ScalarReplAggregates.cpp revision 396a0567cf959d86a8a1ad185e54d84f5dacbacf
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/DominanceFrontier.h" 34#include "llvm/Analysis/ValueTracking.h" 35#include "llvm/Target/TargetData.h" 36#include "llvm/Transforms/Utils/PromoteMemToReg.h" 37#include "llvm/Transforms/Utils/Local.h" 38#include "llvm/Transforms/Utils/SSAUpdater.h" 39#include "llvm/Support/CallSite.h" 40#include "llvm/Support/Debug.h" 41#include "llvm/Support/ErrorHandling.h" 42#include "llvm/Support/GetElementPtrTypeIterator.h" 43#include "llvm/Support/IRBuilder.h" 44#include "llvm/Support/MathExtras.h" 45#include "llvm/Support/raw_ostream.h" 46#include "llvm/ADT/SmallVector.h" 47#include "llvm/ADT/Statistic.h" 48using namespace llvm; 49 50STATISTIC(NumReplaced, "Number of allocas broken up"); 51STATISTIC(NumPromoted, "Number of allocas promoted"); 52STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 53STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 54 55namespace { 56 struct SROA : public FunctionPass { 57 SROA(int T, bool hasDF, char &ID) 58 : FunctionPass(ID), HasDomFrontiers(hasDF) { 59 if (T == -1) 60 SRThreshold = 128; 61 else 62 SRThreshold = T; 63 } 64 65 bool runOnFunction(Function &F); 66 67 bool performScalarRepl(Function &F); 68 bool performPromotion(Function &F); 69 70 private: 71 bool HasDomFrontiers; 72 TargetData *TD; 73 74 /// DeadInsts - Keep track of instructions we have made dead, so that 75 /// we can remove them after we are done working. 76 SmallVector<Value*, 32> DeadInsts; 77 78 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 79 /// information about the uses. All these fields are initialized to false 80 /// and set to true when something is learned. 81 struct AllocaInfo { 82 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 83 bool isUnsafe : 1; 84 85 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 86 bool isMemCpySrc : 1; 87 88 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 89 bool isMemCpyDst : 1; 90 91 /// hasSubelementAccess - This is true if a subelement of the alloca is 92 /// ever accessed, or false if the alloca is only accessed with mem 93 /// intrinsics or load/store that only access the entire alloca at once. 94 bool hasSubelementAccess : 1; 95 96 /// hasALoadOrStore - This is true if there are any loads or stores to it. 97 /// The alloca may just be accessed with memcpy, for example, which would 98 /// not set this. 99 bool hasALoadOrStore : 1; 100 101 AllocaInfo() 102 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 103 hasSubelementAccess(false), hasALoadOrStore(false) {} 104 }; 105 106 unsigned SRThreshold; 107 108 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 109 110 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 111 112 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 113 AllocaInfo &Info); 114 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset, 115 AllocaInfo &Info); 116 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 117 const Type *MemOpType, bool isStore, AllocaInfo &Info); 118 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 119 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 120 const Type *&IdxTy); 121 122 void DoScalarReplacement(AllocaInst *AI, 123 std::vector<AllocaInst*> &WorkList); 124 void DeleteDeadInstructions(); 125 126 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 127 SmallVector<AllocaInst*, 32> &NewElts); 128 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 129 SmallVector<AllocaInst*, 32> &NewElts); 130 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 131 SmallVector<AllocaInst*, 32> &NewElts); 132 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 133 AllocaInst *AI, 134 SmallVector<AllocaInst*, 32> &NewElts); 135 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 136 SmallVector<AllocaInst*, 32> &NewElts); 137 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 138 SmallVector<AllocaInst*, 32> &NewElts); 139 140 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 141 }; 142 143 // SROA_DF - SROA that uses DominanceFrontier. 144 struct SROA_DF : public SROA { 145 static char ID; 146 public: 147 SROA_DF(int T = -1) : SROA(T, true, ID) { 148 initializeSROA_DFPass(*PassRegistry::getPassRegistry()); 149 } 150 151 // getAnalysisUsage - This pass does not require any passes, but we know it 152 // will not alter the CFG, so say so. 153 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 154 AU.addRequired<DominatorTree>(); 155 AU.addRequired<DominanceFrontier>(); 156 AU.setPreservesCFG(); 157 } 158 }; 159 160 // SROA_SSAUp - SROA that uses SSAUpdater. 161 struct SROA_SSAUp : public SROA { 162 static char ID; 163 public: 164 SROA_SSAUp(int T = -1) : SROA(T, false, ID) { 165 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 166 } 167 168 // getAnalysisUsage - This pass does not require any passes, but we know it 169 // will not alter the CFG, so say so. 170 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 171 AU.setPreservesCFG(); 172 } 173 }; 174 175} 176 177char SROA_DF::ID = 0; 178char SROA_SSAUp::ID = 0; 179 180INITIALIZE_PASS_BEGIN(SROA_DF, "scalarrepl", 181 "Scalar Replacement of Aggregates (DF)", false, false) 182INITIALIZE_PASS_DEPENDENCY(DominatorTree) 183INITIALIZE_PASS_DEPENDENCY(DominanceFrontier) 184INITIALIZE_PASS_END(SROA_DF, "scalarrepl", 185 "Scalar Replacement of Aggregates (DF)", false, false) 186 187INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 188 "Scalar Replacement of Aggregates (SSAUp)", false, false) 189INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 190 "Scalar Replacement of Aggregates (SSAUp)", false, false) 191 192// Public interface to the ScalarReplAggregates pass 193FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 194 bool UseDomFrontier) { 195 if (UseDomFrontier) 196 return new SROA_DF(Threshold); 197 return new SROA_SSAUp(Threshold); 198} 199 200 201//===----------------------------------------------------------------------===// 202// Convert To Scalar Optimization. 