ScalarReplAggregates.cpp revision 0488fb649a56b7fc89a5814df5308813f9e5a85d
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 // Don't touch MMX operations. 325 if (LI->getType()->isX86_MMXTy()) 326 return false; 327 MergeInType(LI->getType(), Offset); 328 continue; 329 } 330 331 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 332 // Storing the pointer, not into the value? 333 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 334 // Don't touch MMX operations. 335 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 336 return false; 337 MergeInType(SI->getOperand(0)->getType(), Offset); 338 continue; 339 } 340 341 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 342 IsNotTrivial = true; // Can't be mem2reg'd. 343 if (!CanConvertToScalar(BCI, Offset)) 344 return false; 345 continue; 346 } 347 348 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 349 // If this is a GEP with a variable indices, we can't handle it. 350 if (!GEP->hasAllConstantIndices()) 351 return false; 352 353 // Compute the offset that this GEP adds to the pointer. 354 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 355 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 356 &Indices[0], Indices.size()); 357 // See if all uses can be converted. 358 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 359 return false; 360 IsNotTrivial = true; // Can't be mem2reg'd. 361 continue; 362 } 363 364 // If this is a constant sized memset of a constant value (e.g. 0) we can 365 // handle it. 366 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 367 // Store of constant value and constant size. 368 if (!isa<ConstantInt>(MSI->getValue()) || 369 !isa<ConstantInt>(MSI->getLength())) 370 return false; 371 IsNotTrivial = true; // Can't be mem2reg'd. 372 continue; 373 } 374 375 // If this is a memcpy or memmove into or out of the whole allocation, we 376 // can handle it like a load or store of the scalar type. 377 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 378 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 379 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 380 return false; 381 382 IsNotTrivial = true; // Can't be mem2reg'd. 383 continue; 384 } 385 386 // Otherwise, we cannot handle this! 387 return false; 388 } 389 390 return true; 391} 392 393/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 394/// directly. This happens when we are converting an "integer union" to a 395/// single integer scalar, or when we are converting a "vector union" to a 396/// vector with insert/extractelement instructions. 397/// 398/// Offset is an offset from the original alloca, in bits that need to be 399/// shifted to the right. By the end of this, there should be no uses of Ptr. 400void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 401 uint64_t Offset) { 402 while (!Ptr->use_empty()) { 403 Instruction *User = cast<Instruction>(Ptr->use_back()); 404 405 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 406 ConvertUsesToScalar(CI, NewAI, Offset); 407 CI->eraseFromParent(); 408 continue; 409 } 410 411 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 412 // Compute the offset that this GEP adds to the pointer. 413 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 414 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 415 &Indices[0], Indices.size()); 416 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 417 GEP->eraseFromParent(); 418 continue; 419 } 420 421 IRBuilder<> Builder(User->getParent(), User); 422 423 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 424 // The load is a bit extract from NewAI shifted right by Offset bits. 425 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 426 Value *NewLoadVal 427 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 428 LI->replaceAllUsesWith(NewLoadVal); 429 LI->eraseFromParent(); 430 continue; 431 } 432 433 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 434 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 435 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 436 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 437 Builder); 438 Builder.CreateStore(New, NewAI); 439 SI->eraseFromParent(); 440 441 // If the load we just inserted is now dead, then the inserted store 442 // overwrote the entire thing. 443 if (Old->use_empty()) 444 Old->eraseFromParent(); 445 continue; 446 } 447 448 // If this is a constant sized memset of a constant value (e.g. 0) we can 449 // transform it into a store of the expanded constant value. 450 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 451 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 452 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 453 if (NumBytes != 0) { 454 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 455 456 // Compute the value replicated the right number of times. 457 APInt APVal(NumBytes*8, Val); 458 459 // Splat the value if non-zero. 460 if (Val) 461 for (unsigned i = 1; i != NumBytes; ++i) 462 APVal |= APVal << 8; 463 464 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 465 Value *New = ConvertScalar_InsertValue( 466 ConstantInt::get(User->getContext(), APVal), 467 Old, Offset, Builder); 468 Builder.CreateStore(New, NewAI); 469 470 // If the load we just inserted is now dead, then the memset overwrote 471 // the entire thing. 472 if (Old->use_empty()) 473 Old->eraseFromParent(); 474 } 475 MSI->eraseFromParent(); 476 continue; 477 } 478 479 // If this is a memcpy or memmove into or out of the whole allocation, we 480 // can handle it like a load or store of the scalar type. 481 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 482 assert(Offset == 0 && "must be store to start of alloca"); 483 484 // If the source and destination are both to the same alloca, then this is 485 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 486 // as appropriate. 487 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject(0)); 488 489 if (MTI->getSource()->getUnderlyingObject(0) != OrigAI) { 490 // Dest must be OrigAI, change this to be a load from the original 491 // pointer (bitcasted), then a store to our new alloca. 492 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 493 Value *SrcPtr = MTI->getSource(); 494 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 495 496 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 497 SrcVal->setAlignment(MTI->getAlignment()); 498 Builder.CreateStore(SrcVal, NewAI); 499 } else if (MTI->getDest()->getUnderlyingObject(0) != OrigAI) { 500 // Src must be OrigAI, change this to be a load from NewAI then a store 501 // through the original dest pointer (bitcasted). 502 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 503 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 504 505 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 506 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 507 NewStore->setAlignment(MTI->getAlignment()); 508 } else { 509 // Noop transfer. Src == Dst 510 } 511 512 MTI->eraseFromParent(); 513 continue; 514 } 515 516 llvm_unreachable("Unsupported operation!"); 517 } 518} 519 520/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 521/// or vector value FromVal, extracting the bits from the offset specified by 522/// Offset. This returns the value, which is of type ToType. 