ScalarReplAggregates.cpp revision 2ab36d350293c77fc8941ce1023e4899df7e3a82
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_BEGIN(SROA, "scalarrepl", 139 "Scalar Replacement of Aggregates", false, false) 140INITIALIZE_PASS_DEPENDENCY(DominatorTree) 141INITIALIZE_PASS_DEPENDENCY(DominanceFrontier) 142INITIALIZE_PASS_END(SROA, "scalarrepl", 143 "Scalar Replacement of Aggregates", false, false) 144 145// Public interface to the ScalarReplAggregates pass 146FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 147 return new SROA(Threshold); 148} 149 150 151//===----------------------------------------------------------------------===// 152// Convert To Scalar Optimization. 153//===----------------------------------------------------------------------===// 154 155namespace { 156/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 157/// optimization, which scans the uses of an alloca and determines if it can 158/// rewrite it in terms of a single new alloca that can be mem2reg'd. 159class ConvertToScalarInfo { 160 /// AllocaSize - The size of the alloca being considered. 161 unsigned AllocaSize; 162 const TargetData &TD; 163 164 /// IsNotTrivial - This is set to true if there is some access to the object 165 /// which means that mem2reg can't promote it. 166 bool IsNotTrivial; 167 168 /// VectorTy - This tracks the type that we should promote the vector to if 169 /// it is possible to turn it into a vector. This starts out null, and if it 170 /// isn't possible to turn into a vector type, it gets set to VoidTy. 171 const Type *VectorTy; 172 173 /// HadAVector - True if there is at least one vector access to the alloca. 174 /// We don't want to turn random arrays into vectors and use vector element 175 /// insert/extract, but if there are element accesses to something that is 176 /// also declared as a vector, we do want to promote to a vector. 177 bool HadAVector; 178 179public: 180 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 181 : AllocaSize(Size), TD(td) { 182 IsNotTrivial = false; 183 VectorTy = 0; 184 HadAVector = false; 185 } 186 187 AllocaInst *TryConvert(AllocaInst *AI); 188 189private: 190 bool CanConvertToScalar(Value *V, uint64_t Offset); 191 void MergeInType(const Type *In, uint64_t Offset); 192 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 193 194 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 195 uint64_t Offset, IRBuilder<> &Builder); 196 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 197 uint64_t Offset, IRBuilder<> &Builder); 198}; 199} // end anonymous namespace. 200 201 202/// IsVerbotenVectorType - Return true if this is a vector type ScalarRepl isn't 203/// allowed to form. We do this to avoid MMX types, which is a complete hack, 204/// but is required until the backend is fixed. 205static bool IsVerbotenVectorType(const VectorType *VTy, const Instruction *I) { 206 StringRef Triple(I->getParent()->getParent()->getParent()->getTargetTriple()); 207 if (!Triple.startswith("i386") && 208 !Triple.startswith("x86_64")) 209 return false; 210 211 // Reject all the MMX vector types. 212 switch (VTy->getNumElements()) { 213 default: return false; 214 case 1: return VTy->getElementType()->isIntegerTy(64); 215 case 2: return VTy->getElementType()->isIntegerTy(32); 216 case 4: return VTy->getElementType()->isIntegerTy(16); 217 case 8: return VTy->getElementType()->isIntegerTy(8); 218 } 219} 220 221 222/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 223/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 224/// alloca if possible or null if not. 225AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 226 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 227 // out. 228 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 229 return 0; 230 231 // If we were able to find a vector type that can handle this with 232 // insert/extract elements, and if there was at least one use that had 233 // a vector type, promote this to a vector. We don't want to promote 234 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 235 // we just get a lot of insert/extracts. If at least one vector is 236 // involved, then we probably really do have a union of vector/array. 237 const Type *NewTy; 238 if (VectorTy && VectorTy->isVectorTy() && HadAVector && 239 !IsVerbotenVectorType(cast<VectorType>(VectorTy), AI)) { 240 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 241 << *VectorTy << '\n'); 242 NewTy = VectorTy; // Use the vector type. 243 } else { 244 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 245 // Create and insert the integer alloca. 246 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 247 } 248 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 249 ConvertUsesToScalar(AI, NewAI, 0); 250 return NewAI; 251} 252 253/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy) 254/// so far at the offset specified by Offset (which is specified in bytes). 255/// 256/// There are two cases we handle here: 257/// 1) A union of vector types of the same size and potentially its elements. 258/// Here we turn element accesses into insert/extract element operations. 259/// This promotes a <4 x float> with a store of float to the third element 260/// into a <4 x float> that uses insert element. 261/// 2) A fully general blob of memory, which we turn into some (potentially 262/// large) integer type with extract and insert operations where the loads 263/// and stores would mutate the memory. We mark this by setting VectorTy 264/// to VoidTy. 265void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) { 266 // If we already decided to turn this into a blob of integer memory, there is 267 // nothing to be done. 268 if (VectorTy && VectorTy->isVoidTy()) 269 return; 270 271 // If this could be contributing to a vector, analyze it. 272 273 // If the In type is a vector that is the same size as the alloca, see if it 274 // matches the existing VecTy. 275 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 276 // Remember if we saw a vector type. 277 HadAVector = true; 278 279 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 280 // If we're storing/loading a vector of the right size, allow it as a 281 // vector. If this the first vector we see, remember the type so that 282 // we know the element size. If this is a subsequent access, ignore it 283 // even if it is a differing type but the same size. Worst case we can 284 // bitcast the resultant vectors. 285 if (VectorTy == 0) 286 VectorTy = VInTy; 287 return; 288 } 289 } else if (In->isFloatTy() || In->isDoubleTy() || 290 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 291 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 292 // If we're accessing something that could be an element of a vector, see 293 // if the implied vector agrees with what we already have and if Offset is 294 // compatible with it. 295 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 296 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 297 (VectorTy == 0 || 298 cast<VectorType>(VectorTy)->getElementType() 299 ->getPrimitiveSizeInBits()/8 == EltSize)) { 300 if (VectorTy == 0) 301 VectorTy = VectorType::get(In, AllocaSize/EltSize); 302 return; 303 } 304 } 305 306 // Otherwise, we have a case that we can't handle with an optimized vector 307 // form. We can still turn this into a large integer. 308 VectorTy = Type::getVoidTy(In->getContext()); 309} 310 311/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 312/// its accesses to a single vector type, return true and set VecTy to 313/// the new type. If we could convert the alloca into a single promotable 314/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 315/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 316/// is the current offset from the base of the alloca being analyzed. 317/// 318/// If we see at least one access to the value that is as a vector type, set the 319/// SawVec flag. 320bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 321 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 322 Instruction *User = cast<Instruction>(*UI); 323 324 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 325 // Don't break volatile loads. 326 if (LI->isVolatile()) 327 return false; 328 // Don't touch MMX operations. 