ScalarReplAggregates.cpp revision f601d6df6f43bb833461cbcee475c36998e6c259
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/CallSite.h" 38#include "llvm/Support/Debug.h" 39#include "llvm/Support/ErrorHandling.h" 40#include "llvm/Support/GetElementPtrTypeIterator.h" 41#include "llvm/Support/IRBuilder.h" 42#include "llvm/Support/MathExtras.h" 43#include "llvm/Support/raw_ostream.h" 44#include "llvm/ADT/SmallVector.h" 45#include "llvm/ADT/Statistic.h" 46using namespace llvm; 47 48STATISTIC(NumReplaced, "Number of allocas broken up"); 49STATISTIC(NumPromoted, "Number of allocas promoted"); 50STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 51STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 52 53namespace { 54 struct SROA : public FunctionPass { 55 static char ID; // Pass identification, replacement for typeid 56 explicit SROA(signed T = -1) : FunctionPass(ID) { 57 initializeSROAPass(*PassRegistry::getPassRegistry()); 58 if (T == -1) 59 SRThreshold = 128; 60 else 61 SRThreshold = T; 62 } 63 64 bool runOnFunction(Function &F); 65 66 bool performScalarRepl(Function &F); 67 bool performPromotion(Function &F); 68 69 // getAnalysisUsage - This pass does not require any passes, but we know it 70 // will not alter the CFG, so say so. 71 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 72 AU.addRequired<DominatorTree>(); 73 AU.addRequired<DominanceFrontier>(); 74 AU.setPreservesCFG(); 75 } 76 77 private: 78 TargetData *TD; 79 80 /// DeadInsts - Keep track of instructions we have made dead, so that 81 /// we can remove them after we are done working. 82 SmallVector<Value*, 32> DeadInsts; 83 84 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 85 /// information about the uses. All these fields are initialized to false 86 /// and set to true when something is learned. 87 struct AllocaInfo { 88 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 89 bool isUnsafe : 1; 90 91 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 92 bool isMemCpySrc : 1; 93 94 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 95 bool isMemCpyDst : 1; 96 97 AllocaInfo() 98 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false) {} 99 }; 100 101 unsigned SRThreshold; 102 103 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 104 105 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 106 107 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 108 AllocaInfo &Info); 109 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset, 110 AllocaInfo &Info); 111 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 112 const Type *MemOpType, bool isStore, AllocaInfo &Info); 113 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 114 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 115 const Type *&IdxTy); 116 117 void DoScalarReplacement(AllocaInst *AI, 118 std::vector<AllocaInst*> &WorkList); 119 void DeleteDeadInstructions(); 120 121 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 122 SmallVector<AllocaInst*, 32> &NewElts); 123 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 124 SmallVector<AllocaInst*, 32> &NewElts); 125 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 126 SmallVector<AllocaInst*, 32> &NewElts); 127 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 128 AllocaInst *AI, 129 SmallVector<AllocaInst*, 32> &NewElts); 130 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 131 SmallVector<AllocaInst*, 32> &NewElts); 132 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 133 SmallVector<AllocaInst*, 32> &NewElts); 134 135 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 136 }; 137} 138 139char SROA::ID = 0; 140INITIALIZE_PASS_BEGIN(SROA, "scalarrepl", 141 "Scalar Replacement of Aggregates", false, false) 142INITIALIZE_PASS_DEPENDENCY(DominatorTree) 143INITIALIZE_PASS_DEPENDENCY(DominanceFrontier) 144INITIALIZE_PASS_END(SROA, "scalarrepl", 145 "Scalar Replacement of Aggregates", false, false) 146 147// Public interface to the ScalarReplAggregates pass 148FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 149 return new SROA(Threshold); 150} 151 152 153//===----------------------------------------------------------------------===// 154// Convert To Scalar Optimization. 155//===----------------------------------------------------------------------===// 156 157namespace { 158/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 159/// optimization, which scans the uses of an alloca and determines if it can 160/// rewrite it in terms of a single new alloca that can be mem2reg'd. 161class ConvertToScalarInfo { 162 /// AllocaSize - The size of the alloca being considered. 163 unsigned AllocaSize; 164 const TargetData &TD; 165 166 /// IsNotTrivial - This is set to true if there is some access to the object 167 /// which means that mem2reg can't promote it. 168 bool IsNotTrivial; 169 170 /// VectorTy - This tracks the type that we should promote the vector to if 171 /// it is possible to turn it into a vector. This starts out null, and if it 172 /// isn't possible to turn into a vector type, it gets set to VoidTy. 173 const Type *VectorTy; 174 175 /// HadAVector - True if there is at least one vector access to the alloca. 176 /// We don't want to turn random arrays into vectors and use vector element 177 /// insert/extract, but if there are element accesses to something that is 178 /// also declared as a vector, we do want to promote to a vector. 179 bool HadAVector; 180 181public: 182 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 183 : AllocaSize(Size), TD(td) { 184 IsNotTrivial = false; 185 VectorTy = 0; 186 HadAVector = false; 187 } 188 189 AllocaInst *TryConvert(AllocaInst *AI); 190 191private: 192 bool CanConvertToScalar(Value *V, uint64_t Offset); 193 void MergeInType(const Type *In, uint64_t Offset); 194 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 195 196 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 197 uint64_t Offset, IRBuilder<> &Builder); 198 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 199 uint64_t Offset, IRBuilder<> &Builder); 200}; 201} // end anonymous namespace. 202 203 204/// IsVerbotenVectorType - Return true if this is a vector type ScalarRepl isn't 205/// allowed to form. We do this to avoid MMX types, which is a complete hack, 206/// but is required until the backend is fixed. 207static bool IsVerbotenVectorType(const VectorType *VTy, const Instruction *I) { 208 StringRef Triple(I->getParent()->getParent()->getParent()->getTargetTriple()); 209 if (!Triple.startswith("i386") && 210 !Triple.startswith("x86_64")) 211 return false; 212 213 // Reject all the MMX vector types. 214 switch (VTy->getNumElements()) { 215 default: return false; 216 case 1: return VTy->getElementType()->isIntegerTy(64); 217 case 2: return VTy->getElementType()->isIntegerTy(32); 218 case 4: return VTy->getElementType()->isIntegerTy(16); 219 case 8: return VTy->getElementType()->isIntegerTy(8); 220 } 221} 222 223 224/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 225/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 226/// alloca if possible or null if not. 227AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 228 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 229 // out. 230 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 231 return 0; 232 233 // If we were able to find a vector type that can handle this with 234 // insert/extract elements, and if there was at least one use that had 235 // a vector type, promote this to a vector. We don't want to promote 236 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 237 // we just get a lot of insert/extracts. If at least one vector is 238 // involved, then we probably really do have a union of vector/array. 239 const Type *NewTy; 240 if (VectorTy && VectorTy->isVectorTy() && HadAVector && 241 !IsVerbotenVectorType(cast<VectorType>(VectorTy), AI)) { 242 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 243 << *VectorTy << '\n'); 244 NewTy = VectorTy; // Use the vector type. 245 } else { 246 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 247 // Create and insert the integer alloca. 248 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 249 } 250 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 251 ConvertUsesToScalar(AI, NewAI, 0); 252 return NewAI; 253} 254 255/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy) 256/// so far at the offset specified by Offset (which is specified in bytes). 257/// 258/// There are two cases we handle here: 259/// 1) A union of vector types of the same size and potentially its elements. 260/// Here we turn element accesses into insert/extract element operations. 261/// This promotes a <4 x float> with a store of float to the third element 262/// into a <4 x float> that uses insert element. 