523/// 524/// This happens when we are converting an "integer union" to a single 525/// integer scalar, or when we are converting a "vector union" to a vector with 526/// insert/extractelement instructions. 527/// 528/// Offset is an offset from the original alloca, in bits that need to be 529/// shifted to the right. 530Value *ConvertToScalarInfo:: 531ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 532 uint64_t Offset, IRBuilder<> &Builder) { 533 // If the load is of the whole new alloca, no conversion is needed. 534 if (FromVal->getType() == ToType && Offset == 0) 535 return FromVal; 536 537 // If the result alloca is a vector type, this is either an element 538 // access or a bitcast to another vector type of the same size. 539 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 540 if (ToType->isVectorTy()) 541 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 542 543 // Otherwise it must be an element access. 544 unsigned Elt = 0; 545 if (Offset) { 546 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 547 Elt = Offset/EltSize; 548 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 549 } 550 // Return the element extracted out of it. 551 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 552 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 553 if (V->getType() != ToType) 554 V = Builder.CreateBitCast(V, ToType, "tmp"); 555 return V; 556 } 557 558 // If ToType is a first class aggregate, extract out each of the pieces and 559 // use insertvalue's to form the FCA. 560 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 561 const StructLayout &Layout = *TD.getStructLayout(ST); 562 Value *Res = UndefValue::get(ST); 563 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 564 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 565 Offset+Layout.getElementOffsetInBits(i), 566 Builder); 567 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 568 } 569 return Res; 570 } 571 572 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 573 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 574 Value *Res = UndefValue::get(AT); 575 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 576 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 577 Offset+i*EltSize, Builder); 578 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 579 } 580 return Res; 581 } 582 583 // Otherwise, this must be a union that was converted to an integer value. 584 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 585 586 // If this is a big-endian system and the load is narrower than the 587 // full alloca type, we need to do a shift to get the right bits. 588 int ShAmt = 0; 589 if (TD.isBigEndian()) { 590 // On big-endian machines, the lowest bit is stored at the bit offset 591 // from the pointer given by getTypeStoreSizeInBits. This matters for 592 // integers with a bitwidth that is not a multiple of 8. 593 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 594 TD.getTypeStoreSizeInBits(ToType) - Offset; 595 } else { 596 ShAmt = Offset; 597 } 598 599 // Note: we support negative bitwidths (with shl) which are not defined. 600 // We do this to support (f.e.) loads off the end of a structure where 601 // only some bits are used. 602 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 603 FromVal = Builder.CreateLShr(FromVal, 604 ConstantInt::get(FromVal->getType(), 605 ShAmt), "tmp"); 606 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 607 FromVal = Builder.CreateShl(FromVal, 608 ConstantInt::get(FromVal->getType(), 609 -ShAmt), "tmp"); 610 611 // Finally, unconditionally truncate the integer to the right width. 612 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 613 if (LIBitWidth < NTy->getBitWidth()) 614 FromVal = 615 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 616 LIBitWidth), "tmp"); 617 else if (LIBitWidth > NTy->getBitWidth()) 618 FromVal = 619 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 620 LIBitWidth), "tmp"); 621 622 // If the result is an integer, this is a trunc or bitcast. 623 if (ToType->isIntegerTy()) { 624 // Should be done. 625 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 626 // Just do a bitcast, we know the sizes match up. 627 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 628 } else { 629 // Otherwise must be a pointer. 630 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 631 } 632 assert(FromVal->getType() == ToType && "Didn't convert right?"); 633 return FromVal; 634} 635 636/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 637/// or vector value "Old" at the offset specified by Offset. 638/// 639/// This happens when we are converting an "integer union" to a 640/// single integer scalar, or when we are converting a "vector union" to a 641/// vector with insert/extractelement instructions. 642/// 643/// Offset is an offset from the original alloca, in bits that need to be 644/// shifted to the right. 645Value *ConvertToScalarInfo:: 646ConvertScalar_InsertValue(Value *SV, Value *Old, 647 uint64_t Offset, IRBuilder<> &Builder) { 648 // Convert the stored type to the actual type, shift it left to insert 649 // then 'or' into place. 650 const Type *AllocaType = Old->getType(); 651 LLVMContext &Context = Old->getContext(); 652 653 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 654 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 655 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 656 657 // Changing the whole vector with memset or with an access of a different 658 // vector type? 659 if (ValSize == VecSize) 660 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 661 662 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 663 664 // Must be an element insertion. 665 unsigned Elt = Offset/EltSize; 666 667 if (SV->getType() != VTy->getElementType()) 668 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 669 670 SV = Builder.CreateInsertElement(Old, SV, 671 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 672 "tmp"); 673 return SV; 674 } 675 676 // If SV is a first-class aggregate value, insert each value recursively. 677 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 678 const StructLayout &Layout = *TD.getStructLayout(ST); 679 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 680 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 681 Old = ConvertScalar_InsertValue(Elt, Old, 682 Offset+Layout.getElementOffsetInBits(i), 683 Builder); 684 } 685 return Old; 686 } 687 688 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 689 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 690 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 691 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 692 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 693 } 694 return Old; 695 } 696 697 // If SV is a float, convert it to the appropriate integer type. 698 // If it is a pointer, do the same. 699 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 700 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 701 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 702 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 703 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 704 SV = Builder.CreateBitCast(SV, 705 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 706 else if (SV->getType()->isPointerTy()) 707 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 708 709 // Zero extend or truncate the value if needed. 