaa0e5eb345a8483608675b86dfcfa465c784c
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/Module.h" 32#include "llvm/Pass.h" 33#include "llvm/Analysis/Dominators.h" 34#include "llvm/Target/TargetData.h" 35#include "llvm/Transforms/Utils/PromoteMemToReg.h" 36#include "llvm/Transforms/Utils/Local.h" 37#include "llvm/Support/Debug.h" 38#include "llvm/Support/ErrorHandling.h" 39#include "llvm/Support/GetElementPtrTypeIterator.h" 40#include "llvm/Support/IRBuilder.h" 41#include "llvm/Support/MathExtras.h" 42#include "llvm/Support/raw_ostream.h" 43#include "llvm/ADT/SmallVector.h" 44#include "llvm/ADT/Statistic.h" 45using namespace llvm; 46 47STATISTIC(NumReplaced, "Number of allocas broken up"); 48STATISTIC(NumPromoted, "Number of allocas promoted"); 49STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 50STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 51 52namespace { 53 struct SROA : public FunctionPass { 54 static char ID; // Pass identification, replacement for typeid 55 explicit SROA(signed T = -1) : FunctionPass(ID) { 56 if (T == -1) 57 SRThreshold = 128; 58 else 59 SRThreshold = T; 60 } 61 62 bool runOnFunction(Function &F); 63 64 bool performScalarRepl(Function &F); 65 bool performPromotion(Function &F); 66 67 // getAnalysisUsage - This pass does not require any passes, but we know it 68 // will not alter the CFG, so say so. 69 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 70 AU.addRequired<DominatorTree>(); 71 AU.addRequired<DominanceFrontier>(); 72 AU.setPreservesCFG(); 73 } 74 75 private: 76 TargetData *TD; 77 78 /// DeadInsts - Keep track of instructions we have made dead, so that 79 /// we can remove them after we are done working. 80 SmallVector<Value*, 32> DeadInsts; 81 82 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 83 /// information about the uses. All these fields are initialized to false 84 /// and set to true when something is learned. 85 struct AllocaInfo { 86 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 87 bool isUnsafe : 1; 88 89 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 90 bool isMemCpySrc : 1; 91 92 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 93 bool isMemCpyDst : 1; 94 95 AllocaInfo() 96 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false) {} 97 }; 98 99 unsigned SRThreshold; 100 101 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 102 103 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 104 105 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 106 AllocaInfo &Info); 107 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset, 108 AllocaInfo &Info); 109 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 110 const Type *MemOpType, bool isStore, AllocaInfo &Info); 111 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 112 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 113 const Type *&IdxTy); 114 115 void DoScalarReplacement(AllocaInst *AI, 116 std::vector<AllocaInst*> &WorkList); 117 void DeleteDeadInstructions(); 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; 138INITIALIZE_PASS(SROA, "scalarrepl", 139 "Scalar Replacement of Aggregates", false, false); 140 141// Public interface to the ScalarReplAggregates pass 142FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 143 return new SROA(Threshold); 144} 145 146 147//===----------------------------------------------------------------------===// 148// Convert To Scalar Optimization. 149//===----------------------------------------------------------------------===// 150 151namespace { 152/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 153/// optimization, which scans the uses of an alloca and determines if it can 154/// rewrite it in terms of a single new alloca that can be mem2reg'd. 155class ConvertToScalarInfo { 156 /// AllocaSize - The size of the alloca being considered. 157 unsigned AllocaSize; 158 const TargetData &TD; 159 160 /// IsNotTrivial - This is set to true if there is some access to the object 161 /// which means that mem2reg can't promote it. 162 bool IsNotTrivial; 163 164 /// VectorTy - This tracks the type that we should promote the vector to if 165 /// it is possible to turn it into a vector. This starts out null, and if it 166 /// isn't possible to turn into a vector type, it gets set to VoidTy. 167 const Type *VectorTy; 168 169 /// HadAVector - True if there is at least one vector access to the alloca. 170 /// We don't want to turn random arrays into vectors and use vector element 171 /// insert/extract, but if there are element accesses to something that is 172 /// also declared as a vector, we do want to promote to a vector. 173 bool HadAVector; 174 175public: 176 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 177 : AllocaSize(Size), TD(td) { 178 IsNotTrivial = false; 179 VectorTy = 0; 180 HadAVector = false; 181 } 182 183 AllocaInst *TryConvert(AllocaInst *AI); 184 185private: 186 bool CanConvertToScalar(Value *V, uint64_t Offset); 187 void MergeInType(const Type *In, uint64_t Offset); 188 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 189 190 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 191 uint64_t Offset, IRBuilder<> &Builder); 192 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 193 uint64_t Offset, IRBuilder<> &Builder); 194}; 195} // end anonymous namespace. 196 197 198/// IsVerbotenVectorType - Return true if this is a vector type ScalarRepl isn't 199/// allowed to form. We do this to avoid MMX types, which is a complete hack, 200/// but is required until the backend is fixed. 201static bool IsVerbotenVectorType(const VectorType *VTy, const Instruction *I) { 202 StringRef Triple(I->getParent()->getParent()->getParent()->getTargetTriple()); 203 if (!Triple.startswith("i386") && 204 !Triple.startswith("x86_64")) 205 return false; 206 207 // Reject all the MMX vector types. 208 switch (VTy->getNumElements()) { 209 default: return false; 210 case 1: return VTy->getElementType()->isIntegerTy(64); 211 case 2: return VTy->getElementType()->isIntegerTy(32); 212 case 4: return VTy->getElementType()->isIntegerTy(16); 213 case 8: return VTy->getElementType()->isIntegerTy(8); 214 } 215} 216 217 218/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 219/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 220/// alloca if possible or null if not. 221AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 222 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 223 // out. 224 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 225 return 0; 226 227 // If we were able to find a vector type that can handle this with 228 // insert/extract elements, and if there was at least one use that had 229 // a vector type, promote this to a vector. We don't want to promote 230 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 231 // we just get a lot of insert/extracts. If at least one vector is 232 // involved, then we probably really do have a union of vector/array. 233 const Type *NewTy; 234 if (VectorTy && VectorTy->isVectorTy() && HadAVector && 235 !IsVerbotenVectorType(cast<VectorType>(VectorTy), AI)) { 236 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 237 << *VectorTy << '\n'); 238 NewTy = VectorTy; // Use the vector type. 239 } else { 240 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 241 // Create and insert the integer alloca. 242 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 243 } 244 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 245 ConvertUsesToScalar(AI, NewAI, 0); 246 return NewAI; 247} 248 249/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy) 250/// so far at the offset specified by Offset (which is specified in bytes). 251/// 252/// There are two cases we handle here: 253/// 1) A union of vector types of the same size and potentially its elements. 254/// Here we turn element accesses into insert/extract element operations. 255/// This promotes a <4 x float> with a store of float to the third element 256/// into a <4 x float> that uses insert element. 257/// 2) A fully general blob of memory, which we turn into some (potentially 258/// large) integer type with extract and insert operations where the loads 259/// and stores would mutate the memory. We mark this by setting VectorTy 260/// to VoidTy. 261void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) { 262 // If we already decided to turn this into a blob of integer memory, there is 263 // nothing to be done. 264 if (VectorTy && VectorTy->isVoidTy()) 265 return; 266 267 // If this could be contributing to a vector, analyze it. 268 269 // If the In type is a vector that is the same size as the alloca, see if it 270 // matches the existing VecTy. 