203//===----------------------------------------------------------------------===// 204 205namespace { 206/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 207/// optimization, which scans the uses of an alloca and determines if it can 208/// rewrite it in terms of a single new alloca that can be mem2reg'd. 209class ConvertToScalarInfo { 210 /// AllocaSize - The size of the alloca being considered. 211 unsigned AllocaSize; 212 const TargetData &TD; 213 214 /// IsNotTrivial - This is set to true if there is some access to the object 215 /// which means that mem2reg can't promote it. 216 bool IsNotTrivial; 217 218 /// VectorTy - This tracks the type that we should promote the vector to if 219 /// it is possible to turn it into a vector. This starts out null, and if it 220 /// isn't possible to turn into a vector type, it gets set to VoidTy. 221 const Type *VectorTy; 222 223 /// HadAVector - True if there is at least one vector access to the alloca. 224 /// We don't want to turn random arrays into vectors and use vector element 225 /// insert/extract, but if there are element accesses to something that is 226 /// also declared as a vector, we do want to promote to a vector. 227 bool HadAVector; 228 229public: 230 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 231 : AllocaSize(Size), TD(td) { 232 IsNotTrivial = false; 233 VectorTy = 0; 234 HadAVector = false; 235 } 236 237 AllocaInst *TryConvert(AllocaInst *AI); 238 239private: 240 bool CanConvertToScalar(Value *V, uint64_t Offset); 241 void MergeInType(const Type *In, uint64_t Offset); 242 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 243 244 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 245 uint64_t Offset, IRBuilder<> &Builder); 246 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 247 uint64_t Offset, IRBuilder<> &Builder); 248}; 249} // end anonymous namespace. 250 251 252/// IsVerbotenVectorType - Return true if this is a vector type ScalarRepl isn't 253/// allowed to form. We do this to avoid MMX types, which is a complete hack, 254/// but is required until the backend is fixed. 255static bool IsVerbotenVectorType(const VectorType *VTy, const Instruction *I) { 256 StringRef Triple(I->getParent()->getParent()->getParent()->getTargetTriple()); 257 if (!Triple.startswith("i386") && 258 !Triple.startswith("x86_64")) 259 return false; 260 261 // Reject all the MMX vector types. 262 switch (VTy->getNumElements()) { 263 default: return false; 264 case 1: return VTy->getElementType()->isIntegerTy(64); 265 case 2: return VTy->getElementType()->isIntegerTy(32); 266 case 4: return VTy->getElementType()->isIntegerTy(16); 267 case 8: return VTy->getElementType()->isIntegerTy(8); 268 } 269} 270 271 272/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 273/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 274/// alloca if possible or null if not. 275AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 276 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 277 // out. 278 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 279 return 0; 280 281 // If we were able to find a vector type that can handle this with 282 // insert/extract elements, and if there was at least one use that had 283 // a vector type, promote this to a vector. We don't want to promote 284 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 285 // we just get a lot of insert/extracts. If at least one vector is 286 // involved, then we probably really do have a union of vector/array. 287 const Type *NewTy; 288 if (VectorTy && VectorTy->isVectorTy() && HadAVector && 289 !IsVerbotenVectorType(cast<VectorType>(VectorTy), AI)) { 290 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 291 << *VectorTy << '\n'); 292 NewTy = VectorTy; // Use the vector type. 293 } else { 294 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 295 // Create and insert the integer alloca. 296 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 297 } 298 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 299 ConvertUsesToScalar(AI, NewAI, 0); 300 return NewAI; 301} 302 303/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy) 304/// so far at the offset specified by Offset (which is specified in bytes). 305/// 306/// There are two cases we handle here: 307/// 1) A union of vector types of the same size and potentially its elements. 308/// Here we turn element accesses into insert/extract element operations. 309/// This promotes a <4 x float> with a store of float to the third element 310/// into a <4 x float> that uses insert element. 311/// 2) A fully general blob of memory, which we turn into some (potentially 312/// large) integer type with extract and insert operations where the loads 313/// and stores would mutate the memory. We mark this by setting VectorTy 314/// to VoidTy. 