329 if (LI->getType()->isX86_MMXTy()) 330 return false; 331 MergeInType(LI->getType(), Offset); 332 continue; 333 } 334 335 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 336 // Storing the pointer, not into the value? 337 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 338 // Don't touch MMX operations. 339 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 340 return false; 341 MergeInType(SI->getOperand(0)->getType(), Offset); 342 continue; 343 } 344 345 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 346 IsNotTrivial = true; // Can't be mem2reg'd. 347 if (!CanConvertToScalar(BCI, Offset)) 348 return false; 349 continue; 350 } 351 352 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 353 // If this is a GEP with a variable indices, we can't handle it. 354 if (!GEP->hasAllConstantIndices()) 355 return false; 356 357 // Compute the offset that this GEP adds to the pointer. 358 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 359 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 360 &Indices[0], Indices.size()); 361 // See if all uses can be converted. 362 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 363 return false; 364 IsNotTrivial = true; // Can't be mem2reg'd. 365 continue; 366 } 367 368 // If this is a constant sized memset of a constant value (e.g. 0) we can 369 // handle it. 370 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 371 // Store of constant value and constant size. 372 if (!isa<ConstantInt>(MSI->getValue()) || 373 !isa<ConstantInt>(MSI->getLength())) 374 return false; 375 IsNotTrivial = true; // Can't be mem2reg'd. 376 continue; 377 } 378 379 // If this is a memcpy or memmove into or out of the whole allocation, we 380 // can handle it like a load or store of the scalar type. 381 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 382 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 383 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 384 return false; 385 386 IsNotTrivial = true; // Can't be mem2reg'd. 387 continue; 388 } 389 390 // Otherwise, we cannot handle this! 391 return false; 392 } 393 394 return true; 395} 396 397/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 398/// directly. This happens when we are converting an "integer union" to a 399/// single integer scalar, or when we are converting a "vector union" to a 400/// vector with insert/extractelement instructions. 401/// 402/// Offset is an offset from the original alloca, in bits that need to be 403/// shifted to the right. By the end of this, there should be no uses of Ptr. 404void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 405 uint64_t Offset) { 406 while (!Ptr->use_empty()) { 407 Instruction *User = cast<Instruction>(Ptr->use_back()); 408 409 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 410 ConvertUsesToScalar(CI, NewAI, Offset); 411 CI->eraseFromParent(); 412 continue; 413 } 414 415 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 416 // Compute the offset that this GEP adds to the pointer. 417 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 418 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 419 &Indices[0], Indices.size()); 420 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 421 GEP->eraseFromParent(); 422 continue; 423 } 424 425 IRBuilder<> Builder(User->getParent(), User); 426 427 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 428 // The load is a bit extract from NewAI shifted right by Offset bits. 429 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 430 Value *NewLoadVal 431 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 432 LI->replaceAllUsesWith(NewLoadVal); 433 LI->eraseFromParent(); 434 continue; 435 } 436 437 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 438 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 439 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 440 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 441 Builder); 442 Builder.CreateStore(New, NewAI); 443 SI->eraseFromParent(); 444 445 // If the load we just inserted is now dead, then the inserted store 446 // overwrote the entire thing. 447 if (Old->use_empty()) 448 Old->eraseFromParent(); 449 continue; 450 } 451 452 // If this is a constant sized memset of a constant value (e.g. 0) we can 453 // transform it into a store of the expanded constant value. 454 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 455 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 456 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 457 if (NumBytes != 0) { 458 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 459 460 // Compute the value replicated the right number of times. 461 APInt APVal(NumBytes*8, Val); 462 463 // Splat the value if non-zero. 464 if (Val) 465 for (unsigned i = 1; i != NumBytes; ++i) 466 APVal |= APVal << 8; 467 468 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 469 Value *New = ConvertScalar_InsertValue( 470 ConstantInt::get(User->getContext(), APVal), 471 Old, Offset, Builder); 472 Builder.CreateStore(New, NewAI); 473 474 // If the load we just inserted is now dead, then the memset overwrote 475 // the entire thing. 476 if (Old->use_empty()) 477 Old->eraseFromParent(); 478 } 479 MSI->eraseFromParent(); 480 continue; 481 } 482 483 // If this is a memcpy or memmove into or out of the whole allocation, we 484 // can handle it like a load or store of the scalar type. 485 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 486 assert(Offset == 0 && "must be store to start of alloca"); 487 488 // If the source and destination are both to the same alloca, then this is 489 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 490 // as appropriate. 491 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject(0)); 492 493 if (MTI->getSource()->getUnderlyingObject(0) != OrigAI) { 494 // Dest must be OrigAI, change this to be a load from the original 495 // pointer (bitcasted), then a store to our new alloca. 496 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 497 Value *SrcPtr = MTI->getSource(); 498 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 499 500 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 501 SrcVal->setAlignment(MTI->getAlignment()); 502 Builder.CreateStore(SrcVal, NewAI); 503 } else if (MTI->getDest()->getUnderlyingObject(0) != OrigAI) { 504 // Src must be OrigAI, change this to be a load from NewAI then a store 505 // through the original dest pointer (bitcasted). 506 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 507 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 508 509 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 510 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 511 NewStore->setAlignment(MTI->getAlignment()); 512 } else { 513 // Noop transfer. Src == Dst 514 } 515 516 MTI->eraseFromParent(); 517 continue; 518 } 519 520 llvm_unreachable("Unsupported operation!"); 521 } 522} 523 524/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 525/// or vector value FromVal, extracting the bits from the offset specified by 526/// Offset. This returns the value, which is of type ToType. 527/// 528/// This happens when we are converting an "integer union" to a single 529/// integer scalar, or when we are converting a "vector union" to a vector with 530/// insert/extractelement instructions. 531/// 532/// Offset is an offset from the original alloca, in bits that need to be 533/// shifted to the right. 534Value *ConvertToScalarInfo:: 535ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 536 uint64_t Offset, IRBuilder<> &Builder) { 537 // If the load is of the whole new alloca, no conversion is needed. 538 if (FromVal->getType() == ToType && Offset == 0) 539 return FromVal; 540 541 // If the result alloca is a vector type, this is either an element 542 // access or a bitcast to another vector type of the same size. 543 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 544 if (ToType->isVectorTy()) 545 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 546 547 // Otherwise it must be an element access. 