263/// 2) A fully general blob of memory, which we turn into some (potentially 264/// large) integer type with extract and insert operations where the loads 265/// and stores would mutate the memory. We mark this by setting VectorTy 266/// to VoidTy. 267void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) { 268 // If we already decided to turn this into a blob of integer memory, there is 269 // nothing to be done. 270 if (VectorTy && VectorTy->isVoidTy()) 271 return; 272 273 // If this could be contributing to a vector, analyze it. 274 275 // If the In type is a vector that is the same size as the alloca, see if it 276 // matches the existing VecTy. 277 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 278 // Remember if we saw a vector type. 279 HadAVector = true; 280 281 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 282 // If we're storing/loading a vector of the right size, allow it as a 283 // vector. If this the first vector we see, remember the type so that 284 // we know the element size. If this is a subsequent access, ignore it 285 // even if it is a differing type but the same size. Worst case we can 286 // bitcast the resultant vectors. 287 if (VectorTy == 0) 288 VectorTy = VInTy; 289 return; 290 } 291 } else if (In->isFloatTy() || In->isDoubleTy() || 292 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 293 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 294 // If we're accessing something that could be an element of a vector, see 295 // if the implied vector agrees with what we already have and if Offset is 296 // compatible with it. 297 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 298 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 299 (VectorTy == 0 || 300 cast<VectorType>(VectorTy)->getElementType() 301 ->getPrimitiveSizeInBits()/8 == EltSize)) { 302 if (VectorTy == 0) 303 VectorTy = VectorType::get(In, AllocaSize/EltSize); 304 return; 305 } 306 } 307 308 // Otherwise, we have a case that we can't handle with an optimized vector 309 // form. We can still turn this into a large integer. 310 VectorTy = Type::getVoidTy(In->getContext()); 311} 312 313/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 314/// its accesses to a single vector type, return true and set VecTy to 315/// the new type. If we could convert the alloca into a single promotable 316/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 317/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 318/// is the current offset from the base of the alloca being analyzed. 319/// 320/// If we see at least one access to the value that is as a vector type, set the 321/// SawVec flag. 322bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 323 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 324 Instruction *User = cast<Instruction>(*UI); 325 326 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 327 // Don't break volatile loads. 328 if (LI->isVolatile()) 329 return false; 330 // Don't touch MMX operations. 331 if (LI->getType()->isX86_MMXTy()) 332 return false; 333 MergeInType(LI->getType(), Offset); 334 continue; 335 } 336 337 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 338 // Storing the pointer, not into the value? 339 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 340 // Don't touch MMX operations. 341 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 342 return false; 343 MergeInType(SI->getOperand(0)->getType(), Offset); 344 continue; 345 } 346 347 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 348 IsNotTrivial = true; // Can't be mem2reg'd. 349 if (!CanConvertToScalar(BCI, Offset)) 350 return false; 351 continue; 352 } 353 354 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 355 // If this is a GEP with a variable indices, we can't handle it. 356 if (!GEP->hasAllConstantIndices()) 357 return false; 358 359 // Compute the offset that this GEP adds to the pointer. 360 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 361 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 362 &Indices[0], Indices.size()); 363 // See if all uses can be converted. 364 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 365 return false; 366 IsNotTrivial = true; // Can't be mem2reg'd. 367 continue; 368 } 369 370 // If this is a constant sized memset of a constant value (e.g. 0) we can 371 // handle it. 372 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 373 // Store of constant value and constant size. 374 if (!isa<ConstantInt>(MSI->getValue()) || 375 !isa<ConstantInt>(MSI->getLength())) 376 return false; 377 IsNotTrivial = true; // Can't be mem2reg'd. 378 continue; 379 } 380 381 // If this is a memcpy or memmove into or out of the whole allocation, we 382 // can handle it like a load or store of the scalar type. 383 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 384 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 385 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 386 return false; 387 388 IsNotTrivial = true; // Can't be mem2reg'd. 389 continue; 390 } 391 392 // Otherwise, we cannot handle this! 393 return false; 394 } 395 396 return true; 397} 398 399/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 400/// directly. This happens when we are converting an "integer union" to a 401/// single integer scalar, or when we are converting a "vector union" to a 402/// vector with insert/extractelement instructions. 403/// 404/// Offset is an offset from the original alloca, in bits that need to be 405/// shifted to the right. By the end of this, there should be no uses of Ptr. 406void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 407 uint64_t Offset) { 408 while (!Ptr->use_empty()) { 409 Instruction *User = cast<Instruction>(Ptr->use_back()); 410 411 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 412 ConvertUsesToScalar(CI, NewAI, Offset); 413 CI->eraseFromParent(); 414 continue; 415 } 416 417 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 418 // Compute the offset that this GEP adds to the pointer. 419 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 420 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 421 &Indices[0], Indices.size()); 422 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 423 GEP->eraseFromParent(); 424 continue; 425 } 426 427 IRBuilder<> Builder(User->getParent(), User); 428 429 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 430 // The load is a bit extract from NewAI shifted right by Offset bits. 431 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 432 Value *NewLoadVal 433 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 434 LI->replaceAllUsesWith(NewLoadVal); 435 LI->eraseFromParent(); 436 continue; 437 } 438 439 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 440 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 441 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 442 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 443 Builder); 444 Builder.CreateStore(New, NewAI); 445 SI->eraseFromParent(); 446 447 // If the load we just inserted is now dead, then the inserted store 448 // overwrote the entire thing. 449 if (Old->use_empty()) 450 Old->eraseFromParent(); 451 continue; 452 } 453 454 // If this is a constant sized memset of a constant value (e.g. 0) we can 455 // transform it into a store of the expanded constant value. 456 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 457 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 458 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 459 if (NumBytes != 0) { 460 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 461 462 // Compute the value replicated the right number of times. 463 APInt APVal(NumBytes*8, Val); 464 465 // Splat the value if non-zero. 466 if (Val) 467 for (unsigned i = 1; i != NumBytes; ++i) 468 APVal |= APVal << 8; 469 470 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 471 Value *New = ConvertScalar_InsertValue( 472 ConstantInt::get(User->getContext(), APVal), 473 Old, Offset, Builder); 474 Builder.CreateStore(New, NewAI); 475 476 // If the load we just inserted is now dead, then the memset overwrote 477 // the entire thing. 478 if (Old->use_empty()) 479 Old->eraseFromParent(); 480 } 481 MSI->eraseFromParent(); 482 continue; 483 } 484 485 // If this is a memcpy or memmove into or out of the whole allocation, we 486 // can handle it like a load or store of the scalar type. 487 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 488 assert(Offset == 0 && "must be store to start of alloca"); 489 490 // If the source and destination are both to the same alloca, then this is 491 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 492 // as appropriate. 493 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject(0)); 494 495 if (MTI->getSource()->getUnderlyingObject(0) != OrigAI) { 496 // Dest must be OrigAI, change this to be a load from the original 497 // pointer (bitcasted), then a store to our new alloca. 