315void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) { 316 // If we already decided to turn this into a blob of integer memory, there is 317 // nothing to be done. 318 if (VectorTy && VectorTy->isVoidTy()) 319 return; 320 321 // If this could be contributing to a vector, analyze it. 322 323 // If the In type is a vector that is the same size as the alloca, see if it 324 // matches the existing VecTy. 325 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 326 // Remember if we saw a vector type. 327 HadAVector = true; 328 329 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 330 // If we're storing/loading a vector of the right size, allow it as a 331 // vector. If this the first vector we see, remember the type so that 332 // we know the element size. If this is a subsequent access, ignore it 333 // even if it is a differing type but the same size. Worst case we can 334 // bitcast the resultant vectors. 335 if (VectorTy == 0) 336 VectorTy = VInTy; 337 return; 338 } 339 } else if (In->isFloatTy() || In->isDoubleTy() || 340 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 341 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 342 // If we're accessing something that could be an element of a vector, see 343 // if the implied vector agrees with what we already have and if Offset is 344 // compatible with it. 345 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 346 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 347 (VectorTy == 0 || 348 cast<VectorType>(VectorTy)->getElementType() 349 ->getPrimitiveSizeInBits()/8 == EltSize)) { 350 if (VectorTy == 0) 351 VectorTy = VectorType::get(In, AllocaSize/EltSize); 352 return; 353 } 354 } 355 356 // Otherwise, we have a case that we can't handle with an optimized vector 357 // form. We can still turn this into a large integer. 358 VectorTy = Type::getVoidTy(In->getContext()); 359} 360 361/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 362/// its accesses to a single vector type, return true and set VecTy to 363/// the new type. If we could convert the alloca into a single promotable 364/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 365/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 366/// is the current offset from the base of the alloca being analyzed. 367/// 368/// If we see at least one access to the value that is as a vector type, set the 369/// SawVec flag. 370bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 371 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 372 Instruction *User = cast<Instruction>(*UI); 373 374 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 375 // Don't break volatile loads. 376 if (LI->isVolatile()) 377 return false; 378 // Don't touch MMX operations. 379 if (LI->getType()->isX86_MMXTy()) 380 return false; 381 MergeInType(LI->getType(), Offset); 382 continue; 383 } 384 385 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 386 // Storing the pointer, not into the value? 387 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 388 // Don't touch MMX operations. 389 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 390 return false; 391 MergeInType(SI->getOperand(0)->getType(), Offset); 392 continue; 393 } 394 395 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 396 IsNotTrivial = true; // Can't be mem2reg'd. 397 if (!CanConvertToScalar(BCI, Offset)) 398 return false; 399 continue; 400 } 401 402 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 403 // If this is a GEP with a variable indices, we can't handle it. 404 if (!GEP->hasAllConstantIndices()) 405 return false; 406 407 // Compute the offset that this GEP adds to the pointer. 408 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 409 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 410 &Indices[0], Indices.size()); 411 // See if all uses can be converted. 412 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 413 return false; 414 IsNotTrivial = true; // Can't be mem2reg'd. 415 continue; 416 } 417 418 // If this is a constant sized memset of a constant value (e.g. 0) we can 419 // handle it. 420 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 421 // Store of constant value and constant size. 422 if (!isa<ConstantInt>(MSI->getValue()) || 423 !isa<ConstantInt>(MSI->getLength())) 424 return false; 425 IsNotTrivial = true; // Can't be mem2reg'd. 426 continue; 427 } 428 429 // If this is a memcpy or memmove into or out of the whole allocation, we 430 // can handle it like a load or store of the scalar type. 431 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 432 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 433 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 434 return false; 435 436 IsNotTrivial = true; // Can't be mem2reg'd. 437 continue; 438 } 439 440 // Otherwise, we cannot handle this! 441 return false; 442 } 443 444 return true; 445} 446 447/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 448/// directly. This happens when we are converting an "integer union" to a 449/// single integer scalar, or when we are converting a "vector union" to a 450/// vector with insert/extractelement instructions. 451/// 452/// Offset is an offset from the original alloca, in bits that need to be 453/// shifted to the right. By the end of this, there should be no uses of Ptr. 454void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 455 uint64_t Offset) { 456 while (!Ptr->use_empty()) { 457 Instruction *User = cast<Instruction>(Ptr->use_back()); 458 459 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 460 ConvertUsesToScalar(CI, NewAI, Offset); 461 CI->eraseFromParent(); 462 continue; 463 } 464 465 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 466 // Compute the offset that this GEP adds to the pointer. 