548 unsigned Elt = 0; 549 if (Offset) { 550 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 551 Elt = Offset/EltSize; 552 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 553 } 554 // Return the element extracted out of it. 555 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 556 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 557 if (V->getType() != ToType) 558 V = Builder.CreateBitCast(V, ToType, "tmp"); 559 return V; 560 } 561 562 // If ToType is a first class aggregate, extract out each of the pieces and 563 // use insertvalue's to form the FCA. 564 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 565 const StructLayout &Layout = *TD.getStructLayout(ST); 566 Value *Res = UndefValue::get(ST); 567 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 568 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 569 Offset+Layout.getElementOffsetInBits(i), 570 Builder); 571 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 572 } 573 return Res; 574 } 575 576 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 577 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 578 Value *Res = UndefValue::get(AT); 579 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 580 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 581 Offset+i*EltSize, Builder); 582 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 583 } 584 return Res; 585 } 586 587 // Otherwise, this must be a union that was converted to an integer value. 588 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 589 590 // If this is a big-endian system and the load is narrower than the 591 // full alloca type, we need to do a shift to get the right bits. 592 int ShAmt = 0; 593 if (TD.isBigEndian()) { 594 // On big-endian machines, the lowest bit is stored at the bit offset 595 // from the pointer given by getTypeStoreSizeInBits. This matters for 596 // integers with a bitwidth that is not a multiple of 8. 597 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 598 TD.getTypeStoreSizeInBits(ToType) - Offset; 599 } else { 600 ShAmt = Offset; 601 } 602 603 // Note: we support negative bitwidths (with shl) which are not defined. 604 // We do this to support (f.e.) loads off the end of a structure where 605 // only some bits are used. 606 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 607 FromVal = Builder.CreateLShr(FromVal, 608 ConstantInt::get(FromVal->getType(), 609 ShAmt), "tmp"); 610 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 611 FromVal = Builder.CreateShl(FromVal, 612 ConstantInt::get(FromVal->getType(), 613 -ShAmt), "tmp"); 614 615 // Finally, unconditionally truncate the integer to the right width. 616 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 617 if (LIBitWidth < NTy->getBitWidth()) 618 FromVal = 619 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 620 LIBitWidth), "tmp"); 621 else if (LIBitWidth > NTy->getBitWidth()) 622 FromVal = 623 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 624 LIBitWidth), "tmp"); 625 626 // If the result is an integer, this is a trunc or bitcast. 627 if (ToType->isIntegerTy()) { 628 // Should be done. 629 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 630 // Just do a bitcast, we know the sizes match up. 631 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 632 } else { 633 // Otherwise must be a pointer. 634 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 635 } 636 assert(FromVal->getType() == ToType && "Didn't convert right?"); 637 return FromVal; 638} 639 640/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 641/// or vector value "Old" at the offset specified by Offset. 642/// 643/// This happens when we are converting an "integer union" to a 644/// single integer scalar, or when we are converting a "vector union" to a 645/// vector with insert/extractelement instructions. 646/// 647/// Offset is an offset from the original alloca, in bits that need to be 648/// shifted to the right. 649Value *ConvertToScalarInfo:: 650ConvertScalar_InsertValue(Value *SV, Value *Old, 651 uint64_t Offset, IRBuilder<> &Builder) { 652 // Convert the stored type to the actual type, shift it left to insert 653 // then 'or' into place. 654 const Type *AllocaType = Old->getType(); 655 LLVMContext &Context = Old->getContext(); 656 657 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 658 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 659 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 660 661 // Changing the whole vector with memset or with an access of a different 662 // vector type? 663 if (ValSize == VecSize) 664 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 665 666 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 667 668 // Must be an element insertion. 669 unsigned Elt = Offset/EltSize; 670 671 if (SV->getType() != VTy->getElementType()) 672 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 673 674 SV = Builder.CreateInsertElement(Old, SV, 675 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 676 "tmp"); 677 return SV; 678 } 679 680 // If SV is a first-class aggregate value, insert each value recursively. 681 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 682 const StructLayout &Layout = *TD.getStructLayout(ST); 683 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 684 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 685 Old = ConvertScalar_InsertValue(Elt, Old, 686 Offset+Layout.getElementOffsetInBits(i), 687 Builder); 688 } 689 return Old; 690 } 691 692 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 693 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 694 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 695 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 696 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 697 } 698 return Old; 699 } 700 701 // If SV is a float, convert it to the appropriate integer type. 702 // If it is a pointer, do the same. 703 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 704 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 705 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 706 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 707 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 708 SV = Builder.CreateBitCast(SV, 709 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 710 else if (SV->getType()->isPointerTy()) 711 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 712 713 // Zero extend or truncate the value if needed. 714 if (SV->getType() != AllocaType) { 715 if (SV->getType()->getPrimitiveSizeInBits() < 716 AllocaType->getPrimitiveSizeInBits()) 717 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 718 else { 719 // Truncation may be needed if storing more than the alloca can hold 720 // (undefined behavior). 721 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 722 SrcWidth = DestWidth; 723 SrcStoreWidth = DestStoreWidth; 724 } 725 } 726 727 // If this is a big-endian system and the store is narrower than the 728 // full alloca type, we need to do a shift to get the right bits. 729 int ShAmt = 0; 730 if (TD.isBigEndian()) { 731 // On big-endian machines, the lowest bit is stored at the bit offset 732 // from the pointer given by getTypeStoreSizeInBits. This matters for 733 // integers with a bitwidth that is not a multiple of 8. 734 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 735 } else { 736 ShAmt = Offset; 737 } 738 739 // Note: we support negative bitwidths (with shr) which are not defined. 