271 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 272 // Remember if we saw a vector type. 273 HadAVector = true; 274 275 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 276 // If we're storing/loading a vector of the right size, allow it as a 277 // vector. If this the first vector we see, remember the type so that 278 // we know the element size. If this is a subsequent access, ignore it 279 // even if it is a differing type but the same size. Worst case we can 280 // bitcast the resultant vectors. 281 if (VectorTy == 0) 282 VectorTy = VInTy; 283 return; 284 } 285 } else if (In->isFloatTy() || In->isDoubleTy() || 286 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 287 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 288 // If we're accessing something that could be an element of a vector, see 289 // if the implied vector agrees with what we already have and if Offset is 290 // compatible with it. 291 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 292 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 293 (VectorTy == 0 || 294 cast<VectorType>(VectorTy)->getElementType() 295 ->getPrimitiveSizeInBits()/8 == EltSize)) { 296 if (VectorTy == 0) 297 VectorTy = VectorType::get(In, AllocaSize/EltSize); 298 return; 299 } 300 } 301 302 // Otherwise, we have a case that we can't handle with an optimized vector 303 // form. We can still turn this into a large integer. 304 VectorTy = Type::getVoidTy(In->getContext()); 305} 306 307/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 308/// its accesses to a single vector type, return true and set VecTy to 309/// the new type. If we could convert the alloca into a single promotable 310/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 311/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 312/// is the current offset from the base of the alloca being analyzed. 313/// 314/// If we see at least one access to the value that is as a vector type, set the 315/// SawVec flag. 316bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 317 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 318 Instruction *User = cast<Instruction>(*UI); 319 320 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 321 // Don't break volatile loads. 322 if (LI->isVolatile()) 323 return false; 324 MergeInType(LI->getType(), Offset); 325 continue; 326 } 327 328 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 329 // Storing the pointer, not into the value? 330 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 331 MergeInType(SI->getOperand(0)->getType(), Offset); 332 continue; 333 } 334 335 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 336 IsNotTrivial = true; // Can't be mem2reg'd. 337 if (!CanConvertToScalar(BCI, Offset)) 338 return false; 339 continue; 340 } 341 342 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 343 // If this is a GEP with a variable indices, we can't handle it. 344 if (!GEP->hasAllConstantIndices()) 345 return false; 346 347 // Compute the offset that this GEP adds to the pointer. 348 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 349 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 350 &Indices[0], Indices.size()); 351 // See if all uses can be converted. 352 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 353 return false; 354 IsNotTrivial = true; // Can't be mem2reg'd. 355 continue; 356 } 357 358 // If this is a constant sized memset of a constant value (e.g. 0) we can 359 // handle it. 360 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 361 // Store of constant value and constant size. 362 if (!isa<ConstantInt>(MSI->getValue()) || 363 !isa<ConstantInt>(MSI->getLength())) 364 return false; 365 IsNotTrivial = true; // Can't be mem2reg'd. 366 continue; 367 } 368 369 // If this is a memcpy or memmove into or out of the whole allocation, we 370 // can handle it like a load or store of the scalar type. 371 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 372 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 373 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 374 return false; 375 376 IsNotTrivial = true; // Can't be mem2reg'd. 377 continue; 378 } 379 380 // Otherwise, we cannot handle this! 381 return false; 382 } 383 384 return true; 385} 386 387/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 388/// directly. This happens when we are converting an "integer union" to a 389/// single integer scalar, or when we are converting a "vector union" to a 390/// vector with insert/extractelement instructions. 391/// 392/// Offset is an offset from the original alloca, in bits that need to be 393/// shifted to the right. By the end of this, there should be no uses of Ptr. 394void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 395 uint64_t Offset) { 396 while (!Ptr->use_empty()) { 397 Instruction *User = cast<Instruction>(Ptr->use_back()); 398 399 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 400 ConvertUsesToScalar(CI, NewAI, Offset); 401 CI->eraseFromParent(); 402 continue; 403 } 404 405 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 406 // Compute the offset that this GEP adds to the pointer. 407 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 408 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 409 &Indices[0], Indices.size()); 410 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 411 GEP->eraseFromParent(); 412 continue; 413 } 414 415 IRBuilder<> Builder(User->getParent(), User); 416 417 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 418 // The load is a bit extract from NewAI shifted right by Offset bits. 419 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 420 Value *NewLoadVal 421 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 422 LI->replaceAllUsesWith(NewLoadVal); 423 LI->eraseFromParent(); 424 continue; 425 } 426 427 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 428 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 429 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 430 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 431 Builder); 432 Builder.CreateStore(New, NewAI); 433 SI->eraseFromParent(); 434 435 // If the load we just inserted is now dead, then the inserted store 436 // overwrote the entire thing. 437 if (Old->use_empty()) 438 Old->eraseFromParent(); 439 continue; 440 } 441 442 // If this is a constant sized memset of a constant value (e.g. 0) we can 443 // transform it into a store of the expanded constant value. 444 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 445 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 446 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 447 if (NumBytes != 0) { 448 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 449 450 // Compute the value replicated the right number of times. 451 APInt APVal(NumBytes*8, Val); 452 453 // Splat the value if non-zero. 454 if (Val) 455 for (unsigned i = 1; i != NumBytes; ++i) 456 APVal |= APVal << 8; 457 458 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 459 Value *New = ConvertScalar_InsertValue( 460 ConstantInt::get(User->getContext(), APVal), 461 Old, Offset, Builder); 462 Builder.CreateStore(New, NewAI); 463 464 // If the load we just inserted is now dead, then the memset overwrote 465 // the entire thing. 466 if (Old->use_empty()) 467 Old->eraseFromParent(); 468 } 469 MSI->eraseFromParent(); 470 continue; 471 } 472 473 // If this is a memcpy or memmove into or out of the whole allocation, we 474 // can handle it like a load or store of the scalar type. 475 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 476 assert(Offset == 0 && "must be store to start of alloca"); 477 478 // If the source and destination are both to the same alloca, then this is 479 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 480 // as appropriate. 481 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject(0)); 482 483 if (MTI->getSource()->getUnderlyingObject(0) != OrigAI) { 484 // Dest must be OrigAI, change this to be a load from the original 485 // pointer (bitcasted), then a store to our new alloca. 486 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 487 Value *SrcPtr = MTI->getSource(); 488 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 489 490 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 491 SrcVal->setAlignment(MTI->getAlignment()); 492 Builder.CreateStore(SrcVal, NewAI); 493 } else if (MTI->getDest()->getUnderlyingObject(0) != OrigAI) { 494 // Src must be OrigAI, change this to be a load from NewAI then a store 495 // through the original dest pointer (bitcasted). 