710 if (SV->getType() != AllocaType) { 711 if (SV->getType()->getPrimitiveSizeInBits() < 712 AllocaType->getPrimitiveSizeInBits()) 713 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 714 else { 715 // Truncation may be needed if storing more than the alloca can hold 716 // (undefined behavior). 717 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 718 SrcWidth = DestWidth; 719 SrcStoreWidth = DestStoreWidth; 720 } 721 } 722 723 // If this is a big-endian system and the store is narrower than the 724 // full alloca type, we need to do a shift to get the right bits. 725 int ShAmt = 0; 726 if (TD.isBigEndian()) { 727 // On big-endian machines, the lowest bit is stored at the bit offset 728 // from the pointer given by getTypeStoreSizeInBits. This matters for 729 // integers with a bitwidth that is not a multiple of 8. 730 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 731 } else { 732 ShAmt = Offset; 733 } 734 735 // Note: we support negative bitwidths (with shr) which are not defined. 736 // We do this to support (f.e.) stores off the end of a structure where 737 // only some bits in the structure are set. 738 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 739 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 740 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 741 ShAmt), "tmp"); 742 Mask <<= ShAmt; 743 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 744 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 745 -ShAmt), "tmp"); 746 Mask = Mask.lshr(-ShAmt); 747 } 748 749 // Mask out the bits we are about to insert from the old value, and or 750 // in the new bits. 751 if (SrcWidth != DestWidth) { 752 assert(DestWidth > SrcWidth); 753 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 754 SV = Builder.CreateOr(Old, SV, "ins"); 755 } 756 return SV; 757} 758 759 760//===----------------------------------------------------------------------===// 761// SRoA Driver 762//===----------------------------------------------------------------------===// 763 764 765bool SROA::runOnFunction(Function &F) { 766 TD = getAnalysisIfAvailable<TargetData>(); 767 768 bool Changed = performPromotion(F); 769 770 // FIXME: ScalarRepl currently depends on TargetData more than it 771 // theoretically needs to. It should be refactored in order to support 772 // target-independent IR. Until this is done, just skip the actual 773 // scalar-replacement portion of this pass. 774 if (!TD) return Changed; 775 776 while (1) { 777 bool LocalChange = performScalarRepl(F); 778 if (!LocalChange) break; // No need to repromote if no scalarrepl 779 Changed = true; 780 LocalChange = performPromotion(F); 781 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 782 } 783 784 return Changed; 785} 786 787 788bool SROA::performPromotion(Function &F) { 789 std::vector<AllocaInst*> Allocas; 790 DominatorTree &DT = getAnalysis<DominatorTree>(); 791 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 792 793 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 794 795 bool Changed = false; 796 797 while (1) { 798 Allocas.clear(); 799 800 // Find allocas that are safe to promote, by looking at all instructions in 801 // the entry node 802 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 803 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 804 if (isAllocaPromotable(AI)) 805 Allocas.push_back(AI); 806 807 if (Allocas.empty()) break; 808 809 PromoteMemToReg(Allocas, DT, DF); 810 NumPromoted += Allocas.size(); 811 Changed = true; 812 } 813 814 return Changed; 815} 816 817 818/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 819/// SROA. It must be a struct or array type with a small number of elements. 820static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 821 const Type *T = AI->getAllocatedType(); 822 // Do not promote any struct into more than 32 separate vars. 823 if (const StructType *ST = dyn_cast<StructType>(T)) 824 return ST->getNumElements() <= 32; 825 // Arrays are much less likely to be safe for SROA; only consider 826 // them if they are very small. 827 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 828 return AT->getNumElements() <= 8; 829 return false; 830} 831 832 833// performScalarRepl - This algorithm is a simple worklist driven algorithm, 834// which runs on all of the malloc/alloca instructions in the function, removing 835// them if they are only used by getelementptr instructions. 836// 837bool SROA::performScalarRepl(Function &F) { 838 std::vector<AllocaInst*> WorkList; 839 840 // Scan the entry basic block, adding allocas to the worklist. 841 BasicBlock &BB = F.getEntryBlock(); 842 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 843 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 844 WorkList.push_back(A); 845 846 // Process the worklist 847 bool Changed = false; 848 while (!WorkList.empty()) { 849 AllocaInst *AI = WorkList.back(); 850 WorkList.pop_back(); 851 852 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 853 // with unused elements. 854 if (AI->use_empty()) { 855 AI->eraseFromParent(); 856 Changed = true; 857 continue; 858 } 859 860 // If this alloca is impossible for us to promote, reject it early. 861 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 862 continue; 863 864 // Check to see if this allocation is only modified by a memcpy/memmove from 865 // a constant global. If this is the case, we can change all users to use 866 // the constant global instead. This is commonly produced by the CFE by 867 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 868 // is only subsequently read. 869 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 870 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 871 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 872 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 873 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 874 TheCopy->eraseFromParent(); // Don't mutate the global. 875 AI->eraseFromParent(); 876 ++NumGlobals; 877 Changed = true; 878 continue; 879 } 880 881 // Check to see if we can perform the core SROA transformation. We cannot 882 // transform the allocation instruction if it is an array allocation 883 // (allocations OF arrays are ok though), and an allocation of a scalar 884 // value cannot be decomposed at all. 885 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 886 887 // Do not promote [0 x %struct]. 888 if (AllocaSize == 0) continue; 889 890 // Do not promote any struct whose size is too big. 891 if (AllocaSize > SRThreshold) continue; 892 893 // If the alloca looks like a good candidate for scalar replacement, and if 894 // all its users can be transformed, then split up the aggregate into its 895 // separate elements. 896 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 897 DoScalarReplacement(AI, WorkList); 898 Changed = true; 899 continue; 900 } 901 902 // If we can turn this aggregate value (potentially with casts) into a 903 // simple scalar value that can be mem2reg'd into a register value. 