740 // We do this to support (f.e.) stores off the end of a structure where 741 // only some bits in the structure are set. 742 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 743 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 744 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 745 ShAmt), "tmp"); 746 Mask <<= ShAmt; 747 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 748 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 749 -ShAmt), "tmp"); 750 Mask = Mask.lshr(-ShAmt); 751 } 752 753 // Mask out the bits we are about to insert from the old value, and or 754 // in the new bits. 755 if (SrcWidth != DestWidth) { 756 assert(DestWidth > SrcWidth); 757 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 758 SV = Builder.CreateOr(Old, SV, "ins"); 759 } 760 return SV; 761} 762 763 764//===----------------------------------------------------------------------===// 765// SRoA Driver 766//===----------------------------------------------------------------------===// 767 768 769bool SROA::runOnFunction(Function &F) { 770 TD = getAnalysisIfAvailable<TargetData>(); 771 772 bool Changed = performPromotion(F); 773 774 // FIXME: ScalarRepl currently depends on TargetData more than it 775 // theoretically needs to. It should be refactored in order to support 776 // target-independent IR. Until this is done, just skip the actual 777 // scalar-replacement portion of this pass. 778 if (!TD) return Changed; 779 780 while (1) { 781 bool LocalChange = performScalarRepl(F); 782 if (!LocalChange) break; // No need to repromote if no scalarrepl 783 Changed = true; 784 LocalChange = performPromotion(F); 785 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 786 } 787 788 return Changed; 789} 790 791 792bool SROA::performPromotion(Function &F) { 793 std::vector<AllocaInst*> Allocas; 794 DominatorTree &DT = getAnalysis<DominatorTree>(); 795 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 796 797 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 798 799 bool Changed = false; 800 801 while (1) { 802 Allocas.clear(); 803 804 // Find allocas that are safe to promote, by looking at all instructions in 805 // the entry node 806 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 807 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 808 if (isAllocaPromotable(AI)) 809 Allocas.push_back(AI); 810 811 if (Allocas.empty()) break; 812 813 PromoteMemToReg(Allocas, DT, DF); 814 NumPromoted += Allocas.size(); 815 Changed = true; 816 } 817 818 return Changed; 819} 820 821 822/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 823/// SROA. It must be a struct or array type with a small number of elements. 824static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 825 const Type *T = AI->getAllocatedType(); 826 // Do not promote any struct into more than 32 separate vars. 827 if (const StructType *ST = dyn_cast<StructType>(T)) 828 return ST->getNumElements() <= 32; 829 // Arrays are much less likely to be safe for SROA; only consider 830 // them if they are very small. 831 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 832 return AT->getNumElements() <= 8; 833 return false; 834} 835 836 837// performScalarRepl - This algorithm is a simple worklist driven algorithm, 838// which runs on all of the malloc/alloca instructions in the function, removing 839// them if they are only used by getelementptr instructions. 840// 841bool SROA::performScalarRepl(Function &F) { 842 std::vector<AllocaInst*> WorkList; 843 844 // Scan the entry basic block, adding allocas to the worklist. 845 BasicBlock &BB = F.getEntryBlock(); 846 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 847 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 848 WorkList.push_back(A); 849 850 // Process the worklist 851 bool Changed = false; 852 while (!WorkList.empty()) { 853 AllocaInst *AI = WorkList.back(); 854 WorkList.pop_back(); 855 856 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 857 // with unused elements. 858 if (AI->use_empty()) { 859 AI->eraseFromParent(); 860 Changed = true; 861 continue; 862 } 863 864 // If this alloca is impossible for us to promote, reject it early. 865 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 866 continue; 867 868 // Check to see if this allocation is only modified by a memcpy/memmove from 869 // a constant global. If this is the case, we can change all users to use 870 // the constant global instead. This is commonly produced by the CFE by 871 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 872 // is only subsequently read. 873 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 874 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 875 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 876 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 877 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 878 TheCopy->eraseFromParent(); // Don't mutate the global. 879 AI->eraseFromParent(); 880 ++NumGlobals; 881 Changed = true; 882 continue; 883 } 884 885 // Check to see if we can perform the core SROA transformation. We cannot 886 // transform the allocation instruction if it is an array allocation 887 // (allocations OF arrays are ok though), and an allocation of a scalar 888 // value cannot be decomposed at all. 889 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 890 891 // Do not promote [0 x %struct]. 892 if (AllocaSize == 0) continue; 893 894 // Do not promote any struct whose size is too big. 895 if (AllocaSize > SRThreshold) continue; 896 897 // If the alloca looks like a good candidate for scalar replacement, and if 898 // all its users can be transformed, then split up the aggregate into its 899 // separate elements. 900 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 901 DoScalarReplacement(AI, WorkList); 902 Changed = true; 903 continue; 904 } 905 906 // If we can turn this aggregate value (potentially with casts) into a 907 // simple scalar value that can be mem2reg'd into a register value. 908 // IsNotTrivial tracks whether this is something that mem2reg could have 909 // promoted itself. If so, we don't want to transform it needlessly. Note 910 // that we can't just check based on the type: the alloca may be of an i32 911 // but that has pointer arithmetic to set byte 3 of it or something. 912 if (AllocaInst *NewAI = 913 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 914 NewAI->takeName(AI); 915 AI->eraseFromParent(); 916 ++NumConverted; 917 Changed = true; 918 continue; 919 } 920 921 // Otherwise, couldn't process this alloca. 922 } 923 924 return Changed; 925} 926 927/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 928/// predicate, do SROA now. 929void SROA::DoScalarReplacement(AllocaInst *AI, 930 std::vector<AllocaInst*> &WorkList) { 931 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 932 SmallVector<AllocaInst*, 32> ElementAllocas; 933 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 934 ElementAllocas.reserve(ST->getNumContainedTypes()); 935 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 936 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 937 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 } else { 943 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 944 ElementAllocas.reserve(AT->getNumElements()); 945 const Type *ElTy = AT->getElementType(); 946 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 947 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 948 AI->getName() + "." + Twine(i), AI); 949 ElementAllocas.push_back(NA); 950 WorkList.push_back(NA); // Add to worklist for recursive processing 951 } 952 } 953 954 // Now that we have created the new alloca instructions, rewrite all the 955 // uses of the old alloca. 956 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 957 958 // Now erase any instructions that were made dead while rewriting the alloca. 959 DeleteDeadInstructions(); 960 AI->eraseFromParent(); 961 962 ++NumReplaced; 963} 964 965/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 966/// recursively including all their operands that become trivially dead. 967void SROA::DeleteDeadInstructions() { 968 while (!DeadInsts.empty()) { 969 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 970 971 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 972 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 973 // Zero out the operand and see if it becomes trivially dead. 974 // (But, don't add allocas to the dead instruction list -- they are 975 // already on the worklist and will be deleted separately.) 976 *OI = 0; 977 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 978 DeadInsts.push_back(U); 979 } 980 981 I->eraseFromParent(); 982 } 983} 984 985/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 986/// performing scalar replacement of alloca AI. The results are flagged in 987/// the Info parameter. Offset indicates the position within AI that is 988/// referenced by this instruction. 