496 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 497 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 498 499 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 500 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 501 NewStore->setAlignment(MTI->getAlignment()); 502 } else { 503 // Noop transfer. Src == Dst 504 } 505 506 MTI->eraseFromParent(); 507 continue; 508 } 509 510 llvm_unreachable("Unsupported operation!"); 511 } 512} 513 514/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 515/// or vector value FromVal, extracting the bits from the offset specified by 516/// Offset. This returns the value, which is of type ToType. 517/// 518/// This happens when we are converting an "integer union" to a single 519/// integer scalar, or when we are converting a "vector union" to a vector with 520/// insert/extractelement instructions. 521/// 522/// Offset is an offset from the original alloca, in bits that need to be 523/// shifted to the right. 524Value *ConvertToScalarInfo:: 525ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 526 uint64_t Offset, IRBuilder<> &Builder) { 527 // If the load is of the whole new alloca, no conversion is needed. 528 if (FromVal->getType() == ToType && Offset == 0) 529 return FromVal; 530 531 // If the result alloca is a vector type, this is either an element 532 // access or a bitcast to another vector type of the same size. 533 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 534 if (ToType->isVectorTy()) 535 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 536 537 // Otherwise it must be an element access. 538 unsigned Elt = 0; 539 if (Offset) { 540 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 541 Elt = Offset/EltSize; 542 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 543 } 544 // Return the element extracted out of it. 545 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 546 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 547 if (V->getType() != ToType) 548 V = Builder.CreateBitCast(V, ToType, "tmp"); 549 return V; 550 } 551 552 // If ToType is a first class aggregate, extract out each of the pieces and 553 // use insertvalue's to form the FCA. 554 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 555 const StructLayout &Layout = *TD.getStructLayout(ST); 556 Value *Res = UndefValue::get(ST); 557 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 558 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 559 Offset+Layout.getElementOffsetInBits(i), 560 Builder); 561 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 562 } 563 return Res; 564 } 565 566 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 567 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 568 Value *Res = UndefValue::get(AT); 569 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 570 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 571 Offset+i*EltSize, Builder); 572 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 573 } 574 return Res; 575 } 576 577 // Otherwise, this must be a union that was converted to an integer value. 578 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 579 580 // If this is a big-endian system and the load is narrower than the 581 // full alloca type, we need to do a shift to get the right bits. 582 int ShAmt = 0; 583 if (TD.isBigEndian()) { 584 // On big-endian machines, the lowest bit is stored at the bit offset 585 // from the pointer given by getTypeStoreSizeInBits. This matters for 586 // integers with a bitwidth that is not a multiple of 8. 587 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 588 TD.getTypeStoreSizeInBits(ToType) - Offset; 589 } else { 590 ShAmt = Offset; 591 } 592 593 // Note: we support negative bitwidths (with shl) which are not defined. 594 // We do this to support (f.e.) loads off the end of a structure where 595 // only some bits are used. 596 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 597 FromVal = Builder.CreateLShr(FromVal, 598 ConstantInt::get(FromVal->getType(), 599 ShAmt), "tmp"); 600 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 601 FromVal = Builder.CreateShl(FromVal, 602 ConstantInt::get(FromVal->getType(), 603 -ShAmt), "tmp"); 604 605 // Finally, unconditionally truncate the integer to the right width. 606 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 607 if (LIBitWidth < NTy->getBitWidth()) 608 FromVal = 609 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 610 LIBitWidth), "tmp"); 611 else if (LIBitWidth > NTy->getBitWidth()) 612 FromVal = 613 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 614 LIBitWidth), "tmp"); 615 616 // If the result is an integer, this is a trunc or bitcast. 617 if (ToType->isIntegerTy()) { 618 // Should be done. 619 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 620 // Just do a bitcast, we know the sizes match up. 621 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 622 } else { 623 // Otherwise must be a pointer. 624 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 625 } 626 assert(FromVal->getType() == ToType && "Didn't convert right?"); 627 return FromVal; 628} 629 630/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 631/// or vector value "Old" at the offset specified by Offset. 632/// 633/// This happens when we are converting an "integer union" to a 634/// single integer scalar, or when we are converting a "vector union" to a 635/// vector with insert/extractelement instructions. 636/// 637/// Offset is an offset from the original alloca, in bits that need to be 638/// shifted to the right. 639Value *ConvertToScalarInfo:: 640ConvertScalar_InsertValue(Value *SV, Value *Old, 641 uint64_t Offset, IRBuilder<> &Builder) { 642 // Convert the stored type to the actual type, shift it left to insert 643 // then 'or' into place. 644 const Type *AllocaType = Old->getType(); 645 LLVMContext &Context = Old->getContext(); 646 647 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 648 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 649 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 650 651 // Changing the whole vector with memset or with an access of a different 652 // vector type? 653 if (ValSize == VecSize) 654 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 655 656 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 657 658 // Must be an element insertion. 659 unsigned Elt = Offset/EltSize; 660 661 if (SV->getType() != VTy->getElementType()) 662 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 663 664 SV = Builder.CreateInsertElement(Old, SV, 665 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 666 "tmp"); 667 return SV; 668 } 669 670 // If SV is a first-class aggregate value, insert each value recursively. 671 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 672 const StructLayout &Layout = *TD.getStructLayout(ST); 673 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 674 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 675 Old = ConvertScalar_InsertValue(Elt, Old, 676 Offset+Layout.getElementOffsetInBits(i), 677 Builder); 678 } 679 return Old; 680 } 681 682 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 683 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 684 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 685 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 686 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 687 } 688 return Old; 689 } 690 691 // If SV is a float, convert it to the appropriate integer type. 692 // If it is a pointer, do the same. 693 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 694 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 695 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 696 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 697 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 698 SV = Builder.CreateBitCast(SV, 699 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 700 else if (SV->getType()->isPointerTy()) 701 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 702 703 // Zero extend or truncate the value if needed. 704 if (SV->getType() != AllocaType) { 705 if (SV->getType()->getPrimitiveSizeInBits() < 706 AllocaType->getPrimitiveSizeInBits()) 707 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 708 else { 709 // Truncation may be needed if storing more than the alloca can hold 710 // (undefined behavior). 