498 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 499 Value *SrcPtr = MTI->getSource(); 500 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 501 502 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 503 SrcVal->setAlignment(MTI->getAlignment()); 504 Builder.CreateStore(SrcVal, NewAI); 505 } else if (MTI->getDest()->getUnderlyingObject(0) != OrigAI) { 506 // Src must be OrigAI, change this to be a load from NewAI then a store 507 // through the original dest pointer (bitcasted). 508 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 509 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 510 511 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 512 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 513 NewStore->setAlignment(MTI->getAlignment()); 514 } else { 515 // Noop transfer. Src == Dst 516 } 517 518 MTI->eraseFromParent(); 519 continue; 520 } 521 522 llvm_unreachable("Unsupported operation!"); 523 } 524} 525 526/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 527/// or vector value FromVal, extracting the bits from the offset specified by 528/// Offset. This returns the value, which is of type ToType. 529/// 530/// This happens when we are converting an "integer union" to a single 531/// integer scalar, or when we are converting a "vector union" to a vector with 532/// insert/extractelement instructions. 533/// 534/// Offset is an offset from the original alloca, in bits that need to be 535/// shifted to the right. 536Value *ConvertToScalarInfo:: 537ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 538 uint64_t Offset, IRBuilder<> &Builder) { 539 // If the load is of the whole new alloca, no conversion is needed. 540 if (FromVal->getType() == ToType && Offset == 0) 541 return FromVal; 542 543 // If the result alloca is a vector type, this is either an element 544 // access or a bitcast to another vector type of the same size. 545 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 546 if (ToType->isVectorTy()) 547 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 548 549 // Otherwise it must be an element access. 550 unsigned Elt = 0; 551 if (Offset) { 552 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 553 Elt = Offset/EltSize; 554 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 555 } 556 // Return the element extracted out of it. 557 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 558 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 559 if (V->getType() != ToType) 560 V = Builder.CreateBitCast(V, ToType, "tmp"); 561 return V; 562 } 563 564 // If ToType is a first class aggregate, extract out each of the pieces and 565 // use insertvalue's to form the FCA. 566 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 567 const StructLayout &Layout = *TD.getStructLayout(ST); 568 Value *Res = UndefValue::get(ST); 569 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 570 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 571 Offset+Layout.getElementOffsetInBits(i), 572 Builder); 573 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 574 } 575 return Res; 576 } 577 578 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 579 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 580 Value *Res = UndefValue::get(AT); 581 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 582 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 583 Offset+i*EltSize, Builder); 584 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 585 } 586 return Res; 587 } 588 589 // Otherwise, this must be a union that was converted to an integer value. 590 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 591 592 // If this is a big-endian system and the load is narrower than the 593 // full alloca type, we need to do a shift to get the right bits. 594 int ShAmt = 0; 595 if (TD.isBigEndian()) { 596 // On big-endian machines, the lowest bit is stored at the bit offset 597 // from the pointer given by getTypeStoreSizeInBits. This matters for 598 // integers with a bitwidth that is not a multiple of 8. 599 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 600 TD.getTypeStoreSizeInBits(ToType) - Offset; 601 } else { 602 ShAmt = Offset; 603 } 604 605 // Note: we support negative bitwidths (with shl) which are not defined. 606 // We do this to support (f.e.) loads off the end of a structure where 607 // only some bits are used. 608 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 609 FromVal = Builder.CreateLShr(FromVal, 610 ConstantInt::get(FromVal->getType(), 611 ShAmt), "tmp"); 612 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 613 FromVal = Builder.CreateShl(FromVal, 614 ConstantInt::get(FromVal->getType(), 615 -ShAmt), "tmp"); 616 617 // Finally, unconditionally truncate the integer to the right width. 618 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 619 if (LIBitWidth < NTy->getBitWidth()) 620 FromVal = 621 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 622 LIBitWidth), "tmp"); 623 else if (LIBitWidth > NTy->getBitWidth()) 624 FromVal = 625 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 626 LIBitWidth), "tmp"); 627 628 // If the result is an integer, this is a trunc or bitcast. 629 if (ToType->isIntegerTy()) { 630 // Should be done. 631 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 632 // Just do a bitcast, we know the sizes match up. 633 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 634 } else { 635 // Otherwise must be a pointer. 636 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 637 } 638 assert(FromVal->getType() == ToType && "Didn't convert right?"); 639 return FromVal; 640} 641 642/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 643/// or vector value "Old" at the offset specified by Offset. 644/// 645/// This happens when we are converting an "integer union" to a 646/// single integer scalar, or when we are converting a "vector union" to a 647/// vector with insert/extractelement instructions. 648/// 649/// Offset is an offset from the original alloca, in bits that need to be 650/// shifted to the right. 651Value *ConvertToScalarInfo:: 652ConvertScalar_InsertValue(Value *SV, Value *Old, 653 uint64_t Offset, IRBuilder<> &Builder) { 654 // Convert the stored type to the actual type, shift it left to insert 655 // then 'or' into place. 656 const Type *AllocaType = Old->getType(); 657 LLVMContext &Context = Old->getContext(); 658 659 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 660 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 661 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 662 663 // Changing the whole vector with memset or with an access of a different 664 // vector type? 665 if (ValSize == VecSize) 666 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 667 668 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 669 670 // Must be an element insertion. 671 unsigned Elt = Offset/EltSize; 672 673 if (SV->getType() != VTy->getElementType()) 674 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 675 676 SV = Builder.CreateInsertElement(Old, SV, 677 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 678 "tmp"); 679 return SV; 680 } 681 682 // If SV is a first-class aggregate value, insert each value recursively. 683 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 684 const StructLayout &Layout = *TD.getStructLayout(ST); 685 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 686 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 687 Old = ConvertScalar_InsertValue(Elt, Old, 688 Offset+Layout.getElementOffsetInBits(i), 689 Builder); 690 } 691 return Old; 692 } 693 694 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 695 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 696 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 697 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 698 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 699 } 700 return Old; 701 } 702 703 // If SV is a float, convert it to the appropriate integer type. 704 // If it is a pointer, do the same. 705 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 706 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 707 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 708 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 709 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 710 SV = Builder.CreateBitCast(SV, 711 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 712 else if (SV->getType()->isPointerTy()) 713 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 714 715 // Zero extend or truncate the value if needed. 