904 // IsNotTrivial tracks whether this is something that mem2reg could have 905 // promoted itself. If so, we don't want to transform it needlessly. Note 906 // that we can't just check based on the type: the alloca may be of an i32 907 // but that has pointer arithmetic to set byte 3 of it or something. 908 if (AllocaInst *NewAI = 909 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 910 NewAI->takeName(AI); 911 AI->eraseFromParent(); 912 ++NumConverted; 913 Changed = true; 914 continue; 915 } 916 917 // Otherwise, couldn't process this alloca. 918 } 919 920 return Changed; 921} 922 923/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 924/// predicate, do SROA now. 925void SROA::DoScalarReplacement(AllocaInst *AI, 926 std::vector<AllocaInst*> &WorkList) { 927 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 928 SmallVector<AllocaInst*, 32> ElementAllocas; 929 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 930 ElementAllocas.reserve(ST->getNumContainedTypes()); 931 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 932 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 933 AI->getAlignment(), 934 AI->getName() + "." + Twine(i), AI); 935 ElementAllocas.push_back(NA); 936 WorkList.push_back(NA); // Add to worklist for recursive processing 937 } 938 } else { 939 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 940 ElementAllocas.reserve(AT->getNumElements()); 941 const Type *ElTy = AT->getElementType(); 942 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 943 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 944 AI->getName() + "." + Twine(i), AI); 945 ElementAllocas.push_back(NA); 946 WorkList.push_back(NA); // Add to worklist for recursive processing 947 } 948 } 949 950 // Now that we have created the new alloca instructions, rewrite all the 951 // uses of the old alloca. 952 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 953 954 // Now erase any instructions that were made dead while rewriting the alloca. 955 DeleteDeadInstructions(); 956 AI->eraseFromParent(); 957 958 ++NumReplaced; 959} 960 961/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 962/// recursively including all their operands that become trivially dead. 963void SROA::DeleteDeadInstructions() { 964 while (!DeadInsts.empty()) { 965 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 966 967 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 968 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 969 // Zero out the operand and see if it becomes trivially dead. 970 // (But, don't add allocas to the dead instruction list -- they are 971 // already on the worklist and will be deleted separately.) 972 *OI = 0; 973 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 974 DeadInsts.push_back(U); 975 } 976 977 I->eraseFromParent(); 978 } 979} 980 981/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 982/// performing scalar replacement of alloca AI. The results are flagged in 983/// the Info parameter. Offset indicates the position within AI that is 984/// referenced by this instruction. 985void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 986 AllocaInfo &Info) { 987 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 988 Instruction *User = cast<Instruction>(*UI); 989 990 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 991 isSafeForScalarRepl(BC, AI, Offset, Info); 992 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 993 uint64_t GEPOffset = Offset; 994 isSafeGEP(GEPI, AI, GEPOffset, Info); 995 if (!Info.isUnsafe) 996 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 997 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 998 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 999 if (Length) 1000 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 1001 UI.getOperandNo() == 0, Info); 1002 else 1003 MarkUnsafe(Info); 1004 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1005 if (!LI->isVolatile()) { 1006 const Type *LIType = LI->getType(); 1007 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 1008 LIType, false, Info); 1009 } else 1010 MarkUnsafe(Info); 1011 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1012 // Store is ok if storing INTO the pointer, not storing the pointer 1013 if (!SI->isVolatile() && SI->getOperand(0) != I) { 1014 const Type *SIType = SI->getOperand(0)->getType(); 1015 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 1016 SIType, true, Info); 1017 } else 1018 MarkUnsafe(Info); 1019 } else { 1020 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 1021 MarkUnsafe(Info); 1022 } 1023 if (Info.isUnsafe) return; 1024 } 1025} 1026 1027/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1028/// replacement. It is safe when all the indices are constant, in-bounds 1029/// references, and when the resulting offset corresponds to an element within 1030/// the alloca type. The results are flagged in the Info parameter. Upon 1031/// return, Offset is adjusted as specified by the GEP indices. 1032void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 1033 uint64_t &Offset, AllocaInfo &Info) { 1034 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1035 if (GEPIt == E) 1036 return; 1037 1038 // Walk through the GEP type indices, checking the types that this indexes 1039 // into. 1040 for (; GEPIt != E; ++GEPIt) { 1041 // Ignore struct elements, no extra checking needed for these. 1042 if ((*GEPIt)->isStructTy()) 1043 continue; 1044 1045 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1046 if (!IdxVal) 1047 return MarkUnsafe(Info); 1048 } 1049 1050 // Compute the offset due to this GEP and check if the alloca has a 1051 // component element at that offset. 1052 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1053 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1054 &Indices[0], Indices.size()); 1055 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 1056 MarkUnsafe(Info); 1057} 1058 1059/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1060/// alloca or has an offset and size that corresponds to a component element 1061/// within it. The offset checked here may have been formed from a GEP with a 1062/// pointer bitcasted to a different type. 1063void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 1064 const Type *MemOpType, bool isStore, 1065 AllocaInfo &Info) { 1066 // Check if this is a load/store of the entire alloca. 1067 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 1068 bool UsesAggregateType = (MemOpType == AI->getAllocatedType()); 1069 // This is safe for MemIntrinsics (where MemOpType is 0), integer types 1070 // (which are essentially the same as the MemIntrinsics, especially with 1071 // regard to copying padding between elements), or references using the 1072 // aggregate type of the alloca. 