711 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 712 SrcWidth = DestWidth; 713 SrcStoreWidth = DestStoreWidth; 714 } 715 } 716 717 // If this is a big-endian system and the store is narrower than the 718 // full alloca type, we need to do a shift to get the right bits. 719 int ShAmt = 0; 720 if (TD.isBigEndian()) { 721 // On big-endian machines, the lowest bit is stored at the bit offset 722 // from the pointer given by getTypeStoreSizeInBits. This matters for 723 // integers with a bitwidth that is not a multiple of 8. 724 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 725 } else { 726 ShAmt = Offset; 727 } 728 729 // Note: we support negative bitwidths (with shr) which are not defined. 730 // We do this to support (f.e.) stores off the end of a structure where 731 // only some bits in the structure are set. 732 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 733 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 734 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 735 ShAmt), "tmp"); 736 Mask <<= ShAmt; 737 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 738 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 739 -ShAmt), "tmp"); 740 Mask = Mask.lshr(-ShAmt); 741 } 742 743 // Mask out the bits we are about to insert from the old value, and or 744 // in the new bits. 745 if (SrcWidth != DestWidth) { 746 assert(DestWidth > SrcWidth); 747 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 748 SV = Builder.CreateOr(Old, SV, "ins"); 749 } 750 return SV; 751} 752 753 754//===----------------------------------------------------------------------===// 755// SRoA Driver 756//===----------------------------------------------------------------------===// 757 758 759bool SROA::runOnFunction(Function &F) { 760 TD = getAnalysisIfAvailable<TargetData>(); 761 762 bool Changed = performPromotion(F); 763 764 // FIXME: ScalarRepl currently depends on TargetData more than it 765 // theoretically needs to. It should be refactored in order to support 766 // target-independent IR. Until this is done, just skip the actual 767 // scalar-replacement portion of this pass. 768 if (!TD) return Changed; 769 770 while (1) { 771 bool LocalChange = performScalarRepl(F); 772 if (!LocalChange) break; // No need to repromote if no scalarrepl 773 Changed = true; 774 LocalChange = performPromotion(F); 775 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 776 } 777 778 return Changed; 779} 780 781 782bool SROA::performPromotion(Function &F) { 783 std::vector<AllocaInst*> Allocas; 784 DominatorTree &DT = getAnalysis<DominatorTree>(); 785 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 786 787 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 788 789 bool Changed = false; 790 791 while (1) { 792 Allocas.clear(); 793 794 // Find allocas that are safe to promote, by looking at all instructions in 795 // the entry node 796 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 797 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 798 if (isAllocaPromotable(AI)) 799 Allocas.push_back(AI); 800 801 if (Allocas.empty()) break; 802 803 PromoteMemToReg(Allocas, DT, DF); 804 NumPromoted += Allocas.size(); 805 Changed = true; 806 } 807 808 return Changed; 809} 810 811 812/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 813/// SROA. It must be a struct or array type with a small number of elements. 814static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 815 const Type *T = AI->getAllocatedType(); 816 // Do not promote any struct into more than 32 separate vars. 817 if (const StructType *ST = dyn_cast<StructType>(T)) 818 return ST->getNumElements() <= 32; 819 // Arrays are much less likely to be safe for SROA; only consider 820 // them if they are very small. 821 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 822 return AT->getNumElements() <= 8; 823 return false; 824} 825 826 827// performScalarRepl - This algorithm is a simple worklist driven algorithm, 828// which runs on all of the malloc/alloca instructions in the function, removing 829// them if they are only used by getelementptr instructions. 830// 831bool SROA::performScalarRepl(Function &F) { 832 std::vector<AllocaInst*> WorkList; 833 834 // Scan the entry basic block, adding allocas to the worklist. 835 BasicBlock &BB = F.getEntryBlock(); 836 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 837 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 838 WorkList.push_back(A); 839 840 // Process the worklist 841 bool Changed = false; 842 while (!WorkList.empty()) { 843 AllocaInst *AI = WorkList.back(); 844 WorkList.pop_back(); 845 846 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 847 // with unused elements. 848 if (AI->use_empty()) { 849 AI->eraseFromParent(); 850 Changed = true; 851 continue; 852 } 853 854 // If this alloca is impossible for us to promote, reject it early. 855 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 856 continue; 857 858 // Check to see if this allocation is only modified by a memcpy/memmove from 859 // a constant global. If this is the case, we can change all users to use 860 // the constant global instead. This is commonly produced by the CFE by 861 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 862 // is only subsequently read. 863 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 864 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 865 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 866 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 867 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 868 TheCopy->eraseFromParent(); // Don't mutate the global. 869 AI->eraseFromParent(); 870 ++NumGlobals; 871 Changed = true; 872 continue; 873 } 874 875 // Check to see if we can perform the core SROA transformation. We cannot 876 // transform the allocation instruction if it is an array allocation 877 // (allocations OF arrays are ok though), and an allocation of a scalar 878 // value cannot be decomposed at all. 879 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 880 881 // Do not promote [0 x %struct]. 882 if (AllocaSize == 0) continue; 883 884 // Do not promote any struct whose size is too big. 885 if (AllocaSize > SRThreshold) continue; 886 887 // If the alloca looks like a good candidate for scalar replacement, and if 888 // all its users can be transformed, then split up the aggregate into its 889 // separate elements. 890 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 891 DoScalarReplacement(AI, WorkList); 892 Changed = true; 893 continue; 894 } 895 896 // If we can turn this aggregate value (potentially with casts) into a 897 // simple scalar value that can be mem2reg'd into a register value. 898 // IsNotTrivial tracks whether this is something that mem2reg could have 899 // promoted itself. If so, we don't want to transform it needlessly. Note 900 // that we can't just check based on the type: the alloca may be of an i32 901 // but that has pointer arithmetic to set byte 3 of it or something. 902 if (AllocaInst *NewAI = 903 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 904 NewAI->takeName(AI); 905 AI->eraseFromParent(); 906 ++NumConverted; 907 Changed = true; 908 continue; 909 } 910 911 // Otherwise, couldn't process this alloca. 912 } 913 914 return Changed; 915} 916 917/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 918/// predicate, do SROA now. 919void SROA::DoScalarReplacement(AllocaInst *AI, 920 std::vector<AllocaInst*> &WorkList) { 921 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 922 SmallVector<AllocaInst*, 32> ElementAllocas; 923 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 924 ElementAllocas.reserve(ST->getNumContainedTypes()); 925 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 926 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 927 AI->getAlignment(), 928 AI->getName() + "." + Twine(i), AI); 929 ElementAllocas.push_back(NA); 930 WorkList.push_back(NA); // Add to worklist for recursive processing 931 } 932 } else { 933 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 934 ElementAllocas.reserve(AT->getNumElements()); 935 const Type *ElTy = AT->getElementType(); 936 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 937 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 938 AI->getName() + "." + Twine(i), AI); 939 ElementAllocas.push_back(NA); 940 WorkList.push_back(NA); // Add to worklist for recursive processing 941 } 942 } 943 944 // Now that we have created the new alloca instructions, rewrite all the 945 // uses of the old alloca. 946 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 947 948 // Now erase any instructions that were made dead while rewriting the alloca. 