989void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 990 AllocaInfo &Info) { 991 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 992 Instruction *User = cast<Instruction>(*UI); 993 994 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 995 isSafeForScalarRepl(BC, AI, Offset, Info); 996 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 997 uint64_t GEPOffset = Offset; 998 isSafeGEP(GEPI, AI, GEPOffset, Info); 999 if (!Info.isUnsafe) 1000 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 1001 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1002 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1003 if (Length) 1004 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 1005 UI.getOperandNo() == 0, Info); 1006 else 1007 MarkUnsafe(Info); 1008 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1009 if (!LI->isVolatile()) { 1010 const Type *LIType = LI->getType(); 1011 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 1012 LIType, false, Info); 1013 } else 1014 MarkUnsafe(Info); 1015 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1016 // Store is ok if storing INTO the pointer, not storing the pointer 1017 if (!SI->isVolatile() && SI->getOperand(0) != I) { 1018 const Type *SIType = SI->getOperand(0)->getType(); 1019 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 1020 SIType, true, Info); 1021 } else 1022 MarkUnsafe(Info); 1023 } else { 1024 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 1025 MarkUnsafe(Info); 1026 } 1027 if (Info.isUnsafe) return; 1028 } 1029} 1030 1031/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1032/// replacement. It is safe when all the indices are constant, in-bounds 1033/// references, and when the resulting offset corresponds to an element within 1034/// the alloca type. The results are flagged in the Info parameter. Upon 1035/// return, Offset is adjusted as specified by the GEP indices. 1036void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 1037 uint64_t &Offset, AllocaInfo &Info) { 1038 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1039 if (GEPIt == E) 1040 return; 1041 1042 // Walk through the GEP type indices, checking the types that this indexes 1043 // into. 1044 for (; GEPIt != E; ++GEPIt) { 1045 // Ignore struct elements, no extra checking needed for these. 1046 if ((*GEPIt)->isStructTy()) 1047 continue; 1048 1049 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1050 if (!IdxVal) 1051 return MarkUnsafe(Info); 1052 } 1053 1054 // Compute the offset due to this GEP and check if the alloca has a 1055 // component element at that offset. 1056 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1057 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1058 &Indices[0], Indices.size()); 1059 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 1060 MarkUnsafe(Info); 1061} 1062 1063/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1064/// alloca or has an offset and size that corresponds to a component element 1065/// within it. The offset checked here may have been formed from a GEP with a 1066/// pointer bitcasted to a different type. 1067void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 1068 const Type *MemOpType, bool isStore, 1069 AllocaInfo &Info) { 1070 // Check if this is a load/store of the entire alloca. 1071 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 1072 bool UsesAggregateType = (MemOpType == AI->getAllocatedType()); 1073 // This is safe for MemIntrinsics (where MemOpType is 0), integer types 1074 // (which are essentially the same as the MemIntrinsics, especially with 1075 // regard to copying padding between elements), or references using the 1076 // aggregate type of the alloca. 1077 if (!MemOpType || MemOpType->isIntegerTy() || UsesAggregateType) { 1078 if (!UsesAggregateType) { 1079 if (isStore) 1080 Info.isMemCpyDst = true; 1081 else 1082 Info.isMemCpySrc = true; 1083 } 1084 return; 1085 } 1086 } 1087 // Check if the offset/size correspond to a component within the alloca type. 1088 const Type *T = AI->getAllocatedType(); 1089 if (TypeHasComponent(T, Offset, MemSize)) 1090 return; 1091 1092 return MarkUnsafe(Info); 1093} 1094 1095/// TypeHasComponent - Return true if T has a component type with the 1096/// specified offset and size. If Size is zero, do not check the size. 1097bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1098 const Type *EltTy; 1099 uint64_t EltSize; 1100 if (const StructType *ST = dyn_cast<StructType>(T)) { 1101 const StructLayout *Layout = TD->getStructLayout(ST); 1102 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1103 EltTy = ST->getContainedType(EltIdx); 1104 EltSize = TD->getTypeAllocSize(EltTy); 1105 Offset -= Layout->getElementOffset(EltIdx); 1106 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1107 EltTy = AT->getElementType(); 1108 EltSize = TD->getTypeAllocSize(EltTy); 1109 if (Offset >= AT->getNumElements() * EltSize) 1110 return false; 1111 Offset %= EltSize; 1112 } else { 1113 return false; 1114 } 1115 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1116 return true; 1117 // Check if the component spans multiple elements. 1118 if (Offset + Size > EltSize) 1119 return false; 1120 return TypeHasComponent(EltTy, Offset, Size); 1121} 1122 1123/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1124/// the instruction I, which references it, to use the separate elements. 1125/// Offset indicates the position within AI that is referenced by this 1126/// instruction. 1127void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1128 SmallVector<AllocaInst*, 32> &NewElts) { 1129 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1130 Instruction *User = cast<Instruction>(*UI); 1131 1132 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1133 RewriteBitCast(BC, AI, Offset, NewElts); 1134 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1135 RewriteGEP(GEPI, AI, Offset, NewElts); 1136 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1137 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1138 uint64_t MemSize = Length->getZExtValue(); 1139 if (Offset == 0 && 1140 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1141 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1142 // Otherwise the intrinsic can only touch a single element and the 1143 // address operand will be updated, so nothing else needs to be done. 1144 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1145 const Type *LIType = LI->getType(); 1146 if (LIType == AI->getAllocatedType()) { 1147 // Replace: 1148 // %res = load { i32, i32 }* %alloc 1149 // with: 1150 // %load.0 = load i32* %alloc.0 1151 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1152 // %load.1 = load i32* %alloc.1 1153 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1154 // (Also works for arrays instead of structs) 1155 Value *Insert = UndefValue::get(LIType); 1156 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1157 Value *Load = new LoadInst(NewElts[i], "load", LI); 1158 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1159 } 1160 LI->replaceAllUsesWith(Insert); 1161 DeadInsts.push_back(LI); 1162 } else if (LIType->isIntegerTy() && 1163 TD->getTypeAllocSize(LIType) == 1164 TD->getTypeAllocSize(AI->getAllocatedType())) { 1165 // If this is a load of the entire alloca to an integer, rewrite it. 1166 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1167 } 1168 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1169 Value *Val = SI->getOperand(0); 1170 const Type *SIType = Val->getType(); 1171 if (SIType == AI->getAllocatedType()) { 1172 // Replace: 1173 // store { i32, i32 } %val, { i32, i32 }* %alloc 1174 // with: 1175 // %val.0 = extractvalue { i32, i32 } %val, 0 1176 // store i32 %val.0, i32* %alloc.0 1177 // %val.1 = extractvalue { i32, i32 } %val, 1 1178 // store i32 %val.1, i32* %alloc.1 1179 // (Also works for arrays instead of structs) 1180 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1181 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1182 new StoreInst(Extract, NewElts[i], SI); 1183 } 1184 DeadInsts.push_back(SI); 1185 } else if (SIType->isIntegerTy() && 1186 TD->getTypeAllocSize(SIType) == 1187 TD->getTypeAllocSize(AI->getAllocatedType())) { 1188 // If this is a store of the entire alloca from an integer, rewrite it. 1189 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1190 } 1191 } 1192 } 1193} 1194 1195/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1196/// and recursively continue updating all of its uses. 1197void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1198 SmallVector<AllocaInst*, 32> &NewElts) { 1199 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1200 if (BC->getOperand(0) != AI) 1201 return; 1202 1203 // The bitcast references the original alloca. Replace its uses with 1204 // references to the first new element alloca. 