716 if (SV->getType() != AllocaType) { 717 if (SV->getType()->getPrimitiveSizeInBits() < 718 AllocaType->getPrimitiveSizeInBits()) 719 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 720 else { 721 // Truncation may be needed if storing more than the alloca can hold 722 // (undefined behavior). 723 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 724 SrcWidth = DestWidth; 725 SrcStoreWidth = DestStoreWidth; 726 } 727 } 728 729 // If this is a big-endian system and the store is narrower than the 730 // full alloca type, we need to do a shift to get the right bits. 731 int ShAmt = 0; 732 if (TD.isBigEndian()) { 733 // On big-endian machines, the lowest bit is stored at the bit offset 734 // from the pointer given by getTypeStoreSizeInBits. This matters for 735 // integers with a bitwidth that is not a multiple of 8. 736 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 737 } else { 738 ShAmt = Offset; 739 } 740 741 // Note: we support negative bitwidths (with shr) which are not defined. 742 // We do this to support (f.e.) stores off the end of a structure where 743 // only some bits in the structure are set. 744 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 745 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 746 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 747 ShAmt), "tmp"); 748 Mask <<= ShAmt; 749 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 750 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 751 -ShAmt), "tmp"); 752 Mask = Mask.lshr(-ShAmt); 753 } 754 755 // Mask out the bits we are about to insert from the old value, and or 756 // in the new bits. 757 if (SrcWidth != DestWidth) { 758 assert(DestWidth > SrcWidth); 759 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 760 SV = Builder.CreateOr(Old, SV, "ins"); 761 } 762 return SV; 763} 764 765 766//===----------------------------------------------------------------------===// 767// SRoA Driver 768//===----------------------------------------------------------------------===// 769 770 771bool SROA::runOnFunction(Function &F) { 772 TD = getAnalysisIfAvailable<TargetData>(); 773 774 bool Changed = performPromotion(F); 775 776 // FIXME: ScalarRepl currently depends on TargetData more than it 777 // theoretically needs to. It should be refactored in order to support 778 // target-independent IR. Until this is done, just skip the actual 779 // scalar-replacement portion of this pass. 780 if (!TD) return Changed; 781 782 while (1) { 783 bool LocalChange = performScalarRepl(F); 784 if (!LocalChange) break; // No need to repromote if no scalarrepl 785 Changed = true; 786 LocalChange = performPromotion(F); 787 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 788 } 789 790 return Changed; 791} 792 793 794bool SROA::performPromotion(Function &F) { 795 std::vector<AllocaInst*> Allocas; 796 DominatorTree &DT = getAnalysis<DominatorTree>(); 797 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 798 799 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 800 801 bool Changed = false; 802 803 while (1) { 804 Allocas.clear(); 805 806 // Find allocas that are safe to promote, by looking at all instructions in 807 // the entry node 808 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 809 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 810 if (isAllocaPromotable(AI)) 811 Allocas.push_back(AI); 812 813 if (Allocas.empty()) break; 814 815 PromoteMemToReg(Allocas, DT, DF); 816 NumPromoted += Allocas.size(); 817 Changed = true; 818 } 819 820 return Changed; 821} 822 823 824/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 825/// SROA. It must be a struct or array type with a small number of elements. 826static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 827 const Type *T = AI->getAllocatedType(); 828 // Do not promote any struct into more than 32 separate vars. 829 if (const StructType *ST = dyn_cast<StructType>(T)) 830 return ST->getNumElements() <= 32; 831 // Arrays are much less likely to be safe for SROA; only consider 832 // them if they are very small. 833 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 834 return AT->getNumElements() <= 8; 835 return false; 836} 837 838 839// performScalarRepl - This algorithm is a simple worklist driven algorithm, 840// which runs on all of the malloc/alloca instructions in the function, removing 841// them if they are only used by getelementptr instructions. 842// 843bool SROA::performScalarRepl(Function &F) { 844 std::vector<AllocaInst*> WorkList; 845 846 // Scan the entry basic block, adding allocas to the worklist. 847 BasicBlock &BB = F.getEntryBlock(); 848 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 849 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 850 WorkList.push_back(A); 851 852 // Process the worklist 853 bool Changed = false; 854 while (!WorkList.empty()) { 855 AllocaInst *AI = WorkList.back(); 856 WorkList.pop_back(); 857 858 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 859 // with unused elements. 860 if (AI->use_empty()) { 861 AI->eraseFromParent(); 862 Changed = true; 863 continue; 864 } 865 866 // If this alloca is impossible for us to promote, reject it early. 867 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 868 continue; 869 870 // Check to see if this allocation is only modified by a memcpy/memmove from 871 // a constant global. If this is the case, we can change all users to use 872 // the constant global instead. This is commonly produced by the CFE by 873 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 874 // is only subsequently read. 875 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 876 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 877 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 878 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 879 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 880 TheCopy->eraseFromParent(); // Don't mutate the global. 881 AI->eraseFromParent(); 882 ++NumGlobals; 883 Changed = true; 884 continue; 885 } 886 887 // Check to see if we can perform the core SROA transformation. We cannot 888 // transform the allocation instruction if it is an array allocation 889 // (allocations OF arrays are ok though), and an allocation of a scalar 890 // value cannot be decomposed at all. 891 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 892 893 // Do not promote [0 x %struct]. 894 if (AllocaSize == 0) continue; 895 896 // Do not promote any struct whose size is too big. 897 if (AllocaSize > SRThreshold) continue; 898 899 // If the alloca looks like a good candidate for scalar replacement, and if 900 // all its users can be transformed, then split up the aggregate into its 901 // separate elements. 902 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 903 DoScalarReplacement(AI, WorkList); 904 Changed = true; 905 continue; 906 } 907 908 // If we can turn this aggregate value (potentially with casts) into a 909 // simple scalar value that can be mem2reg'd into a register value. 910 // IsNotTrivial tracks whether this is something that mem2reg could have 911 // promoted itself. If so, we don't want to transform it needlessly. Note 912 // that we can't just check based on the type: the alloca may be of an i32 913 // but that has pointer arithmetic to set byte 3 of it or something. 914 if (AllocaInst *NewAI = 915 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 916 NewAI->takeName(AI); 917 AI->eraseFromParent(); 918 ++NumConverted; 919 Changed = true; 920 continue; 921 } 922 923 // Otherwise, couldn't process this alloca. 924 } 925 926 return Changed; 927} 928 929/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 930/// predicate, do SROA now. 931void SROA::DoScalarReplacement(AllocaInst *AI, 932 std::vector<AllocaInst*> &WorkList) { 933 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 934 SmallVector<AllocaInst*, 32> ElementAllocas; 935 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 936 ElementAllocas.reserve(ST->getNumContainedTypes()); 937 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 938 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 939 AI->getAlignment(), 940 AI->getName() + "." + Twine(i), AI); 941 ElementAllocas.push_back(NA); 942 WorkList.push_back(NA); // Add to worklist for recursive processing 943 } 944 } else { 945 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 946 ElementAllocas.reserve(AT->getNumElements()); 947 const Type *ElTy = AT->getElementType(); 948 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 949 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 950 AI->getName() + "." + Twine(i), AI); 951 ElementAllocas.push_back(NA); 952 WorkList.push_back(NA); // Add to worklist for recursive processing 953 } 954 } 955 956 // Now that we have created the new alloca instructions, rewrite all the 957 // uses of the old alloca. 958 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 959 960 // Now erase any instructions that were made dead while rewriting the alloca. 