1073 if (!MemOpType || MemOpType->isIntegerTy() || UsesAggregateType) { 1074 if (!UsesAggregateType) { 1075 if (isStore) 1076 Info.isMemCpyDst = true; 1077 else 1078 Info.isMemCpySrc = true; 1079 } 1080 return; 1081 } 1082 } 1083 // Check if the offset/size correspond to a component within the alloca type. 1084 const Type *T = AI->getAllocatedType(); 1085 if (TypeHasComponent(T, Offset, MemSize)) 1086 return; 1087 1088 return MarkUnsafe(Info); 1089} 1090 1091/// TypeHasComponent - Return true if T has a component type with the 1092/// specified offset and size. If Size is zero, do not check the size. 1093bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1094 const Type *EltTy; 1095 uint64_t EltSize; 1096 if (const StructType *ST = dyn_cast<StructType>(T)) { 1097 const StructLayout *Layout = TD->getStructLayout(ST); 1098 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1099 EltTy = ST->getContainedType(EltIdx); 1100 EltSize = TD->getTypeAllocSize(EltTy); 1101 Offset -= Layout->getElementOffset(EltIdx); 1102 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1103 EltTy = AT->getElementType(); 1104 EltSize = TD->getTypeAllocSize(EltTy); 1105 if (Offset >= AT->getNumElements() * EltSize) 1106 return false; 1107 Offset %= EltSize; 1108 } else { 1109 return false; 1110 } 1111 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1112 return true; 1113 // Check if the component spans multiple elements. 1114 if (Offset + Size > EltSize) 1115 return false; 1116 return TypeHasComponent(EltTy, Offset, Size); 1117} 1118 1119/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1120/// the instruction I, which references it, to use the separate elements. 1121/// Offset indicates the position within AI that is referenced by this 1122/// instruction. 1123void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1124 SmallVector<AllocaInst*, 32> &NewElts) { 1125 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1126 Instruction *User = cast<Instruction>(*UI); 1127 1128 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1129 RewriteBitCast(BC, AI, Offset, NewElts); 1130 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1131 RewriteGEP(GEPI, AI, Offset, NewElts); 1132 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1133 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1134 uint64_t MemSize = Length->getZExtValue(); 1135 if (Offset == 0 && 1136 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1137 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1138 // Otherwise the intrinsic can only touch a single element and the 1139 // address operand will be updated, so nothing else needs to be done. 1140 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1141 const Type *LIType = LI->getType(); 1142 if (LIType == AI->getAllocatedType()) { 1143 // Replace: 1144 // %res = load { i32, i32 }* %alloc 1145 // with: 1146 // %load.0 = load i32* %alloc.0 1147 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1148 // %load.1 = load i32* %alloc.1 1149 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1150 // (Also works for arrays instead of structs) 1151 Value *Insert = UndefValue::get(LIType); 1152 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1153 Value *Load = new LoadInst(NewElts[i], "load", LI); 1154 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1155 } 1156 LI->replaceAllUsesWith(Insert); 1157 DeadInsts.push_back(LI); 1158 } else if (LIType->isIntegerTy() && 1159 TD->getTypeAllocSize(LIType) == 1160 TD->getTypeAllocSize(AI->getAllocatedType())) { 1161 // If this is a load of the entire alloca to an integer, rewrite it. 1162 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1163 } 1164 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1165 Value *Val = SI->getOperand(0); 1166 const Type *SIType = Val->getType(); 1167 if (SIType == AI->getAllocatedType()) { 1168 // Replace: 1169 // store { i32, i32 } %val, { i32, i32 }* %alloc 1170 // with: 1171 // %val.0 = extractvalue { i32, i32 } %val, 0 1172 // store i32 %val.0, i32* %alloc.0 1173 // %val.1 = extractvalue { i32, i32 } %val, 1 1174 // store i32 %val.1, i32* %alloc.1 1175 // (Also works for arrays instead of structs) 1176 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1177 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1178 new StoreInst(Extract, NewElts[i], SI); 1179 } 1180 DeadInsts.push_back(SI); 1181 } else if (SIType->isIntegerTy() && 1182 TD->getTypeAllocSize(SIType) == 1183 TD->getTypeAllocSize(AI->getAllocatedType())) { 1184 // If this is a store of the entire alloca from an integer, rewrite it. 1185 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1186 } 1187 } 1188 } 1189} 1190 1191/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1192/// and recursively continue updating all of its uses. 1193void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1194 SmallVector<AllocaInst*, 32> &NewElts) { 1195 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1196 if (BC->getOperand(0) != AI) 1197 return; 1198 1199 // The bitcast references the original alloca. Replace its uses with 1200 // references to the first new element alloca. 1201 Instruction *Val = NewElts[0]; 1202 if (Val->getType() != BC->getDestTy()) { 1203 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1204 Val->takeName(BC); 1205 } 1206 BC->replaceAllUsesWith(Val); 1207 DeadInsts.push_back(BC); 1208} 1209 1210/// FindElementAndOffset - Return the index of the element containing Offset 1211/// within the specified type, which must be either a struct or an array. 1212/// Sets T to the type of the element and Offset to the offset within that 1213/// element. IdxTy is set to the type of the index result to be used in a 1214/// GEP instruction. 1215uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1216 const Type *&IdxTy) { 1217 uint64_t Idx = 0; 1218 if (const StructType *ST = dyn_cast<StructType>(T)) { 1219 const StructLayout *Layout = TD->getStructLayout(ST); 1220 Idx = Layout->getElementContainingOffset(Offset); 1221 T = ST->getContainedType(Idx); 1222 Offset -= Layout->getElementOffset(Idx); 1223 IdxTy = Type::getInt32Ty(T->getContext()); 1224 return Idx; 1225 } 1226 const ArrayType *AT = cast<ArrayType>(T); 1227 T = AT->getElementType(); 1228 uint64_t EltSize = TD->getTypeAllocSize(T); 1229 Idx = Offset / EltSize; 1230 Offset -= Idx * EltSize; 1231 IdxTy = Type::getInt64Ty(T->getContext()); 1232 return Idx; 1233} 1234 1235/// RewriteGEP - Check if this GEP instruction moves the pointer across 1236/// elements of the alloca that are being split apart, and if so, rewrite 1237/// the GEP to be relative to the new element. 1238void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1239 SmallVector<AllocaInst*, 32> &NewElts) { 1240 uint64_t OldOffset = Offset; 1241 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1242 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1243 &Indices[0], Indices.size()); 1244 1245 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1246 1247 const Type *T = AI->getAllocatedType(); 1248 const Type *IdxTy; 1249 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1250 if (GEPI->getOperand(0) == AI) 1251 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1252 1253 T = AI->getAllocatedType(); 1254 uint64_t EltOffset = Offset; 1255 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1256 1257 // If this GEP does not move the pointer across elements of the alloca 1258 // being split, then it does not needs to be rewritten. 