949 DeleteDeadInstructions(); 950 AI->eraseFromParent(); 951 952 ++NumReplaced; 953} 954 955/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 956/// recursively including all their operands that become trivially dead. 957void SROA::DeleteDeadInstructions() { 958 while (!DeadInsts.empty()) { 959 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 960 961 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 962 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 963 // Zero out the operand and see if it becomes trivially dead. 964 // (But, don't add allocas to the dead instruction list -- they are 965 // already on the worklist and will be deleted separately.) 966 *OI = 0; 967 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 968 DeadInsts.push_back(U); 969 } 970 971 I->eraseFromParent(); 972 } 973} 974 975/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 976/// performing scalar replacement of alloca AI. The results are flagged in 977/// the Info parameter. Offset indicates the position within AI that is 978/// referenced by this instruction. 979void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 980 AllocaInfo &Info) { 981 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 982 Instruction *User = cast<Instruction>(*UI); 983 984 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 985 isSafeForScalarRepl(BC, AI, Offset, Info); 986 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 987 uint64_t GEPOffset = Offset; 988 isSafeGEP(GEPI, AI, GEPOffset, Info); 989 if (!Info.isUnsafe) 990 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 991 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 992 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 993 if (Length) 994 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 995 UI.getOperandNo() == 0, Info); 996 else 997 MarkUnsafe(Info); 998 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 999 if (!LI->isVolatile()) { 1000 const Type *LIType = LI->getType(); 1001 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 1002 LIType, false, Info); 1003 } else 1004 MarkUnsafe(Info); 1005 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1006 // Store is ok if storing INTO the pointer, not storing the pointer 1007 if (!SI->isVolatile() && SI->getOperand(0) != I) { 1008 const Type *SIType = SI->getOperand(0)->getType(); 1009 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 1010 SIType, true, Info); 1011 } else 1012 MarkUnsafe(Info); 1013 } else { 1014 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 1015 MarkUnsafe(Info); 1016 } 1017 if (Info.isUnsafe) return; 1018 } 1019} 1020 1021/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1022/// replacement. It is safe when all the indices are constant, in-bounds 1023/// references, and when the resulting offset corresponds to an element within 1024/// the alloca type. The results are flagged in the Info parameter. Upon 1025/// return, Offset is adjusted as specified by the GEP indices. 1026void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 1027 uint64_t &Offset, AllocaInfo &Info) { 1028 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1029 if (GEPIt == E) 1030 return; 1031 1032 // Walk through the GEP type indices, checking the types that this indexes 1033 // into. 1034 for (; GEPIt != E; ++GEPIt) { 1035 // Ignore struct elements, no extra checking needed for these. 1036 if ((*GEPIt)->isStructTy()) 1037 continue; 1038 1039 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1040 if (!IdxVal) 1041 return MarkUnsafe(Info); 1042 } 1043 1044 // Compute the offset due to this GEP and check if the alloca has a 1045 // component element at that offset. 1046 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1047 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1048 &Indices[0], Indices.size()); 1049 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 1050 MarkUnsafe(Info); 1051} 1052 1053/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1054/// alloca or has an offset and size that corresponds to a component element 1055/// within it. The offset checked here may have been formed from a GEP with a 1056/// pointer bitcasted to a different type. 1057void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 1058 const Type *MemOpType, bool isStore, 1059 AllocaInfo &Info) { 1060 // Check if this is a load/store of the entire alloca. 1061 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 1062 bool UsesAggregateType = (MemOpType == AI->getAllocatedType()); 1063 // This is safe for MemIntrinsics (where MemOpType is 0), integer types 1064 // (which are essentially the same as the MemIntrinsics, especially with 1065 // regard to copying padding between elements), or references using the 1066 // aggregate type of the alloca. 1067 if (!MemOpType || MemOpType->isIntegerTy() || UsesAggregateType) { 1068 if (!UsesAggregateType) { 1069 if (isStore) 1070 Info.isMemCpyDst = true; 1071 else 1072 Info.isMemCpySrc = true; 1073 } 1074 return; 1075 } 1076 } 1077 // Check if the offset/size correspond to a component within the alloca type. 1078 const Type *T = AI->getAllocatedType(); 1079 if (TypeHasComponent(T, Offset, MemSize)) 1080 return; 1081 1082 return MarkUnsafe(Info); 1083} 1084 1085/// TypeHasComponent - Return true if T has a component type with the 1086/// specified offset and size. If Size is zero, do not check the size. 1087bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1088 const Type *EltTy; 1089 uint64_t EltSize; 1090 if (const StructType *ST = dyn_cast<StructType>(T)) { 1091 const StructLayout *Layout = TD->getStructLayout(ST); 1092 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1093 EltTy = ST->getContainedType(EltIdx); 1094 EltSize = TD->getTypeAllocSize(EltTy); 1095 Offset -= Layout->getElementOffset(EltIdx); 1096 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1097 EltTy = AT->getElementType(); 1098 EltSize = TD->getTypeAllocSize(EltTy); 1099 if (Offset >= AT->getNumElements() * EltSize) 1100 return false; 1101 Offset %= EltSize; 1102 } else { 1103 return false; 1104 } 1105 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1106 return true; 1107 // Check if the component spans multiple elements. 1108 if (Offset + Size > EltSize) 1109 return false; 1110 return TypeHasComponent(EltTy, Offset, Size); 1111} 1112 1113/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1114/// the instruction I, which references it, to use the separate elements. 1115/// Offset indicates the position within AI that is referenced by this 1116/// instruction. 1117void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1118 SmallVector<AllocaInst*, 32> &NewElts) { 1119 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1120 Instruction *User = cast<Instruction>(*UI); 1121 1122 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1123 RewriteBitCast(BC, AI, Offset, NewElts); 1124 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1125 RewriteGEP(GEPI, AI, Offset, NewElts); 1126 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1127 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1128 uint64_t MemSize = Length->getZExtValue(); 1129 if (Offset == 0 && 1130 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1131 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1132 // Otherwise the intrinsic can only touch a single element and the 1133 // address operand will be updated, so nothing else needs to be done. 1134 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1135 const Type *LIType = LI->getType(); 1136 if (LIType == AI->getAllocatedType()) { 1137 // Replace: 1138 // %res = load { i32, i32 }* %alloc 1139 // with: 1140 // %load.0 = load i32* %alloc.0 1141 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1142 // %load.1 = load i32* %alloc.1 1143 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1144 // (Also works for arrays instead of structs) 1145 Value *Insert = UndefValue::get(LIType); 1146 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1147 Value *Load = new LoadInst(NewElts[i], "load", LI); 1148 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1149 } 1150 LI->replaceAllUsesWith(Insert); 1151 DeadInsts.push_back(LI); 1152 } else if (LIType->isIntegerTy() && 1153 TD->getTypeAllocSize(LIType) == 1154 TD->getTypeAllocSize(AI->getAllocatedType())) { 1155 // If this is a load of the entire alloca to an integer, rewrite it. 1156 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1157 } 1158 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1159 Value *Val = SI->getOperand(0); 1160 const Type *SIType = Val->getType(); 1161 if (SIType == AI->getAllocatedType()) { 1162 // Replace: 1163 // store { i32, i32 } %val, { i32, i32 }* %alloc 1164 // with: 1165 // %val.0 = extractvalue { i32, i32 } %val, 0 1166 // store i32 %val.0, i32* %alloc.0 1167 // %val.1 = extractvalue { i32, i32 } %val, 1 1168 // store i32 %val.1, i32* %alloc.1 1169 // (Also works for arrays instead of structs) 1170 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1171 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1172 new StoreInst(Extract, NewElts[i], SI); 1173 } 1174 DeadInsts.push_back(SI); 1175 } else if (SIType->isIntegerTy() && 1176 TD->getTypeAllocSize(SIType) == 1177 TD->getTypeAllocSize(AI->getAllocatedType())) { 1178 // If this is a store of the entire alloca from an integer, rewrite it. 1179 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1180 } 1181 } 1182 } 1183} 1184 1185/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1186/// and recursively continue updating all of its uses. 1187void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1188 SmallVector<AllocaInst*, 32> &NewElts) { 1189 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1190 if (BC->getOperand(0) != AI) 1191 return; 1192 1193 // The bitcast references the original alloca. Replace its uses with 1194 // references to the first new element alloca. 