961 DeleteDeadInstructions(); 962 AI->eraseFromParent(); 963 964 ++NumReplaced; 965} 966 967/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 968/// recursively including all their operands that become trivially dead. 969void SROA::DeleteDeadInstructions() { 970 while (!DeadInsts.empty()) { 971 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 972 973 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 974 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 975 // Zero out the operand and see if it becomes trivially dead. 976 // (But, don't add allocas to the dead instruction list -- they are 977 // already on the worklist and will be deleted separately.) 978 *OI = 0; 979 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 980 DeadInsts.push_back(U); 981 } 982 983 I->eraseFromParent(); 984 } 985} 986 987/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 988/// performing scalar replacement of alloca AI. The results are flagged in 989/// the Info parameter. Offset indicates the position within AI that is 990/// referenced by this instruction. 991void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 992 AllocaInfo &Info) { 993 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 994 Instruction *User = cast<Instruction>(*UI); 995 996 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 997 isSafeForScalarRepl(BC, AI, Offset, Info); 998 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 999 uint64_t GEPOffset = Offset; 1000 isSafeGEP(GEPI, AI, GEPOffset, Info); 1001 if (!Info.isUnsafe) 1002 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 1003 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1004 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1005 if (Length) 1006 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 1007 UI.getOperandNo() == 0, Info); 1008 else 1009 MarkUnsafe(Info); 1010 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1011 if (!LI->isVolatile()) { 1012 const Type *LIType = LI->getType(); 1013 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 1014 LIType, false, Info); 1015 } else 1016 MarkUnsafe(Info); 1017 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1018 // Store is ok if storing INTO the pointer, not storing the pointer 1019 if (!SI->isVolatile() && SI->getOperand(0) != I) { 1020 const Type *SIType = SI->getOperand(0)->getType(); 1021 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 1022 SIType, true, Info); 1023 } else 1024 MarkUnsafe(Info); 1025 } else { 1026 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 1027 MarkUnsafe(Info); 1028 } 1029 if (Info.isUnsafe) return; 1030 } 1031} 1032 1033/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1034/// replacement. It is safe when all the indices are constant, in-bounds 1035/// references, and when the resulting offset corresponds to an element within 1036/// the alloca type. The results are flagged in the Info parameter. Upon 1037/// return, Offset is adjusted as specified by the GEP indices. 1038void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 1039 uint64_t &Offset, AllocaInfo &Info) { 1040 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1041 if (GEPIt == E) 1042 return; 1043 1044 // Walk through the GEP type indices, checking the types that this indexes 1045 // into. 1046 for (; GEPIt != E; ++GEPIt) { 1047 // Ignore struct elements, no extra checking needed for these. 1048 if ((*GEPIt)->isStructTy()) 1049 continue; 1050 1051 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1052 if (!IdxVal) 1053 return MarkUnsafe(Info); 1054 } 1055 1056 // Compute the offset due to this GEP and check if the alloca has a 1057 // component element at that offset. 1058 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1059 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1060 &Indices[0], Indices.size()); 1061 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 1062 MarkUnsafe(Info); 1063} 1064 1065/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1066/// alloca or has an offset and size that corresponds to a component element 1067/// within it. The offset checked here may have been formed from a GEP with a 1068/// pointer bitcasted to a different type. 1069void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 1070 const Type *MemOpType, bool isStore, 1071 AllocaInfo &Info) { 1072 // Check if this is a load/store of the entire alloca. 1073 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 1074 bool UsesAggregateType = (MemOpType == AI->getAllocatedType()); 1075 // This is safe for MemIntrinsics (where MemOpType is 0), integer types 1076 // (which are essentially the same as the MemIntrinsics, especially with 1077 // regard to copying padding between elements), or references using the 1078 // aggregate type of the alloca. 1079 if (!MemOpType || MemOpType->isIntegerTy() || UsesAggregateType) { 1080 if (!UsesAggregateType) { 1081 if (isStore) 1082 Info.isMemCpyDst = true; 1083 else 1084 Info.isMemCpySrc = true; 1085 } 1086 return; 1087 } 1088 } 1089 // Check if the offset/size correspond to a component within the alloca type. 1090 const Type *T = AI->getAllocatedType(); 1091 if (TypeHasComponent(T, Offset, MemSize)) 1092 return; 1093 1094 return MarkUnsafe(Info); 1095} 1096 1097/// TypeHasComponent - Return true if T has a component type with the 1098/// specified offset and size. If Size is zero, do not check the size. 1099bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1100 const Type *EltTy; 1101 uint64_t EltSize; 1102 if (const StructType *ST = dyn_cast<StructType>(T)) { 1103 const StructLayout *Layout = TD->getStructLayout(ST); 1104 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1105 EltTy = ST->getContainedType(EltIdx); 1106 EltSize = TD->getTypeAllocSize(EltTy); 1107 Offset -= Layout->getElementOffset(EltIdx); 1108 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1109 EltTy = AT->getElementType(); 1110 EltSize = TD->getTypeAllocSize(EltTy); 1111 if (Offset >= AT->getNumElements() * EltSize) 1112 return false; 1113 Offset %= EltSize; 1114 } else { 1115 return false; 1116 } 1117 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1118 return true; 1119 // Check if the component spans multiple elements. 1120 if (Offset + Size > EltSize) 1121 return false; 1122 return TypeHasComponent(EltTy, Offset, Size); 1123} 1124 1125/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1126/// the instruction I, which references it, to use the separate elements. 1127/// Offset indicates the position within AI that is referenced by this 1128/// instruction. 1129void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1130 SmallVector<AllocaInst*, 32> &NewElts) { 1131 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1132 Instruction *User = cast<Instruction>(*UI); 1133 1134 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1135 RewriteBitCast(BC, AI, Offset, NewElts); 1136 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1137 RewriteGEP(GEPI, AI, Offset, NewElts); 1138 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1139 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1140 uint64_t MemSize = Length->getZExtValue(); 1141 if (Offset == 0 && 1142 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1143 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1144 // Otherwise the intrinsic can only touch a single element and the 1145 // address operand will be updated, so nothing else needs to be done. 1146 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1147 const Type *LIType = LI->getType(); 1148 if (LIType == AI->getAllocatedType()) { 1149 // Replace: 1150 // %res = load { i32, i32 }* %alloc 1151 // with: 1152 // %load.0 = load i32* %alloc.0 1153 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1154 // %load.1 = load i32* %alloc.1 1155 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1156 // (Also works for arrays instead of structs) 1157 Value *Insert = UndefValue::get(LIType); 1158 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1159 Value *Load = new LoadInst(NewElts[i], "load", LI); 1160 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1161 } 1162 LI->replaceAllUsesWith(Insert); 1163 DeadInsts.push_back(LI); 1164 } else if (LIType->isIntegerTy() && 1165 TD->getTypeAllocSize(LIType) == 1166 TD->getTypeAllocSize(AI->getAllocatedType())) { 1167 // If this is a load of the entire alloca to an integer, rewrite it. 1168 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1169 } 1170 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1171 Value *Val = SI->getOperand(0); 1172 const Type *SIType = Val->getType(); 1173 if (SIType == AI->getAllocatedType()) { 1174 // Replace: 1175 // store { i32, i32 } %val, { i32, i32 }* %alloc 1176 // with: 1177 // %val.0 = extractvalue { i32, i32 } %val, 0 1178 // store i32 %val.0, i32* %alloc.0 1179 // %val.1 = extractvalue { i32, i32 } %val, 1 1180 // store i32 %val.1, i32* %alloc.1 1181 // (Also works for arrays instead of structs) 1182 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1183 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1184 new StoreInst(Extract, NewElts[i], SI); 1185 } 1186 DeadInsts.push_back(SI); 1187 } else if (SIType->isIntegerTy() && 1188 TD->getTypeAllocSize(SIType) == 1189 TD->getTypeAllocSize(AI->getAllocatedType())) { 1190 // If this is a store of the entire alloca from an integer, rewrite it. 1191 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1192 } 1193 } 1194 } 1195} 1196 1197/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1198/// and recursively continue updating all of its uses. 1199void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1200 SmallVector<AllocaInst*, 32> &NewElts) { 1201 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1202 if (BC->getOperand(0) != AI) 1203 return; 1204 1205 // The bitcast references the original alloca. Replace its uses with 1206 // references to the first new element alloca. 