1205 Instruction *Val = NewElts[0]; 1206 if (Val->getType() != BC->getDestTy()) { 1207 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1208 Val->takeName(BC); 1209 } 1210 BC->replaceAllUsesWith(Val); 1211 DeadInsts.push_back(BC); 1212} 1213 1214/// FindElementAndOffset - Return the index of the element containing Offset 1215/// within the specified type, which must be either a struct or an array. 1216/// Sets T to the type of the element and Offset to the offset within that 1217/// element. IdxTy is set to the type of the index result to be used in a 1218/// GEP instruction. 1219uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1220 const Type *&IdxTy) { 1221 uint64_t Idx = 0; 1222 if (const StructType *ST = dyn_cast<StructType>(T)) { 1223 const StructLayout *Layout = TD->getStructLayout(ST); 1224 Idx = Layout->getElementContainingOffset(Offset); 1225 T = ST->getContainedType(Idx); 1226 Offset -= Layout->getElementOffset(Idx); 1227 IdxTy = Type::getInt32Ty(T->getContext()); 1228 return Idx; 1229 } 1230 const ArrayType *AT = cast<ArrayType>(T); 1231 T = AT->getElementType(); 1232 uint64_t EltSize = TD->getTypeAllocSize(T); 1233 Idx = Offset / EltSize; 1234 Offset -= Idx * EltSize; 1235 IdxTy = Type::getInt64Ty(T->getContext()); 1236 return Idx; 1237} 1238 1239/// RewriteGEP - Check if this GEP instruction moves the pointer across 1240/// elements of the alloca that are being split apart, and if so, rewrite 1241/// the GEP to be relative to the new element. 1242void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1243 SmallVector<AllocaInst*, 32> &NewElts) { 1244 uint64_t OldOffset = Offset; 1245 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1246 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1247 &Indices[0], Indices.size()); 1248 1249 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1250 1251 const Type *T = AI->getAllocatedType(); 1252 const Type *IdxTy; 1253 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1254 if (GEPI->getOperand(0) == AI) 1255 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1256 1257 T = AI->getAllocatedType(); 1258 uint64_t EltOffset = Offset; 1259 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1260 1261 // If this GEP does not move the pointer across elements of the alloca 1262 // being split, then it does not needs to be rewritten. 1263 if (Idx == OldIdx) 1264 return; 1265 1266 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1267 SmallVector<Value*, 8> NewArgs; 1268 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1269 while (EltOffset != 0) { 1270 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1271 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1272 } 1273 Instruction *Val = NewElts[Idx]; 1274 if (NewArgs.size() > 1) { 1275 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1276 NewArgs.end(), "", GEPI); 1277 Val->takeName(GEPI); 1278 } 1279 if (Val->getType() != GEPI->getType()) 1280 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1281 GEPI->replaceAllUsesWith(Val); 1282 DeadInsts.push_back(GEPI); 1283} 1284 1285/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1286/// Rewrite it to copy or set the elements of the scalarized memory. 1287void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1288 AllocaInst *AI, 1289 SmallVector<AllocaInst*, 32> &NewElts) { 1290 // If this is a memcpy/memmove, construct the other pointer as the 1291 // appropriate type. The "Other" pointer is the pointer that goes to memory 1292 // that doesn't have anything to do with the alloca that we are promoting. For 1293 // memset, this Value* stays null. 1294 Value *OtherPtr = 0; 1295 unsigned MemAlignment = MI->getAlignment(); 1296 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1297 if (Inst == MTI->getRawDest()) 1298 OtherPtr = MTI->getRawSource(); 1299 else { 1300 assert(Inst == MTI->getRawSource()); 1301 OtherPtr = MTI->getRawDest(); 1302 } 1303 } 1304 1305 // If there is an other pointer, we want to convert it to the same pointer 1306 // type as AI has, so we can GEP through it safely. 1307 if (OtherPtr) { 1308 unsigned AddrSpace = 1309 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 1310 1311 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1312 // optimization, but it's also required to detect the corner case where 1313 // both pointer operands are referencing the same memory, and where 1314 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1315 // function is only called for mem intrinsics that access the whole 1316 // aggregate, so non-zero GEPs are not an issue here.) 1317 OtherPtr = OtherPtr->stripPointerCasts(); 1318 1319 // Copying the alloca to itself is a no-op: just delete it. 1320 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1321 // This code will run twice for a no-op memcpy -- once for each operand. 1322 // Put only one reference to MI on the DeadInsts list. 1323 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1324 E = DeadInsts.end(); I != E; ++I) 1325 if (*I == MI) return; 1326 DeadInsts.push_back(MI); 1327 return; 1328 } 1329 1330 // If the pointer is not the right type, insert a bitcast to the right 1331 // type. 1332 const Type *NewTy = 1333 PointerType::get(AI->getType()->getElementType(), AddrSpace); 1334 1335 if (OtherPtr->getType() != NewTy) 1336 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 1337 } 1338 1339 // Process each element of the aggregate. 1340 Value *TheFn = MI->getCalledValue(); 1341 const Type *BytePtrTy = MI->getRawDest()->getType(); 1342 bool SROADest = MI->getRawDest() == Inst; 1343 1344 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1345 1346 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1347 // If this is a memcpy/memmove, emit a GEP of the other element address. 1348 Value *OtherElt = 0; 1349 unsigned OtherEltAlign = MemAlignment; 1350 1351 if (OtherPtr) { 1352 Value *Idx[2] = { Zero, 1353 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1354 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1355 OtherPtr->getName()+"."+Twine(i), 1356 MI); 1357 uint64_t EltOffset; 1358 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1359 const Type *OtherTy = OtherPtrTy->getElementType(); 1360 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1361 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1362 } else { 1363 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 1364 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1365 } 1366 1367 // The alignment of the other pointer is the guaranteed alignment of the 1368 // element, which is affected by both the known alignment of the whole 1369 // mem intrinsic and the alignment of the element. If the alignment of 1370 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 1371 // known alignment is just 4 bytes. 1372 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 1373 } 1374 1375 Value *EltPtr = NewElts[i]; 1376 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 1377 1378 // If we got down to a scalar, insert a load or store as appropriate. 1379 if (EltTy->isSingleValueType()) { 1380 if (isa<MemTransferInst>(MI)) { 1381 if (SROADest) { 1382 // From Other to Alloca. 1383 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 1384 new StoreInst(Elt, EltPtr, MI); 1385 } else { 1386 // From Alloca to Other. 1387 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 1388 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 1389 } 1390 continue; 1391 } 1392 assert(isa<MemSetInst>(MI)); 1393 1394 // If the stored element is zero (common case), just store a null 1395 // constant. 1396 Constant *StoreVal; 1397 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 1398 if (CI->isZero()) { 1399 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 1400 } else { 1401 // If EltTy is a vector type, get the element type. 1402 const Type *ValTy = EltTy->getScalarType(); 1403 1404 // Construct an integer with the right value. 