467 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 468 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 469 &Indices[0], Indices.size()); 470 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 471 GEP->eraseFromParent(); 472 continue; 473 } 474 475 IRBuilder<> Builder(User); 476 477 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 478 // The load is a bit extract from NewAI shifted right by Offset bits. 479 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 480 Value *NewLoadVal 481 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 482 LI->replaceAllUsesWith(NewLoadVal); 483 LI->eraseFromParent(); 484 continue; 485 } 486 487 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 488 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 489 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 490 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 491 Builder); 492 Builder.CreateStore(New, NewAI); 493 SI->eraseFromParent(); 494 495 // If the load we just inserted is now dead, then the inserted store 496 // overwrote the entire thing. 497 if (Old->use_empty()) 498 Old->eraseFromParent(); 499 continue; 500 } 501 502 // If this is a constant sized memset of a constant value (e.g. 0) we can 503 // transform it into a store of the expanded constant value. 504 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 505 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 506 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 507 if (NumBytes != 0) { 508 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 509 510 // Compute the value replicated the right number of times. 511 APInt APVal(NumBytes*8, Val); 512 513 // Splat the value if non-zero. 514 if (Val) 515 for (unsigned i = 1; i != NumBytes; ++i) 516 APVal |= APVal << 8; 517 518 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 519 Value *New = ConvertScalar_InsertValue( 520 ConstantInt::get(User->getContext(), APVal), 521 Old, Offset, Builder); 522 Builder.CreateStore(New, NewAI); 523 524 // If the load we just inserted is now dead, then the memset overwrote 525 // the entire thing. 526 if (Old->use_empty()) 527 Old->eraseFromParent(); 528 } 529 MSI->eraseFromParent(); 530 continue; 531 } 532 533 // If this is a memcpy or memmove into or out of the whole allocation, we 534 // can handle it like a load or store of the scalar type. 535 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 536 assert(Offset == 0 && "must be store to start of alloca"); 537 538 // If the source and destination are both to the same alloca, then this is 539 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 540 // as appropriate. 541 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, 0)); 542 543 if (GetUnderlyingObject(MTI->getSource(), 0) != OrigAI) { 544 // Dest must be OrigAI, change this to be a load from the original 545 // pointer (bitcasted), then a store to our new alloca. 546 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 547 Value *SrcPtr = MTI->getSource(); 548 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 549 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 550 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 551 AIPTy = PointerType::get(AIPTy->getElementType(), 552 SPTy->getAddressSpace()); 553 } 554 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 555 556 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 557 SrcVal->setAlignment(MTI->getAlignment()); 558 Builder.CreateStore(SrcVal, NewAI); 559 } else if (GetUnderlyingObject(MTI->getDest(), 0) != OrigAI) { 560 // Src must be OrigAI, change this to be a load from NewAI then a store 561 // through the original dest pointer (bitcasted). 562 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 563 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 564 565 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 566 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 567 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 568 AIPTy = PointerType::get(AIPTy->getElementType(), 569 DPTy->getAddressSpace()); 570 } 571 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 572 573 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 574 NewStore->setAlignment(MTI->getAlignment()); 575 } else { 576 // Noop transfer. Src == Dst 577 } 578 579 MTI->eraseFromParent(); 580 continue; 581 } 582 583 llvm_unreachable("Unsupported operation!"); 584 } 585} 586 587/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 588/// or vector value FromVal, extracting the bits from the offset specified by 589/// Offset. This returns the value, which is of type ToType. 590/// 591/// This happens when we are converting an "integer union" to a single 592/// integer scalar, or when we are converting a "vector union" to a vector with 593/// insert/extractelement instructions. 594/// 595/// Offset is an offset from the original alloca, in bits that need to be 596/// shifted to the right. 597Value *ConvertToScalarInfo:: 598ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 599 uint64_t Offset, IRBuilder<> &Builder) { 600 // If the load is of the whole new alloca, no conversion is needed. 