1207 Instruction *Val = NewElts[0]; 1208 if (Val->getType() != BC->getDestTy()) { 1209 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1210 Val->takeName(BC); 1211 } 1212 BC->replaceAllUsesWith(Val); 1213 DeadInsts.push_back(BC); 1214} 1215 1216/// FindElementAndOffset - Return the index of the element containing Offset 1217/// within the specified type, which must be either a struct or an array. 1218/// Sets T to the type of the element and Offset to the offset within that 1219/// element. IdxTy is set to the type of the index result to be used in a 1220/// GEP instruction. 1221uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1222 const Type *&IdxTy) { 1223 uint64_t Idx = 0; 1224 if (const StructType *ST = dyn_cast<StructType>(T)) { 1225 const StructLayout *Layout = TD->getStructLayout(ST); 1226 Idx = Layout->getElementContainingOffset(Offset); 1227 T = ST->getContainedType(Idx); 1228 Offset -= Layout->getElementOffset(Idx); 1229 IdxTy = Type::getInt32Ty(T->getContext()); 1230 return Idx; 1231 } 1232 const ArrayType *AT = cast<ArrayType>(T); 1233 T = AT->getElementType(); 1234 uint64_t EltSize = TD->getTypeAllocSize(T); 1235 Idx = Offset / EltSize; 1236 Offset -= Idx * EltSize; 1237 IdxTy = Type::getInt64Ty(T->getContext()); 1238 return Idx; 1239} 1240 1241/// RewriteGEP - Check if this GEP instruction moves the pointer across 1242/// elements of the alloca that are being split apart, and if so, rewrite 1243/// the GEP to be relative to the new element. 1244void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1245 SmallVector<AllocaInst*, 32> &NewElts) { 1246 uint64_t OldOffset = Offset; 1247 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1248 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1249 &Indices[0], Indices.size()); 1250 1251 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1252 1253 const Type *T = AI->getAllocatedType(); 1254 const Type *IdxTy; 1255 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1256 if (GEPI->getOperand(0) == AI) 1257 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1258 1259 T = AI->getAllocatedType(); 1260 uint64_t EltOffset = Offset; 1261 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1262 1263 // If this GEP does not move the pointer across elements of the alloca 1264 // being split, then it does not needs to be rewritten. 1265 if (Idx == OldIdx) 1266 return; 1267 1268 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1269 SmallVector<Value*, 8> NewArgs; 1270 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1271 while (EltOffset != 0) { 1272 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1273 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1274 } 1275 Instruction *Val = NewElts[Idx]; 1276 if (NewArgs.size() > 1) { 1277 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1278 NewArgs.end(), "", GEPI); 1279 Val->takeName(GEPI); 1280 } 1281 if (Val->getType() != GEPI->getType()) 1282 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1283 GEPI->replaceAllUsesWith(Val); 1284 DeadInsts.push_back(GEPI); 1285} 1286 1287/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1288/// Rewrite it to copy or set the elements of the scalarized memory. 1289void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1290 AllocaInst *AI, 1291 SmallVector<AllocaInst*, 32> &NewElts) { 1292 // If this is a memcpy/memmove, construct the other pointer as the 1293 // appropriate type. The "Other" pointer is the pointer that goes to memory 1294 // that doesn't have anything to do with the alloca that we are promoting. For 1295 // memset, this Value* stays null. 1296 Value *OtherPtr = 0; 1297 unsigned MemAlignment = MI->getAlignment(); 1298 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1299 if (Inst == MTI->getRawDest()) 1300 OtherPtr = MTI->getRawSource(); 1301 else { 1302 assert(Inst == MTI->getRawSource()); 1303 OtherPtr = MTI->getRawDest(); 1304 } 1305 } 1306 1307 // If there is an other pointer, we want to convert it to the same pointer 1308 // type as AI has, so we can GEP through it safely. 1309 if (OtherPtr) { 1310 unsigned AddrSpace = 1311 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 1312 1313 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1314 // optimization, but it's also required to detect the corner case where 1315 // both pointer operands are referencing the same memory, and where 1316 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1317 // function is only called for mem intrinsics that access the whole 1318 // aggregate, so non-zero GEPs are not an issue here.) 1319 OtherPtr = OtherPtr->stripPointerCasts(); 1320 1321 // Copying the alloca to itself is a no-op: just delete it. 1322 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1323 // This code will run twice for a no-op memcpy -- once for each operand. 1324 // Put only one reference to MI on the DeadInsts list. 1325 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1326 E = DeadInsts.end(); I != E; ++I) 1327 if (*I == MI) return; 1328 DeadInsts.push_back(MI); 1329 return; 1330 } 1331 1332 // If the pointer is not the right type, insert a bitcast to the right 1333 // type. 1334 const Type *NewTy = 1335 PointerType::get(AI->getType()->getElementType(), AddrSpace); 1336 1337 if (OtherPtr->getType() != NewTy) 1338 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 1339 } 1340 1341 // Process each element of the aggregate. 1342 Value *TheFn = MI->getCalledValue(); 1343 const Type *BytePtrTy = MI->getRawDest()->getType(); 1344 bool SROADest = MI->getRawDest() == Inst; 1345 1346 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1347 1348 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1349 // If this is a memcpy/memmove, emit a GEP of the other element address. 1350 Value *OtherElt = 0; 1351 unsigned OtherEltAlign = MemAlignment; 1352 1353 if (OtherPtr) { 1354 Value *Idx[2] = { Zero, 1355 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1356 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1357 OtherPtr->getName()+"."+Twine(i), 1358 MI); 1359 uint64_t EltOffset; 1360 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1361 const Type *OtherTy = OtherPtrTy->getElementType(); 1362 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1363 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1364 } else { 1365 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 1366 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1367 } 1368 1369 // The alignment of the other pointer is the guaranteed alignment of the 1370 // element, which is affected by both the known alignment of the whole 1371 // mem intrinsic and the alignment of the element. If the alignment of 1372 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 1373 // known alignment is just 4 bytes. 1374 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 1375 } 1376 1377 Value *EltPtr = NewElts[i]; 1378 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 1379 1380 // If we got down to a scalar, insert a load or store as appropriate. 1381 if (EltTy->isSingleValueType()) { 1382 if (isa<MemTransferInst>(MI)) { 1383 if (SROADest) { 1384 // From Other to Alloca. 1385 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 1386 new StoreInst(Elt, EltPtr, MI); 1387 } else { 1388 // From Alloca to Other. 1389 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 1390 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 1391 } 1392 continue; 1393 } 1394 assert(isa<MemSetInst>(MI)); 1395 1396 // If the stored element is zero (common case), just store a null 1397 // constant. 