1405 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 1406 APInt OneVal(EltSize, CI->getZExtValue()); 1407 APInt TotalVal(OneVal); 1408 // Set each byte. 1409 for (unsigned i = 0; 8*i < EltSize; ++i) { 1410 TotalVal = TotalVal.shl(8); 1411 TotalVal |= OneVal; 1412 } 1413 1414 // Convert the integer value to the appropriate type. 1415 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 1416 if (ValTy->isPointerTy()) 1417 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 1418 else if (ValTy->isFloatingPointTy()) 1419 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 1420 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 1421 1422 // If the requested value was a vector constant, create it. 1423 if (EltTy != ValTy) { 1424 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 1425 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 1426 StoreVal = ConstantVector::get(&Elts[0], NumElts); 1427 } 1428 } 1429 new StoreInst(StoreVal, EltPtr, MI); 1430 continue; 1431 } 1432 // Otherwise, if we're storing a byte variable, use a memset call for 1433 // this element. 1434 } 1435 1436 // Cast the element pointer to BytePtrTy. 1437 if (EltPtr->getType() != BytePtrTy) 1438 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI); 1439 1440 // Cast the other pointer (if we have one) to BytePtrTy. 1441 if (OtherElt && OtherElt->getType() != BytePtrTy) { 1442 // Preserve address space of OtherElt 1443 const PointerType* OtherPTy = cast<PointerType>(OtherElt->getType()); 1444 const PointerType* PTy = cast<PointerType>(BytePtrTy); 1445 if (OtherPTy->getElementType() != PTy->getElementType()) { 1446 Type *NewOtherPTy = PointerType::get(PTy->getElementType(), 1447 OtherPTy->getAddressSpace()); 1448 OtherElt = new BitCastInst(OtherElt, NewOtherPTy, 1449 OtherElt->getNameStr(), MI); 1450 } 1451 } 1452 1453 unsigned EltSize = TD->getTypeAllocSize(EltTy); 1454 1455 // Finally, insert the meminst for this element. 1456 if (isa<MemTransferInst>(MI)) { 1457 Value *Ops[] = { 1458 SROADest ? EltPtr : OtherElt, // Dest ptr 1459 SROADest ? OtherElt : EltPtr, // Src ptr 1460 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size 1461 // Align 1462 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign), 1463 MI->getVolatileCst() 1464 }; 1465 // In case we fold the address space overloaded memcpy of A to B 1466 // with memcpy of B to C, change the function to be a memcpy of A to C. 1467 const Type *Tys[] = { Ops[0]->getType(), Ops[1]->getType(), 1468 Ops[2]->getType() }; 1469 Module *M = MI->getParent()->getParent()->getParent(); 1470 TheFn = Intrinsic::getDeclaration(M, MI->getIntrinsicID(), Tys, 3); 1471 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1472 } else { 1473 assert(isa<MemSetInst>(MI)); 1474 Value *Ops[] = { 1475 EltPtr, MI->getArgOperand(1), // Dest, Value, 1476 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size 1477 Zero, // Align 1478 ConstantInt::get(Type::getInt1Ty(MI->getContext()), 0) // isVolatile 1479 }; 1480 const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() }; 1481 Module *M = MI->getParent()->getParent()->getParent(); 1482 TheFn = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2); 1483 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1484 } 1485 } 1486 DeadInsts.push_back(MI); 1487} 1488 1489/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 1490/// overwrites the entire allocation. Extract out the pieces of the stored 1491/// integer and store them individually. 1492void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 1493 SmallVector<AllocaInst*, 32> &NewElts){ 1494 // Extract each element out of the integer according to its structure offset 1495 // and store the element value to the individual alloca. 1496 Value *SrcVal = SI->getOperand(0); 1497 const Type *AllocaEltTy = AI->getAllocatedType(); 1498 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1499 1500 // Handle tail padding by extending the operand 1501 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 1502 SrcVal = new ZExtInst(SrcVal, 1503 IntegerType::get(SI->getContext(), AllocaSizeBits), 1504 "", SI); 1505 1506 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 1507 << '\n'); 1508 1509 // There are two forms here: AI could be an array or struct. Both cases 1510 // have different ways to compute the element offset. 1511 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1512 const StructLayout *Layout = TD->getStructLayout(EltSTy); 1513 1514 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1515 // Get the number of bits to shift SrcVal to get the value. 1516 const Type *FieldTy = EltSTy->getElementType(i); 1517 uint64_t Shift = Layout->getElementOffsetInBits(i); 1518 1519 if (TD->isBigEndian()) 1520 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 1521 1522 Value *EltVal = SrcVal; 1523 if (Shift) { 1524 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1525 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1526 "sroa.store.elt", SI); 1527 } 1528 1529 // Truncate down to an integer of the right size. 1530 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1531 1532 // Ignore zero sized fields like {}, they obviously contain no data. 1533 if (FieldSizeBits == 0) continue; 1534 1535 if (FieldSizeBits != AllocaSizeBits) 1536 EltVal = new TruncInst(EltVal, 1537 IntegerType::get(SI->getContext(), FieldSizeBits), 1538 "", SI); 1539 Value *DestField = NewElts[i]; 1540 if (EltVal->getType() == FieldTy) { 1541 // Storing to an integer field of this size, just do it. 1542 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 1543 // Bitcast to the right element type (for fp/vector values). 1544 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 1545 } else { 1546 // Otherwise, bitcast the dest pointer (for aggregates). 1547 DestField = new BitCastInst(DestField, 1548 PointerType::getUnqual(EltVal->getType()), 1549 "", SI); 1550 } 1551 new StoreInst(EltVal, DestField, SI); 1552 } 1553 1554 } else { 1555 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1556 const Type *ArrayEltTy = ATy->getElementType(); 1557 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1558 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1559 1560 uint64_t Shift; 1561 1562 if (TD->isBigEndian()) 1563 Shift = AllocaSizeBits-ElementOffset; 1564 else 1565 Shift = 0; 1566 1567 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1568 // Ignore zero sized fields like {}, they obviously contain no data. 1569 if (ElementSizeBits == 0) continue; 1570 1571 Value *EltVal = SrcVal; 1572 if (Shift) { 1573 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1574 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1575 "sroa.store.elt", SI); 1576 } 1577 1578 // Truncate down to an integer of the right size. 1579 if (ElementSizeBits != AllocaSizeBits) 1580 EltVal = new TruncInst(EltVal, 1581 IntegerType::get(SI->getContext(), 1582 ElementSizeBits),"",SI); 1583 Value *DestField = NewElts[i]; 1584 if (EltVal->getType() == ArrayEltTy) { 1585 // Storing to an integer field of this size, just do it. 1586 } else if (ArrayEltTy->isFloatingPointTy() || 1587 ArrayEltTy->isVectorTy()) { 1588 // Bitcast to the right element type (for fp/vector values). 1589 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1590 } else { 1591 // Otherwise, bitcast the dest pointer (for aggregates). 1592 DestField = new BitCastInst(DestField, 1593 PointerType::getUnqual(EltVal->getType()), 1594 "", SI); 1595 } 1596 new StoreInst(EltVal, DestField, SI); 1597 1598 if (TD->isBigEndian()) 1599 Shift -= ElementOffset; 1600 else 1601 Shift += ElementOffset; 1602 } 1603 } 1604 1605 DeadInsts.push_back(SI); 1606} 1607 1608/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1609/// an integer. Load the individual pieces to form the aggregate value. 1610void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1611 SmallVector<AllocaInst*, 32> &NewElts) { 1612 // Extract each element out of the NewElts according to its structure offset 1613 // and form the result value. 1614 const Type *AllocaEltTy = AI->getAllocatedType(); 1615 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1616 1617 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1618 << '\n'); 1619 1620 // There are two forms here: AI could be an array or struct. Both cases 1621 // have different ways to compute the element offset. 