1195 Instruction *Val = NewElts[0]; 1196 if (Val->getType() != BC->getDestTy()) { 1197 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1198 Val->takeName(BC); 1199 } 1200 BC->replaceAllUsesWith(Val); 1201 DeadInsts.push_back(BC); 1202} 1203 1204/// FindElementAndOffset - Return the index of the element containing Offset 1205/// within the specified type, which must be either a struct or an array. 1206/// Sets T to the type of the element and Offset to the offset within that 1207/// element. IdxTy is set to the type of the index result to be used in a 1208/// GEP instruction. 1209uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1210 const Type *&IdxTy) { 1211 uint64_t Idx = 0; 1212 if (const StructType *ST = dyn_cast<StructType>(T)) { 1213 const StructLayout *Layout = TD->getStructLayout(ST); 1214 Idx = Layout->getElementContainingOffset(Offset); 1215 T = ST->getContainedType(Idx); 1216 Offset -= Layout->getElementOffset(Idx); 1217 IdxTy = Type::getInt32Ty(T->getContext()); 1218 return Idx; 1219 } 1220 const ArrayType *AT = cast<ArrayType>(T); 1221 T = AT->getElementType(); 1222 uint64_t EltSize = TD->getTypeAllocSize(T); 1223 Idx = Offset / EltSize; 1224 Offset -= Idx * EltSize; 1225 IdxTy = Type::getInt64Ty(T->getContext()); 1226 return Idx; 1227} 1228 1229/// RewriteGEP - Check if this GEP instruction moves the pointer across 1230/// elements of the alloca that are being split apart, and if so, rewrite 1231/// the GEP to be relative to the new element. 1232void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1233 SmallVector<AllocaInst*, 32> &NewElts) { 1234 uint64_t OldOffset = Offset; 1235 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1236 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1237 &Indices[0], Indices.size()); 1238 1239 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1240 1241 const Type *T = AI->getAllocatedType(); 1242 const Type *IdxTy; 1243 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1244 if (GEPI->getOperand(0) == AI) 1245 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1246 1247 T = AI->getAllocatedType(); 1248 uint64_t EltOffset = Offset; 1249 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1250 1251 // If this GEP does not move the pointer across elements of the alloca 1252 // being split, then it does not needs to be rewritten. 1253 if (Idx == OldIdx) 1254 return; 1255 1256 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1257 SmallVector<Value*, 8> NewArgs; 1258 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1259 while (EltOffset != 0) { 1260 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1261 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1262 } 1263 Instruction *Val = NewElts[Idx]; 1264 if (NewArgs.size() > 1) { 1265 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1266 NewArgs.end(), "", GEPI); 1267 Val->takeName(GEPI); 1268 } 1269 if (Val->getType() != GEPI->getType()) 1270 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1271 GEPI->replaceAllUsesWith(Val); 1272 DeadInsts.push_back(GEPI); 1273} 1274 1275/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1276/// Rewrite it to copy or set the elements of the scalarized memory. 1277void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1278 AllocaInst *AI, 1279 SmallVector<AllocaInst*, 32> &NewElts) { 1280 // If this is a memcpy/memmove, construct the other pointer as the 1281 // appropriate type. The "Other" pointer is the pointer that goes to memory 1282 // that doesn't have anything to do with the alloca that we are promoting. For 1283 // memset, this Value* stays null. 1284 Value *OtherPtr = 0; 1285 unsigned MemAlignment = MI->getAlignment(); 1286 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1287 if (Inst == MTI->getRawDest()) 1288 OtherPtr = MTI->getRawSource(); 1289 else { 1290 assert(Inst == MTI->getRawSource()); 1291 OtherPtr = MTI->getRawDest(); 1292 } 1293 } 1294 1295 // If there is an other pointer, we want to convert it to the same pointer 1296 // type as AI has, so we can GEP through it safely. 1297 if (OtherPtr) { 1298 unsigned AddrSpace = 1299 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 1300 1301 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1302 // optimization, but it's also required to detect the corner case where 1303 // both pointer operands are referencing the same memory, and where 1304 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1305 // function is only called for mem intrinsics that access the whole 1306 // aggregate, so non-zero GEPs are not an issue here.) 1307 OtherPtr = OtherPtr->stripPointerCasts(); 1308 1309 // Copying the alloca to itself is a no-op: just delete it. 1310 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1311 // This code will run twice for a no-op memcpy -- once for each operand. 1312 // Put only one reference to MI on the DeadInsts list. 1313 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1314 E = DeadInsts.end(); I != E; ++I) 1315 if (*I == MI) return; 1316 DeadInsts.push_back(MI); 1317 return; 1318 } 1319 1320 // If the pointer is not the right type, insert a bitcast to the right 1321 // type. 1322 const Type *NewTy = 1323 PointerType::get(AI->getType()->getElementType(), AddrSpace); 1324 1325 if (OtherPtr->getType() != NewTy) 1326 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 1327 } 1328 1329 // Process each element of the aggregate. 1330 Value *TheFn = MI->getCalledValue(); 1331 const Type *BytePtrTy = MI->getRawDest()->getType(); 1332 bool SROADest = MI->getRawDest() == Inst; 1333 1334 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1335 1336 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1337 // If this is a memcpy/memmove, emit a GEP of the other element address. 1338 Value *OtherElt = 0; 1339 unsigned OtherEltAlign = MemAlignment; 1340 1341 if (OtherPtr) { 1342 Value *Idx[2] = { Zero, 1343 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1344 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1345 OtherPtr->getName()+"."+Twine(i), 1346 MI); 1347 uint64_t EltOffset; 1348 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1349 const Type *OtherTy = OtherPtrTy->getElementType(); 1350 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1351 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1352 } else { 1353 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 1354 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1355 } 1356 1357 // The alignment of the other pointer is the guaranteed alignment of the 1358 // element, which is affected by both the known alignment of the whole 1359 // mem intrinsic and the alignment of the element. If the alignment of 1360 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 1361 // known alignment is just 4 bytes. 1362 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 1363 } 1364 1365 Value *EltPtr = NewElts[i]; 1366 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 1367 1368 // If we got down to a scalar, insert a load or store as appropriate. 1369 if (EltTy->isSingleValueType()) { 1370 if (isa<MemTransferInst>(MI)) { 1371 if (SROADest) { 1372 // From Other to Alloca. 1373 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 1374 new StoreInst(Elt, EltPtr, MI); 1375 } else { 1376 // From Alloca to Other. 