1259 if (Idx == OldIdx) 1260 return; 1261 1262 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1263 SmallVector<Value*, 8> NewArgs; 1264 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1265 while (EltOffset != 0) { 1266 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1267 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1268 } 1269 Instruction *Val = NewElts[Idx]; 1270 if (NewArgs.size() > 1) { 1271 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1272 NewArgs.end(), "", GEPI); 1273 Val->takeName(GEPI); 1274 } 1275 if (Val->getType() != GEPI->getType()) 1276 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1277 GEPI->replaceAllUsesWith(Val); 1278 DeadInsts.push_back(GEPI); 1279} 1280 1281/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1282/// Rewrite it to copy or set the elements of the scalarized memory. 1283void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1284 AllocaInst *AI, 1285 SmallVector<AllocaInst*, 32> &NewElts) { 1286 // If this is a memcpy/memmove, construct the other pointer as the 1287 // appropriate type. The "Other" pointer is the pointer that goes to memory 1288 // that doesn't have anything to do with the alloca that we are promoting. For 1289 // memset, this Value* stays null. 1290 Value *OtherPtr = 0; 1291 unsigned MemAlignment = MI->getAlignment(); 1292 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1293 if (Inst == MTI->getRawDest()) 1294 OtherPtr = MTI->getRawSource(); 1295 else { 1296 assert(Inst == MTI->getRawSource()); 1297 OtherPtr = MTI->getRawDest(); 1298 } 1299 } 1300 1301 // If there is an other pointer, we want to convert it to the same pointer 1302 // type as AI has, so we can GEP through it safely. 1303 if (OtherPtr) { 1304 unsigned AddrSpace = 1305 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 1306 1307 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1308 // optimization, but it's also required to detect the corner case where 1309 // both pointer operands are referencing the same memory, and where 1310 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1311 // function is only called for mem intrinsics that access the whole 1312 // aggregate, so non-zero GEPs are not an issue here.) 1313 OtherPtr = OtherPtr->stripPointerCasts(); 1314 1315 // Copying the alloca to itself is a no-op: just delete it. 1316 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1317 // This code will run twice for a no-op memcpy -- once for each operand. 1318 // Put only one reference to MI on the DeadInsts list. 1319 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1320 E = DeadInsts.end(); I != E; ++I) 1321 if (*I == MI) return; 1322 DeadInsts.push_back(MI); 1323 return; 1324 } 1325 1326 // If the pointer is not the right type, insert a bitcast to the right 1327 // type. 1328 const Type *NewTy = 1329 PointerType::get(AI->getType()->getElementType(), AddrSpace); 1330 1331 if (OtherPtr->getType() != NewTy) 1332 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 1333 } 1334 1335 // Process each element of the aggregate. 1336 Value *TheFn = MI->getCalledValue(); 1337 const Type *BytePtrTy = MI->getRawDest()->getType(); 1338 bool SROADest = MI->getRawDest() == Inst; 1339 1340 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1341 1342 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1343 // If this is a memcpy/memmove, emit a GEP of the other element address. 1344 Value *OtherElt = 0; 1345 unsigned OtherEltAlign = MemAlignment; 1346 1347 if (OtherPtr) { 1348 Value *Idx[2] = { Zero, 1349 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1350 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1351 OtherPtr->getName()+"."+Twine(i), 1352 MI); 1353 uint64_t EltOffset; 1354 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1355 const Type *OtherTy = OtherPtrTy->getElementType(); 1356 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1357 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1358 } else { 1359 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 1360 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1361 } 1362 1363 // The alignment of the other pointer is the guaranteed alignment of the 1364 // element, which is affected by both the known alignment of the whole 1365 // mem intrinsic and the alignment of the element. If the alignment of 1366 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 1367 // known alignment is just 4 bytes. 1368 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 1369 } 1370 1371 Value *EltPtr = NewElts[i]; 1372 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 1373 1374 // If we got down to a scalar, insert a load or store as appropriate. 1375 if (EltTy->isSingleValueType()) { 1376 if (isa<MemTransferInst>(MI)) { 1377 if (SROADest) { 1378 // From Other to Alloca. 1379 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 1380 new StoreInst(Elt, EltPtr, MI); 1381 } else { 1382 // From Alloca to Other. 1383 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 1384 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 1385 } 1386 continue; 1387 } 1388 assert(isa<MemSetInst>(MI)); 1389 1390 // If the stored element is zero (common case), just store a null 1391 // constant. 1392 Constant *StoreVal; 1393 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 1394 if (CI->isZero()) { 1395 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 1396 } else { 1397 // If EltTy is a vector type, get the element type. 1398 const Type *ValTy = EltTy->getScalarType(); 1399 1400 // Construct an integer with the right value. 1401 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 1402 APInt OneVal(EltSize, CI->getZExtValue()); 1403 APInt TotalVal(OneVal); 1404 // Set each byte. 1405 for (unsigned i = 0; 8*i < EltSize; ++i) { 1406 TotalVal = TotalVal.shl(8); 1407 TotalVal |= OneVal; 1408 } 1409 1410 // Convert the integer value to the appropriate type. 1411 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 1412 if (ValTy->isPointerTy()) 1413 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 1414 else if (ValTy->isFloatingPointTy()) 1415 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 1416 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 1417 1418 // If the requested value was a vector constant, create it. 1419 if (EltTy != ValTy) { 1420 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 1421 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 1422 StoreVal = ConstantVector::get(&Elts[0], NumElts); 1423 } 1424 } 1425 new StoreInst(StoreVal, EltPtr, MI); 1426 continue; 1427 } 1428 // Otherwise, if we're storing a byte variable, use a memset call for 1429 // this element. 1430 } 1431 1432 // Cast the element pointer to BytePtrTy. 1433 if (EltPtr->getType() != BytePtrTy) 1434 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI); 1435 1436 // Cast the other pointer (if we have one) to BytePtrTy. 1437 if (OtherElt && OtherElt->getType() != BytePtrTy) { 1438 // Preserve address space of OtherElt 1439 const PointerType* OtherPTy = cast<PointerType>(OtherElt->getType()); 1440 const PointerType* PTy = cast<PointerType>(BytePtrTy); 1441 if (OtherPTy->getElementType() != PTy->getElementType()) { 1442 Type *NewOtherPTy = PointerType::get(PTy->getElementType(), 1443 OtherPTy->getAddressSpace()); 1444 OtherElt = new BitCastInst(OtherElt, NewOtherPTy, 1445 OtherElt->getNameStr(), MI); 1446 } 1447 } 1448 1449 unsigned EltSize = TD->getTypeAllocSize(EltTy); 1450 1451 // Finally, insert the meminst for this element. 