1622 const StructLayout *Layout = 0; 1623 uint64_t ArrayEltBitOffset = 0; 1624 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1625 Layout = TD->getStructLayout(EltSTy); 1626 } else { 1627 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1628 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1629 } 1630 1631 Value *ResultVal = 1632 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1633 1634 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1635 // Load the value from the alloca. If the NewElt is an aggregate, cast 1636 // the pointer to an integer of the same size before doing the load. 1637 Value *SrcField = NewElts[i]; 1638 const Type *FieldTy = 1639 cast<PointerType>(SrcField->getType())->getElementType(); 1640 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1641 1642 // Ignore zero sized fields like {}, they obviously contain no data. 1643 if (FieldSizeBits == 0) continue; 1644 1645 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1646 FieldSizeBits); 1647 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1648 !FieldTy->isVectorTy()) 1649 SrcField = new BitCastInst(SrcField, 1650 PointerType::getUnqual(FieldIntTy), 1651 "", LI); 1652 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1653 1654 // If SrcField is a fp or vector of the right size but that isn't an 1655 // integer type, bitcast to an integer so we can shift it. 1656 if (SrcField->getType() != FieldIntTy) 1657 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1658 1659 // Zero extend the field to be the same size as the final alloca so that 1660 // we can shift and insert it. 1661 if (SrcField->getType() != ResultVal->getType()) 1662 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1663 1664 // Determine the number of bits to shift SrcField. 1665 uint64_t Shift; 1666 if (Layout) // Struct case. 1667 Shift = Layout->getElementOffsetInBits(i); 1668 else // Array case. 1669 Shift = i*ArrayEltBitOffset; 1670 1671 if (TD->isBigEndian()) 1672 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1673 1674 if (Shift) { 1675 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1676 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1677 } 1678 1679 // Don't create an 'or x, 0' on the first iteration. 1680 if (!isa<Constant>(ResultVal) || 1681 !cast<Constant>(ResultVal)->isNullValue()) 1682 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1683 else 1684 ResultVal = SrcField; 1685 } 1686 1687 // Handle tail padding by truncating the result 1688 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1689 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1690 1691 LI->replaceAllUsesWith(ResultVal); 1692 DeadInsts.push_back(LI); 1693} 1694 1695/// HasPadding - Return true if the specified type has any structure or 1696/// alignment padding, false otherwise. 1697static bool HasPadding(const Type *Ty, const TargetData &TD) { 1698 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 1699 return HasPadding(ATy->getElementType(), TD); 1700 1701 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) 1702 return HasPadding(VTy->getElementType(), TD); 1703 1704 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1705 const StructLayout *SL = TD.getStructLayout(STy); 1706 unsigned PrevFieldBitOffset = 0; 1707 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1708 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1709 1710 // Padding in sub-elements? 1711 if (HasPadding(STy->getElementType(i), TD)) 1712 return true; 1713 1714 // Check to see if there is any padding between this element and the 1715 // previous one. 1716 if (i) { 1717 unsigned PrevFieldEnd = 1718 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1719 if (PrevFieldEnd < FieldBitOffset) 1720 return true; 1721 } 1722 1723 PrevFieldBitOffset = FieldBitOffset; 1724 } 1725 1726 // Check for tail padding. 1727 if (unsigned EltCount = STy->getNumElements()) { 1728 unsigned PrevFieldEnd = PrevFieldBitOffset + 1729 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1730 if (PrevFieldEnd < SL->getSizeInBits()) 1731 return true; 1732 } 1733 } 1734 1735 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1736} 1737 1738/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1739/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1740/// or 1 if safe after canonicalization has been performed. 1741bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1742 // Loop over the use list of the alloca. We can only transform it if all of 1743 // the users are safe to transform. 1744 AllocaInfo Info; 1745 1746 isSafeForScalarRepl(AI, AI, 0, Info); 1747 if (Info.isUnsafe) { 1748 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1749 return false; 1750 } 1751 1752 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1753 // source and destination, we have to be careful. In particular, the memcpy 1754 // could be moving around elements that live in structure padding of the LLVM 1755 // types, but may actually be used. In these cases, we refuse to promote the 1756 // struct. 1757 if (Info.isMemCpySrc && Info.isMemCpyDst && 1758 HasPadding(AI->getAllocatedType(), *TD)) 1759 return false; 1760 1761 return true; 1762} 1763 1764 1765 1766/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1767/// some part of a constant global variable. This intentionally only accepts 1768/// constant expressions because we don't can't rewrite arbitrary instructions. 1769static bool PointsToConstantGlobal(Value *V) { 1770 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1771 return GV->isConstant(); 1772 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1773 if (CE->getOpcode() == Instruction::BitCast || 1774 CE->getOpcode() == Instruction::GetElementPtr) 1775 return PointsToConstantGlobal(CE->getOperand(0)); 1776 return false; 1777} 1778 1779/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1780/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1781/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1782/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1783/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1784/// the alloca, and if the source pointer is a pointer to a constant global, we 1785/// can optimize this. 1786static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 1787 bool isOffset) { 1788 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1789 User *U = cast<Instruction>(*UI); 1790 1791 if (LoadInst *LI = dyn_cast<LoadInst>(U)) 1792 // Ignore non-volatile loads, they are always ok. 1793 if (!LI->isVolatile()) 1794 continue; 1795 1796 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1797 // If uses of the bitcast are ok, we are ok. 1798 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1799 return false; 1800 continue; 1801 } 1802 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 1803 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1804 // doesn't, it does. 1805 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1806 isOffset || !GEP->hasAllZeroIndices())) 1807 return false; 1808 continue; 1809 } 1810 1811 // If this is isn't our memcpy/memmove, reject it as something we can't 1812 // handle. 1813 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 1814 if (MI == 0) 1815 return false; 1816 1817 // If we already have seen a copy, reject the second one. 1818 if (TheCopy) return false; 1819 1820 // If the pointer has been offset from the start of the alloca, we can't 1821 // safely handle this. 1822 if (isOffset) return false; 1823 1824 // If the memintrinsic isn't using the alloca as the dest, reject it. 1825 if (UI.getOperandNo() != 0) return false; 1826 1827 // If the source of the memcpy/move is not a constant global, reject it. 1828 if (!PointsToConstantGlobal(MI->getSource())) 1829 return false; 1830 1831 // Otherwise, the transform is safe. Remember the copy instruction. 1832 TheCopy = MI; 1833 } 1834 return true; 1835} 1836 1837/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1838/// modified by a copy from a constant global. If we can prove this, we can 1839/// replace any uses of the alloca with uses of the global directly. 1840MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1841 MemTransferInst *TheCopy = 0; 1842 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1843 return TheCopy; 1844 return 0; 1845} 1846