1398 Constant *StoreVal; 1399 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 1400 if (CI->isZero()) { 1401 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 1402 } else { 1403 // If EltTy is a vector type, get the element type. 1404 const Type *ValTy = EltTy->getScalarType(); 1405 1406 // Construct an integer with the right value. 1407 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 1408 APInt OneVal(EltSize, CI->getZExtValue()); 1409 APInt TotalVal(OneVal); 1410 // Set each byte. 1411 for (unsigned i = 0; 8*i < EltSize; ++i) { 1412 TotalVal = TotalVal.shl(8); 1413 TotalVal |= OneVal; 1414 } 1415 1416 // Convert the integer value to the appropriate type. 1417 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 1418 if (ValTy->isPointerTy()) 1419 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 1420 else if (ValTy->isFloatingPointTy()) 1421 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 1422 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 1423 1424 // If the requested value was a vector constant, create it. 1425 if (EltTy != ValTy) { 1426 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 1427 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 1428 StoreVal = ConstantVector::get(&Elts[0], NumElts); 1429 } 1430 } 1431 new StoreInst(StoreVal, EltPtr, MI); 1432 continue; 1433 } 1434 // Otherwise, if we're storing a byte variable, use a memset call for 1435 // this element. 1436 } 1437 1438 // Cast the element pointer to BytePtrTy. 1439 if (EltPtr->getType() != BytePtrTy) 1440 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI); 1441 1442 // Cast the other pointer (if we have one) to BytePtrTy. 1443 if (OtherElt && OtherElt->getType() != BytePtrTy) { 1444 // Preserve address space of OtherElt 1445 const PointerType* OtherPTy = cast<PointerType>(OtherElt->getType()); 1446 const PointerType* PTy = cast<PointerType>(BytePtrTy); 1447 if (OtherPTy->getElementType() != PTy->getElementType()) { 1448 Type *NewOtherPTy = PointerType::get(PTy->getElementType(), 1449 OtherPTy->getAddressSpace()); 1450 OtherElt = new BitCastInst(OtherElt, NewOtherPTy, 1451 OtherElt->getName(), MI); 1452 } 1453 } 1454 1455 unsigned EltSize = TD->getTypeAllocSize(EltTy); 1456 1457 // Finally, insert the meminst for this element. 1458 if (isa<MemTransferInst>(MI)) { 1459 Value *Ops[] = { 1460 SROADest ? EltPtr : OtherElt, // Dest ptr 1461 SROADest ? OtherElt : EltPtr, // Src ptr 1462 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size 1463 // Align 1464 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign), 1465 MI->getVolatileCst() 1466 }; 1467 // In case we fold the address space overloaded memcpy of A to B 1468 // with memcpy of B to C, change the function to be a memcpy of A to C. 1469 const Type *Tys[] = { Ops[0]->getType(), Ops[1]->getType(), 1470 Ops[2]->getType() }; 1471 Module *M = MI->getParent()->getParent()->getParent(); 1472 TheFn = Intrinsic::getDeclaration(M, MI->getIntrinsicID(), Tys, 3); 1473 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1474 } else { 1475 assert(isa<MemSetInst>(MI)); 1476 Value *Ops[] = { 1477 EltPtr, MI->getArgOperand(1), // Dest, Value, 1478 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size 1479 Zero, // Align 1480 ConstantInt::getFalse(MI->getContext()) // isVolatile 1481 }; 1482 const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() }; 1483 Module *M = MI->getParent()->getParent()->getParent(); 1484 TheFn = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2); 1485 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1486 } 1487 } 1488 DeadInsts.push_back(MI); 1489} 1490 1491/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 1492/// overwrites the entire allocation. Extract out the pieces of the stored 1493/// integer and store them individually. 1494void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 1495 SmallVector<AllocaInst*, 32> &NewElts){ 1496 // Extract each element out of the integer according to its structure offset 1497 // and store the element value to the individual alloca. 1498 Value *SrcVal = SI->getOperand(0); 1499 const Type *AllocaEltTy = AI->getAllocatedType(); 1500 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1501 1502 // Handle tail padding by extending the operand 1503 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 1504 SrcVal = new ZExtInst(SrcVal, 1505 IntegerType::get(SI->getContext(), AllocaSizeBits), 1506 "", SI); 1507 1508 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 1509 << '\n'); 1510 1511 // There are two forms here: AI could be an array or struct. Both cases 1512 // have different ways to compute the element offset. 1513 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1514 const StructLayout *Layout = TD->getStructLayout(EltSTy); 1515 1516 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1517 // Get the number of bits to shift SrcVal to get the value. 1518 const Type *FieldTy = EltSTy->getElementType(i); 1519 uint64_t Shift = Layout->getElementOffsetInBits(i); 1520 1521 if (TD->isBigEndian()) 1522 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 1523 1524 Value *EltVal = SrcVal; 1525 if (Shift) { 1526 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1527 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1528 "sroa.store.elt", SI); 1529 } 1530 1531 // Truncate down to an integer of the right size. 1532 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1533 1534 // Ignore zero sized fields like {}, they obviously contain no data. 1535 if (FieldSizeBits == 0) continue; 1536 1537 if (FieldSizeBits != AllocaSizeBits) 1538 EltVal = new TruncInst(EltVal, 1539 IntegerType::get(SI->getContext(), FieldSizeBits), 1540 "", SI); 1541 Value *DestField = NewElts[i]; 1542 if (EltVal->getType() == FieldTy) { 1543 // Storing to an integer field of this size, just do it. 1544 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 1545 // Bitcast to the right element type (for fp/vector values). 1546 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 1547 } else { 1548 // Otherwise, bitcast the dest pointer (for aggregates). 1549 DestField = new BitCastInst(DestField, 1550 PointerType::getUnqual(EltVal->getType()), 1551 "", SI); 1552 } 1553 new StoreInst(EltVal, DestField, SI); 1554 } 1555 1556 } else { 1557 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1558 const Type *ArrayEltTy = ATy->getElementType(); 1559 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1560 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1561 1562 uint64_t Shift; 1563 1564 if (TD->isBigEndian()) 1565 Shift = AllocaSizeBits-ElementOffset; 1566 else 1567 Shift = 0; 1568 1569 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1570 // Ignore zero sized fields like {}, they obviously contain no data. 1571 if (ElementSizeBits == 0) continue; 1572 1573 Value *EltVal = SrcVal; 1574 if (Shift) { 1575 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1576 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1577 "sroa.store.elt", SI); 1578 } 1579 1580 // Truncate down to an integer of the right size. 1581 if (ElementSizeBits != AllocaSizeBits) 1582 EltVal = new TruncInst(EltVal, 1583 IntegerType::get(SI->getContext(), 1584 ElementSizeBits),"",SI); 1585 Value *DestField = NewElts[i]; 1586 if (EltVal->getType() == ArrayEltTy) { 1587 // Storing to an integer field of this size, just do it. 1588 } else if (ArrayEltTy->isFloatingPointTy() || 1589 ArrayEltTy->isVectorTy()) { 1590 // Bitcast to the right element type (for fp/vector values). 1591 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1592 } else { 1593 // Otherwise, bitcast the dest pointer (for aggregates). 1594 DestField = new BitCastInst(DestField, 1595 PointerType::getUnqual(EltVal->getType()), 1596 "", SI); 1597 } 1598 new StoreInst(EltVal, DestField, SI); 1599 1600 if (TD->isBigEndian()) 1601 Shift -= ElementOffset; 1602 else 1603 Shift += ElementOffset; 1604 } 1605 } 1606 1607 DeadInsts.push_back(SI); 1608} 1609 1610/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1611/// an integer. Load the individual pieces to form the aggregate value. 1612void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1613 SmallVector<AllocaInst*, 32> &NewElts) { 1614 // Extract each element out of the NewElts according to its structure offset 1615 // and form the result value. 1616 const Type *AllocaEltTy = AI->getAllocatedType(); 1617 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1618 1619 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1620 << '\n'); 1621 1622 // There are two forms here: AI could be an array or struct. Both cases 1623 // have different ways to compute the element offset. 