1452 if (isa<MemTransferInst>(MI)) { 1453 Value *Ops[] = { 1454 SROADest ? EltPtr : OtherElt, // Dest ptr 1455 SROADest ? OtherElt : EltPtr, // Src ptr 1456 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size 1457 // Align 1458 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign), 1459 MI->getVolatileCst() 1460 }; 1461 // In case we fold the address space overloaded memcpy of A to B 1462 // with memcpy of B to C, change the function to be a memcpy of A to C. 1463 const Type *Tys[] = { Ops[0]->getType(), Ops[1]->getType(), 1464 Ops[2]->getType() }; 1465 Module *M = MI->getParent()->getParent()->getParent(); 1466 TheFn = Intrinsic::getDeclaration(M, MI->getIntrinsicID(), Tys, 3); 1467 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1468 } else { 1469 assert(isa<MemSetInst>(MI)); 1470 Value *Ops[] = { 1471 EltPtr, MI->getArgOperand(1), // Dest, Value, 1472 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size 1473 Zero, // Align 1474 ConstantInt::get(Type::getInt1Ty(MI->getContext()), 0) // isVolatile 1475 }; 1476 const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() }; 1477 Module *M = MI->getParent()->getParent()->getParent(); 1478 TheFn = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2); 1479 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1480 } 1481 } 1482 DeadInsts.push_back(MI); 1483} 1484 1485/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 1486/// overwrites the entire allocation. Extract out the pieces of the stored 1487/// integer and store them individually. 1488void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 1489 SmallVector<AllocaInst*, 32> &NewElts){ 1490 // Extract each element out of the integer according to its structure offset 1491 // and store the element value to the individual alloca. 1492 Value *SrcVal = SI->getOperand(0); 1493 const Type *AllocaEltTy = AI->getAllocatedType(); 1494 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1495 1496 // Handle tail padding by extending the operand 1497 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 1498 SrcVal = new ZExtInst(SrcVal, 1499 IntegerType::get(SI->getContext(), AllocaSizeBits), 1500 "", SI); 1501 1502 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 1503 << '\n'); 1504 1505 // There are two forms here: AI could be an array or struct. Both cases 1506 // have different ways to compute the element offset. 1507 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1508 const StructLayout *Layout = TD->getStructLayout(EltSTy); 1509 1510 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1511 // Get the number of bits to shift SrcVal to get the value. 1512 const Type *FieldTy = EltSTy->getElementType(i); 1513 uint64_t Shift = Layout->getElementOffsetInBits(i); 1514 1515 if (TD->isBigEndian()) 1516 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 1517 1518 Value *EltVal = SrcVal; 1519 if (Shift) { 1520 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1521 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1522 "sroa.store.elt", SI); 1523 } 1524 1525 // Truncate down to an integer of the right size. 1526 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1527 1528 // Ignore zero sized fields like {}, they obviously contain no data. 1529 if (FieldSizeBits == 0) continue; 1530 1531 if (FieldSizeBits != AllocaSizeBits) 1532 EltVal = new TruncInst(EltVal, 1533 IntegerType::get(SI->getContext(), FieldSizeBits), 1534 "", SI); 1535 Value *DestField = NewElts[i]; 1536 if (EltVal->getType() == FieldTy) { 1537 // Storing to an integer field of this size, just do it. 1538 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 1539 // Bitcast to the right element type (for fp/vector values). 1540 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 1541 } else { 1542 // Otherwise, bitcast the dest pointer (for aggregates). 1543 DestField = new BitCastInst(DestField, 1544 PointerType::getUnqual(EltVal->getType()), 1545 "", SI); 1546 } 1547 new StoreInst(EltVal, DestField, SI); 1548 } 1549 1550 } else { 1551 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1552 const Type *ArrayEltTy = ATy->getElementType(); 1553 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1554 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1555 1556 uint64_t Shift; 1557 1558 if (TD->isBigEndian()) 1559 Shift = AllocaSizeBits-ElementOffset; 1560 else 1561 Shift = 0; 1562 1563 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1564 // Ignore zero sized fields like {}, they obviously contain no data. 1565 if (ElementSizeBits == 0) continue; 1566 1567 Value *EltVal = SrcVal; 1568 if (Shift) { 1569 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1570 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1571 "sroa.store.elt", SI); 1572 } 1573 1574 // Truncate down to an integer of the right size. 1575 if (ElementSizeBits != AllocaSizeBits) 1576 EltVal = new TruncInst(EltVal, 1577 IntegerType::get(SI->getContext(), 1578 ElementSizeBits),"",SI); 1579 Value *DestField = NewElts[i]; 1580 if (EltVal->getType() == ArrayEltTy) { 1581 // Storing to an integer field of this size, just do it. 1582 } else if (ArrayEltTy->isFloatingPointTy() || 1583 ArrayEltTy->isVectorTy()) { 1584 // Bitcast to the right element type (for fp/vector values). 1585 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1586 } else { 1587 // Otherwise, bitcast the dest pointer (for aggregates). 1588 DestField = new BitCastInst(DestField, 1589 PointerType::getUnqual(EltVal->getType()), 1590 "", SI); 1591 } 1592 new StoreInst(EltVal, DestField, SI); 1593 1594 if (TD->isBigEndian()) 1595 Shift -= ElementOffset; 1596 else 1597 Shift += ElementOffset; 1598 } 1599 } 1600 1601 DeadInsts.push_back(SI); 1602} 1603 1604/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1605/// an integer. Load the individual pieces to form the aggregate value. 1606void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1607 SmallVector<AllocaInst*, 32> &NewElts) { 1608 // Extract each element out of the NewElts according to its structure offset 1609 // and form the result value. 1610 const Type *AllocaEltTy = AI->getAllocatedType(); 1611 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1612 1613 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1614 << '\n'); 1615 1616 // There are two forms here: AI could be an array or struct. Both cases 1617 // have different ways to compute the element offset. 1618 const StructLayout *Layout = 0; 1619 uint64_t ArrayEltBitOffset = 0; 1620 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1621 Layout = TD->getStructLayout(EltSTy); 1622 } else { 1623 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1624 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1625 } 1626 1627 Value *ResultVal = 1628 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1629 1630 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1631 // Load the value from the alloca. If the NewElt is an aggregate, cast 1632 // the pointer to an integer of the same size before doing the load. 1633 Value *SrcField = NewElts[i]; 1634 const Type *FieldTy = 1635 cast<PointerType>(SrcField->getType())->getElementType(); 1636 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1637 1638 // Ignore zero sized fields like {}, they obviously contain no data. 1639 if (FieldSizeBits == 0) continue; 1640 1641 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1642 FieldSizeBits); 1643 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1644 !FieldTy->isVectorTy()) 1645 SrcField = new BitCastInst(SrcField, 1646 PointerType::getUnqual(FieldIntTy), 1647 "", LI); 1648 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1649 1650 // If SrcField is a fp or vector of the right size but that isn't an 1651 // integer type, bitcast to an integer so we can shift it. 1652 if (SrcField->getType() != FieldIntTy) 1653 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1654 1655 // Zero extend the field to be the same size as the final alloca so that 1656 // we can shift and insert it. 1657 if (SrcField->getType() != ResultVal->getType()) 1658 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1659 1660 // Determine the number of bits to shift SrcField. 