1377 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 1378 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 1379 } 1380 continue; 1381 } 1382 assert(isa<MemSetInst>(MI)); 1383 1384 // If the stored element is zero (common case), just store a null 1385 // constant. 1386 Constant *StoreVal; 1387 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 1388 if (CI->isZero()) { 1389 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 1390 } else { 1391 // If EltTy is a vector type, get the element type. 1392 const Type *ValTy = EltTy->getScalarType(); 1393 1394 // Construct an integer with the right value. 1395 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 1396 APInt OneVal(EltSize, CI->getZExtValue()); 1397 APInt TotalVal(OneVal); 1398 // Set each byte. 1399 for (unsigned i = 0; 8*i < EltSize; ++i) { 1400 TotalVal = TotalVal.shl(8); 1401 TotalVal |= OneVal; 1402 } 1403 1404 // Convert the integer value to the appropriate type. 1405 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 1406 if (ValTy->isPointerTy()) 1407 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 1408 else if (ValTy->isFloatingPointTy()) 1409 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 1410 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 1411 1412 // If the requested value was a vector constant, create it. 1413 if (EltTy != ValTy) { 1414 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 1415 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 1416 StoreVal = ConstantVector::get(&Elts[0], NumElts); 1417 } 1418 } 1419 new StoreInst(StoreVal, EltPtr, MI); 1420 continue; 1421 } 1422 // Otherwise, if we're storing a byte variable, use a memset call for 1423 // this element. 1424 } 1425 1426 // Cast the element pointer to BytePtrTy. 1427 if (EltPtr->getType() != BytePtrTy) 1428 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI); 1429 1430 // Cast the other pointer (if we have one) to BytePtrTy. 1431 if (OtherElt && OtherElt->getType() != BytePtrTy) { 1432 // Preserve address space of OtherElt 1433 const PointerType* OtherPTy = cast<PointerType>(OtherElt->getType()); 1434 const PointerType* PTy = cast<PointerType>(BytePtrTy); 1435 if (OtherPTy->getElementType() != PTy->getElementType()) { 1436 Type *NewOtherPTy = PointerType::get(PTy->getElementType(), 1437 OtherPTy->getAddressSpace()); 1438 OtherElt = new BitCastInst(OtherElt, NewOtherPTy, 1439 OtherElt->getNameStr(), MI); 1440 } 1441 } 1442 1443 unsigned EltSize = TD->getTypeAllocSize(EltTy); 1444 1445 // Finally, insert the meminst for this element. 1446 if (isa<MemTransferInst>(MI)) { 1447 Value *Ops[] = { 1448 SROADest ? EltPtr : OtherElt, // Dest ptr 1449 SROADest ? OtherElt : EltPtr, // Src ptr 1450 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size 1451 // Align 1452 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign), 1453 MI->getVolatileCst() 1454 }; 1455 // In case we fold the address space overloaded memcpy of A to B 1456 // with memcpy of B to C, change the function to be a memcpy of A to C. 1457 const Type *Tys[] = { Ops[0]->getType(), Ops[1]->getType(), 1458 Ops[2]->getType() }; 1459 Module *M = MI->getParent()->getParent()->getParent(); 1460 TheFn = Intrinsic::getDeclaration(M, MI->getIntrinsicID(), Tys, 3); 1461 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1462 } else { 1463 assert(isa<MemSetInst>(MI)); 1464 Value *Ops[] = { 1465 EltPtr, MI->getArgOperand(1), // Dest, Value, 1466 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size 1467 Zero, // Align 1468 ConstantInt::get(Type::getInt1Ty(MI->getContext()), 0) // isVolatile 1469 }; 1470 const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() }; 1471 Module *M = MI->getParent()->getParent()->getParent(); 1472 TheFn = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2); 1473 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1474 } 1475 } 1476 DeadInsts.push_back(MI); 1477} 1478 1479/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 1480/// overwrites the entire allocation. Extract out the pieces of the stored 1481/// integer and store them individually. 1482void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 1483 SmallVector<AllocaInst*, 32> &NewElts){ 1484 // Extract each element out of the integer according to its structure offset 1485 // and store the element value to the individual alloca. 1486 Value *SrcVal = SI->getOperand(0); 1487 const Type *AllocaEltTy = AI->getAllocatedType(); 1488 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1489 1490 // Handle tail padding by extending the operand 1491 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 1492 SrcVal = new ZExtInst(SrcVal, 1493 IntegerType::get(SI->getContext(), AllocaSizeBits), 1494 "", SI); 1495 1496 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 1497 << '\n'); 1498 1499 // There are two forms here: AI could be an array or struct. Both cases 1500 // have different ways to compute the element offset. 1501 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1502 const StructLayout *Layout = TD->getStructLayout(EltSTy); 1503 1504 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1505 // Get the number of bits to shift SrcVal to get the value. 1506 const Type *FieldTy = EltSTy->getElementType(i); 1507 uint64_t Shift = Layout->getElementOffsetInBits(i); 1508 1509 if (TD->isBigEndian()) 1510 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 1511 1512 Value *EltVal = SrcVal; 1513 if (Shift) { 1514 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1515 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1516 "sroa.store.elt", SI); 1517 } 1518 1519 // Truncate down to an integer of the right size. 1520 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1521 1522 // Ignore zero sized fields like {}, they obviously contain no data. 1523 if (FieldSizeBits == 0) continue; 1524 1525 if (FieldSizeBits != AllocaSizeBits) 1526 EltVal = new TruncInst(EltVal, 1527 IntegerType::get(SI->getContext(), FieldSizeBits), 1528 "", SI); 1529 Value *DestField = NewElts[i]; 1530 if (EltVal->getType() == FieldTy) { 1531 // Storing to an integer field of this size, just do it. 1532 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 1533 // Bitcast to the right element type (for fp/vector values). 1534 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 1535 } else { 1536 // Otherwise, bitcast the dest pointer (for aggregates). 1537 DestField = new BitCastInst(DestField, 1538 PointerType::getUnqual(EltVal->getType()), 1539 "", SI); 1540 } 1541 new StoreInst(EltVal, DestField, SI); 1542 } 1543 1544 } else { 1545 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1546 const Type *ArrayEltTy = ATy->getElementType(); 1547 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1548 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1549 1550 uint64_t Shift; 1551 1552 if (TD->isBigEndian()) 1553 Shift = AllocaSizeBits-ElementOffset; 1554 else 1555 Shift = 0; 1556 1557 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1558 // Ignore zero sized fields like {}, they obviously contain no data. 1559 if (ElementSizeBits == 0) continue; 1560 1561 Value *EltVal = SrcVal; 1562 if (Shift) { 1563 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1564 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1565 "sroa.store.elt", SI); 1566 } 1567 1568 // Truncate down to an integer of the right size. 1569 if (ElementSizeBits != AllocaSizeBits) 1570 EltVal = new TruncInst(EltVal, 1571 IntegerType::get(SI->getContext(), 1572 ElementSizeBits),"",SI); 1573 Value *DestField = NewElts[i]; 1574 if (EltVal->getType() == ArrayEltTy) { 1575 // Storing to an integer field of this size, just do it. 1576 } else if (ArrayEltTy->isFloatingPointTy() || 1577 ArrayEltTy->isVectorTy()) { 1578 // Bitcast to the right element type (for fp/vector values). 1579 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1580 } else { 1581 // Otherwise, bitcast the dest pointer (for aggregates). 1582 DestField = new BitCastInst(DestField, 1583 PointerType::getUnqual(EltVal->getType()), 1584 "", SI); 1585 } 1586 new StoreInst(EltVal, DestField, SI); 1587 1588 if (TD->isBigEndian()) 1589 Shift -= ElementOffset; 1590 else 1591 Shift += ElementOffset; 1592 } 1593 } 1594 1595 DeadInsts.push_back(SI); 1596} 1597 1598/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1599/// an integer. Load the individual pieces to form the aggregate value. 1600void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1601 SmallVector<AllocaInst*, 32> &NewElts) { 1602 // Extract each element out of the NewElts according to its structure offset 1603 // and form the result value. 