1624 const StructLayout *Layout = 0; 1625 uint64_t ArrayEltBitOffset = 0; 1626 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1627 Layout = TD->getStructLayout(EltSTy); 1628 } else { 1629 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1630 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1631 } 1632 1633 Value *ResultVal = 1634 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1635 1636 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1637 // Load the value from the alloca. If the NewElt is an aggregate, cast 1638 // the pointer to an integer of the same size before doing the load. 1639 Value *SrcField = NewElts[i]; 1640 const Type *FieldTy = 1641 cast<PointerType>(SrcField->getType())->getElementType(); 1642 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1643 1644 // Ignore zero sized fields like {}, they obviously contain no data. 1645 if (FieldSizeBits == 0) continue; 1646 1647 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1648 FieldSizeBits); 1649 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1650 !FieldTy->isVectorTy()) 1651 SrcField = new BitCastInst(SrcField, 1652 PointerType::getUnqual(FieldIntTy), 1653 "", LI); 1654 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1655 1656 // If SrcField is a fp or vector of the right size but that isn't an 1657 // integer type, bitcast to an integer so we can shift it. 1658 if (SrcField->getType() != FieldIntTy) 1659 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1660 1661 // Zero extend the field to be the same size as the final alloca so that 1662 // we can shift and insert it. 1663 if (SrcField->getType() != ResultVal->getType()) 1664 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1665 1666 // Determine the number of bits to shift SrcField. 1667 uint64_t Shift; 1668 if (Layout) // Struct case. 1669 Shift = Layout->getElementOffsetInBits(i); 1670 else // Array case. 1671 Shift = i*ArrayEltBitOffset; 1672 1673 if (TD->isBigEndian()) 1674 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1675 1676 if (Shift) { 1677 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1678 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1679 } 1680 1681 // Don't create an 'or x, 0' on the first iteration. 1682 if (!isa<Constant>(ResultVal) || 1683 !cast<Constant>(ResultVal)->isNullValue()) 1684 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1685 else 1686 ResultVal = SrcField; 1687 } 1688 1689 // Handle tail padding by truncating the result 1690 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1691 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1692 1693 LI->replaceAllUsesWith(ResultVal); 1694 DeadInsts.push_back(LI); 1695} 1696 1697/// HasPadding - Return true if the specified type has any structure or 1698/// alignment padding, false otherwise. 1699static bool HasPadding(const Type *Ty, const TargetData &TD) { 1700 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 1701 return HasPadding(ATy->getElementType(), TD); 1702 1703 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) 1704 return HasPadding(VTy->getElementType(), TD); 1705 1706 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1707 const StructLayout *SL = TD.getStructLayout(STy); 1708 unsigned PrevFieldBitOffset = 0; 1709 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1710 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1711 1712 // Padding in sub-elements? 1713 if (HasPadding(STy->getElementType(i), TD)) 1714 return true; 1715 1716 // Check to see if there is any padding between this element and the 1717 // previous one. 1718 if (i) { 1719 unsigned PrevFieldEnd = 1720 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1721 if (PrevFieldEnd < FieldBitOffset) 1722 return true; 1723 } 1724 1725 PrevFieldBitOffset = FieldBitOffset; 1726 } 1727 1728 // Check for tail padding. 1729 if (unsigned EltCount = STy->getNumElements()) { 1730 unsigned PrevFieldEnd = PrevFieldBitOffset + 1731 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1732 if (PrevFieldEnd < SL->getSizeInBits()) 1733 return true; 1734 } 1735 } 1736 1737 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1738} 1739 1740/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1741/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1742/// or 1 if safe after canonicalization has been performed. 1743bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1744 // Loop over the use list of the alloca. We can only transform it if all of 1745 // the users are safe to transform. 1746 AllocaInfo Info; 1747 1748 isSafeForScalarRepl(AI, AI, 0, Info); 1749 if (Info.isUnsafe) { 1750 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1751 return false; 1752 } 1753 1754 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1755 // source and destination, we have to be careful. In particular, the memcpy 1756 // could be moving around elements that live in structure padding of the LLVM 1757 // types, but may actually be used. In these cases, we refuse to promote the 1758 // struct. 1759 if (Info.isMemCpySrc && Info.isMemCpyDst && 1760 HasPadding(AI->getAllocatedType(), *TD)) 1761 return false; 1762 1763 return true; 1764} 1765 1766 1767 1768/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1769/// some part of a constant global variable. This intentionally only accepts 1770/// constant expressions because we don't can't rewrite arbitrary instructions. 1771static bool PointsToConstantGlobal(Value *V) { 1772 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1773 return GV->isConstant(); 1774 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1775 if (CE->getOpcode() == Instruction::BitCast || 1776 CE->getOpcode() == Instruction::GetElementPtr) 1777 return PointsToConstantGlobal(CE->getOperand(0)); 1778 return false; 1779} 1780 1781/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1782/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1783/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1784/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1785/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1786/// the alloca, and if the source pointer is a pointer to a constant global, we 1787/// can optimize this. 1788static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 1789 bool isOffset) { 1790 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1791 User *U = cast<Instruction>(*UI); 1792 1793 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1794 // Ignore non-volatile loads, they are always ok. 1795 if (LI->isVolatile()) return false; 1796 continue; 1797 } 1798 1799 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1800 // If uses of the bitcast are ok, we are ok. 1801 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1802 return false; 1803 continue; 1804 } 1805 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 1806 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1807 // doesn't, it does. 1808 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1809 isOffset || !GEP->hasAllZeroIndices())) 1810 return false; 1811 continue; 1812 } 1813 1814 if (CallSite CS = U) { 1815 // If this is a readonly/readnone call site, then we know it is just a 1816 // load and we can ignore it. 1817 if (CS.onlyReadsMemory()) 1818 continue; 1819 1820 // If this is being passed as a byval argument, the caller is making a 1821 // copy, so it is only a read of the alloca. 1822 unsigned ArgNo = CS.getArgumentNo(UI); 1823 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 1824 continue; 1825 } 1826 1827 // If this is isn't our memcpy/memmove, reject it as something we can't 1828 // handle. 1829 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 1830 if (MI == 0) 1831 return false; 1832 1833 // If the transfer is using the alloca as a source of the transfer, then 1834 // ignore it since it is a load (unless the transfer is volatile). 1835 if (UI.getOperandNo() == 1) { 1836 if (MI->isVolatile()) return false; 1837 continue; 1838 } 1839 1840 // If we already have seen a copy, reject the second one. 1841 if (TheCopy) return false; 1842 1843 // If the pointer has been offset from the start of the alloca, we can't 1844 // safely handle this. 1845 if (isOffset) return false; 1846 1847 // If the memintrinsic isn't using the alloca as the dest, reject it. 1848 if (UI.getOperandNo() != 0) return false; 1849 1850 // If the source of the memcpy/move is not a constant global, reject it. 1851 if (!PointsToConstantGlobal(MI->getSource())) 1852 return false; 1853 1854 // Otherwise, the transform is safe. Remember the copy instruction. 1855 TheCopy = MI; 1856 } 1857 return true; 1858} 1859 1860/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1861/// modified by a copy from a constant global. If we can prove this, we can 1862/// replace any uses of the alloca with uses of the global directly. 1863MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1864 MemTransferInst *TheCopy = 0; 1865 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1866 return TheCopy; 1867 return 0; 1868} 1869