1661 uint64_t Shift; 1662 if (Layout) // Struct case. 1663 Shift = Layout->getElementOffsetInBits(i); 1664 else // Array case. 1665 Shift = i*ArrayEltBitOffset; 1666 1667 if (TD->isBigEndian()) 1668 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1669 1670 if (Shift) { 1671 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1672 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1673 } 1674 1675 // Don't create an 'or x, 0' on the first iteration. 1676 if (!isa<Constant>(ResultVal) || 1677 !cast<Constant>(ResultVal)->isNullValue()) 1678 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1679 else 1680 ResultVal = SrcField; 1681 } 1682 1683 // Handle tail padding by truncating the result 1684 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1685 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1686 1687 LI->replaceAllUsesWith(ResultVal); 1688 DeadInsts.push_back(LI); 1689} 1690 1691/// HasPadding - Return true if the specified type has any structure or 1692/// alignment padding, false otherwise. 1693static bool HasPadding(const Type *Ty, const TargetData &TD) { 1694 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 1695 return HasPadding(ATy->getElementType(), TD); 1696 1697 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) 1698 return HasPadding(VTy->getElementType(), TD); 1699 1700 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1701 const StructLayout *SL = TD.getStructLayout(STy); 1702 unsigned PrevFieldBitOffset = 0; 1703 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1704 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1705 1706 // Padding in sub-elements? 1707 if (HasPadding(STy->getElementType(i), TD)) 1708 return true; 1709 1710 // Check to see if there is any padding between this element and the 1711 // previous one. 1712 if (i) { 1713 unsigned PrevFieldEnd = 1714 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1715 if (PrevFieldEnd < FieldBitOffset) 1716 return true; 1717 } 1718 1719 PrevFieldBitOffset = FieldBitOffset; 1720 } 1721 1722 // Check for tail padding. 1723 if (unsigned EltCount = STy->getNumElements()) { 1724 unsigned PrevFieldEnd = PrevFieldBitOffset + 1725 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1726 if (PrevFieldEnd < SL->getSizeInBits()) 1727 return true; 1728 } 1729 } 1730 1731 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1732} 1733 1734/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1735/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1736/// or 1 if safe after canonicalization has been performed. 1737bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1738 // Loop over the use list of the alloca. We can only transform it if all of 1739 // the users are safe to transform. 1740 AllocaInfo Info; 1741 1742 isSafeForScalarRepl(AI, AI, 0, Info); 1743 if (Info.isUnsafe) { 1744 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1745 return false; 1746 } 1747 1748 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1749 // source and destination, we have to be careful. In particular, the memcpy 1750 // could be moving around elements that live in structure padding of the LLVM 1751 // types, but may actually be used. In these cases, we refuse to promote the 1752 // struct. 1753 if (Info.isMemCpySrc && Info.isMemCpyDst && 1754 HasPadding(AI->getAllocatedType(), *TD)) 1755 return false; 1756 1757 return true; 1758} 1759 1760 1761 1762/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1763/// some part of a constant global variable. This intentionally only accepts 1764/// constant expressions because we don't can't rewrite arbitrary instructions. 1765static bool PointsToConstantGlobal(Value *V) { 1766 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1767 return GV->isConstant(); 1768 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1769 if (CE->getOpcode() == Instruction::BitCast || 1770 CE->getOpcode() == Instruction::GetElementPtr) 1771 return PointsToConstantGlobal(CE->getOperand(0)); 1772 return false; 1773} 1774 1775/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1776/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1777/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1778/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1779/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1780/// the alloca, and if the source pointer is a pointer to a constant global, we 1781/// can optimize this. 1782static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 1783 bool isOffset) { 1784 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1785 User *U = cast<Instruction>(*UI); 1786 1787 if (LoadInst *LI = dyn_cast<LoadInst>(U)) 1788 // Ignore non-volatile loads, they are always ok. 1789 if (!LI->isVolatile()) 1790 continue; 1791 1792 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1793 // If uses of the bitcast are ok, we are ok. 1794 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1795 return false; 1796 continue; 1797 } 1798 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 1799 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1800 // doesn't, it does. 1801 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1802 isOffset || !GEP->hasAllZeroIndices())) 1803 return false; 1804 continue; 1805 } 1806 1807 // If this is isn't our memcpy/memmove, reject it as something we can't 1808 // handle. 1809 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 1810 if (MI == 0) 1811 return false; 1812 1813 // If we already have seen a copy, reject the second one. 1814 if (TheCopy) return false; 1815 1816 // If the pointer has been offset from the start of the alloca, we can't 1817 // safely handle this. 1818 if (isOffset) return false; 1819 1820 // If the memintrinsic isn't using the alloca as the dest, reject it. 1821 if (UI.getOperandNo() != 0) return false; 1822 1823 // If the source of the memcpy/move is not a constant global, reject it. 1824 if (!PointsToConstantGlobal(MI->getSource())) 1825 return false; 1826 1827 // Otherwise, the transform is safe. Remember the copy instruction. 1828 TheCopy = MI; 1829 } 1830 return true; 1831} 1832 1833/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1834/// modified by a copy from a constant global. If we can prove this, we can 1835/// replace any uses of the alloca with uses of the global directly. 1836MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1837 MemTransferInst *TheCopy = 0; 1838 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1839 return TheCopy; 1840 return 0; 1841} 1842