1604 const Type *AllocaEltTy = AI->getAllocatedType(); 1605 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1606 1607 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1608 << '\n'); 1609 1610 // There are two forms here: AI could be an array or struct. Both cases 1611 // have different ways to compute the element offset. 1612 const StructLayout *Layout = 0; 1613 uint64_t ArrayEltBitOffset = 0; 1614 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1615 Layout = TD->getStructLayout(EltSTy); 1616 } else { 1617 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1618 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1619 } 1620 1621 Value *ResultVal = 1622 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1623 1624 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1625 // Load the value from the alloca. If the NewElt is an aggregate, cast 1626 // the pointer to an integer of the same size before doing the load. 1627 Value *SrcField = NewElts[i]; 1628 const Type *FieldTy = 1629 cast<PointerType>(SrcField->getType())->getElementType(); 1630 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1631 1632 // Ignore zero sized fields like {}, they obviously contain no data. 1633 if (FieldSizeBits == 0) continue; 1634 1635 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1636 FieldSizeBits); 1637 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1638 !FieldTy->isVectorTy()) 1639 SrcField = new BitCastInst(SrcField, 1640 PointerType::getUnqual(FieldIntTy), 1641 "", LI); 1642 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1643 1644 // If SrcField is a fp or vector of the right size but that isn't an 1645 // integer type, bitcast to an integer so we can shift it. 1646 if (SrcField->getType() != FieldIntTy) 1647 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1648 1649 // Zero extend the field to be the same size as the final alloca so that 1650 // we can shift and insert it. 1651 if (SrcField->getType() != ResultVal->getType()) 1652 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1653 1654 // Determine the number of bits to shift SrcField. 1655 uint64_t Shift; 1656 if (Layout) // Struct case. 1657 Shift = Layout->getElementOffsetInBits(i); 1658 else // Array case. 1659 Shift = i*ArrayEltBitOffset; 1660 1661 if (TD->isBigEndian()) 1662 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1663 1664 if (Shift) { 1665 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1666 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1667 } 1668 1669 // Don't create an 'or x, 0' on the first iteration. 1670 if (!isa<Constant>(ResultVal) || 1671 !cast<Constant>(ResultVal)->isNullValue()) 1672 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1673 else 1674 ResultVal = SrcField; 1675 } 1676 1677 // Handle tail padding by truncating the result 1678 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1679 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1680 1681 LI->replaceAllUsesWith(ResultVal); 1682 DeadInsts.push_back(LI); 1683} 1684 1685/// HasPadding - Return true if the specified type has any structure or 1686/// alignment padding, false otherwise. 1687static bool HasPadding(const Type *Ty, const TargetData &TD) { 1688 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 1689 return HasPadding(ATy->getElementType(), TD); 1690 1691 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) 1692 return HasPadding(VTy->getElementType(), TD); 1693 1694 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1695 const StructLayout *SL = TD.getStructLayout(STy); 1696 unsigned PrevFieldBitOffset = 0; 1697 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1698 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1699 1700 // Padding in sub-elements? 1701 if (HasPadding(STy->getElementType(i), TD)) 1702 return true; 1703 1704 // Check to see if there is any padding between this element and the 1705 // previous one. 1706 if (i) { 1707 unsigned PrevFieldEnd = 1708 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1709 if (PrevFieldEnd < FieldBitOffset) 1710 return true; 1711 } 1712 1713 PrevFieldBitOffset = FieldBitOffset; 1714 } 1715 1716 // Check for tail padding. 1717 if (unsigned EltCount = STy->getNumElements()) { 1718 unsigned PrevFieldEnd = PrevFieldBitOffset + 1719 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1720 if (PrevFieldEnd < SL->getSizeInBits()) 1721 return true; 1722 } 1723 } 1724 1725 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1726} 1727 1728/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1729/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1730/// or 1 if safe after canonicalization has been performed. 1731bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1732 // Loop over the use list of the alloca. We can only transform it if all of 1733 // the users are safe to transform. 1734 AllocaInfo Info; 1735 1736 isSafeForScalarRepl(AI, AI, 0, Info); 1737 if (Info.isUnsafe) { 1738 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1739 return false; 1740 } 1741 1742 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1743 // source and destination, we have to be careful. In particular, the memcpy 1744 // could be moving around elements that live in structure padding of the LLVM 1745 // types, but may actually be used. In these cases, we refuse to promote the 1746 // struct. 1747 if (Info.isMemCpySrc && Info.isMemCpyDst && 1748 HasPadding(AI->getAllocatedType(), *TD)) 1749 return false; 1750 1751 return true; 1752} 1753 1754 1755 1756/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1757/// some part of a constant global variable. This intentionally only accepts 1758/// constant expressions because we don't can't rewrite arbitrary instructions. 1759static bool PointsToConstantGlobal(Value *V) { 1760 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1761 return GV->isConstant(); 1762 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1763 if (CE->getOpcode() == Instruction::BitCast || 1764 CE->getOpcode() == Instruction::GetElementPtr) 1765 return PointsToConstantGlobal(CE->getOperand(0)); 1766 return false; 1767} 1768 1769/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1770/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1771/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1772/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1773/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1774/// the alloca, and if the source pointer is a pointer to a constant global, we 1775/// can optimize this. 1776static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 1777 bool isOffset) { 1778 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1779 User *U = cast<Instruction>(*UI); 1780 1781 if (LoadInst *LI = dyn_cast<LoadInst>(U)) 1782 // Ignore non-volatile loads, they are always ok. 1783 if (!LI->isVolatile()) 1784 continue; 1785 1786 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1787 // If uses of the bitcast are ok, we are ok. 1788 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1789 return false; 1790 continue; 1791 } 1792 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 1793 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1794 // doesn't, it does. 1795 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1796 isOffset || !GEP->hasAllZeroIndices())) 1797 return false; 1798 continue; 1799 } 1800 1801 // If this is isn't our memcpy/memmove, reject it as something we can't 1802 // handle. 1803 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 1804 if (MI == 0) 1805 return false; 1806 1807 // If we already have seen a copy, reject the second one. 1808 if (TheCopy) return false; 1809 1810 // If the pointer has been offset from the start of the alloca, we can't 1811 // safely handle this. 1812 if (isOffset) return false; 1813 1814 // If the memintrinsic isn't using the alloca as the dest, reject it. 1815 if (UI.getOperandNo() != 0) return false; 1816 1817 // If the source of the memcpy/move is not a constant global, reject it. 1818 if (!PointsToConstantGlobal(MI->getSource())) 1819 return false; 1820 1821 // Otherwise, the transform is safe. Remember the copy instruction. 1822 TheCopy = MI; 1823 } 1824 return true; 1825} 1826 1827/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1828/// modified by a copy from a constant global. If we can prove this, we can 1829/// replace any uses of the alloca with uses of the global directly. 1830MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1831 MemTransferInst *TheCopy = 0; 1832 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1833 return TheCopy; 1834 return 0; 1835} 1836