601 if (FromVal->getType() == ToType && Offset == 0) 602 return FromVal; 603 604 // If the result alloca is a vector type, this is either an element 605 // access or a bitcast to another vector type of the same size. 606 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 607 if (ToType->isVectorTy()) 608 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 609 610 // Otherwise it must be an element access. 611 unsigned Elt = 0; 612 if (Offset) { 613 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 614 Elt = Offset/EltSize; 615 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 616 } 617 // Return the element extracted out of it. 618 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 619 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 620 if (V->getType() != ToType) 621 V = Builder.CreateBitCast(V, ToType, "tmp"); 622 return V; 623 } 624 625 // If ToType is a first class aggregate, extract out each of the pieces and 626 // use insertvalue's to form the FCA. 627 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 628 const StructLayout &Layout = *TD.getStructLayout(ST); 629 Value *Res = UndefValue::get(ST); 630 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 631 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 632 Offset+Layout.getElementOffsetInBits(i), 633 Builder); 634 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 635 } 636 return Res; 637 } 638 639 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 640 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 641 Value *Res = UndefValue::get(AT); 642 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 643 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 644 Offset+i*EltSize, Builder); 645 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 646 } 647 return Res; 648 } 649 650 // Otherwise, this must be a union that was converted to an integer value. 651 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 652 653 // If this is a big-endian system and the load is narrower than the 654 // full alloca type, we need to do a shift to get the right bits. 655 int ShAmt = 0; 656 if (TD.isBigEndian()) { 657 // On big-endian machines, the lowest bit is stored at the bit offset 658 // from the pointer given by getTypeStoreSizeInBits. This matters for 659 // integers with a bitwidth that is not a multiple of 8. 660 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 661 TD.getTypeStoreSizeInBits(ToType) - Offset; 662 } else { 663 ShAmt = Offset; 664 } 665 666 // Note: we support negative bitwidths (with shl) which are not defined. 667 // We do this to support (f.e.) loads off the end of a structure where 668 // only some bits are used. 669 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 670 FromVal = Builder.CreateLShr(FromVal, 671 ConstantInt::get(FromVal->getType(), 672 ShAmt), "tmp"); 673 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 674 FromVal = Builder.CreateShl(FromVal, 675 ConstantInt::get(FromVal->getType(), 676 -ShAmt), "tmp"); 677 678 // Finally, unconditionally truncate the integer to the right width. 679 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 680 if (LIBitWidth < NTy->getBitWidth()) 681 FromVal = 682 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 683 LIBitWidth), "tmp"); 684 else if (LIBitWidth > NTy->getBitWidth()) 685 FromVal = 686 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 687 LIBitWidth), "tmp"); 688 689 // If the result is an integer, this is a trunc or bitcast. 690 if (ToType->isIntegerTy()) { 691 // Should be done. 692 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 693 // Just do a bitcast, we know the sizes match up. 694 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 695 } else { 696 // Otherwise must be a pointer. 697 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 698 } 699 assert(FromVal->getType() == ToType && "Didn't convert right?"); 700 return FromVal; 701} 702 703/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 704/// or vector value "Old" at the offset specified by Offset. 705/// 706/// This happens when we are converting an "integer union" to a 707/// single integer scalar, or when we are converting a "vector union" to a 708/// vector with insert/extractelement instructions. 709/// 710/// Offset is an offset from the original alloca, in bits that need to be 711/// shifted to the right. 712Value *ConvertToScalarInfo:: 713ConvertScalar_InsertValue(Value *SV, Value *Old, 714 uint64_t Offset, IRBuilder<> &Builder) { 715 // Convert the stored type to the actual type, shift it left to insert 716 // then 'or' into place. 717 const Type *AllocaType = Old->getType(); 718 LLVMContext &Context = Old->getContext(); 719 720 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 721 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 722 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 723 724 // Changing the whole vector with memset or with an access of a different 725 // vector type? 726 if (ValSize == VecSize) 727 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 728 729 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 730 731 // Must be an element insertion. 732 unsigned Elt = Offset/EltSize; 733 734 if (SV->getType() != VTy->getElementType()) 735 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 736 737 SV = Builder.CreateInsertElement(Old, SV, 738 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 739 "tmp"); 740 return SV; 741 } 742 743 // If SV is a first-class aggregate value, insert each value recursively. 744 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 745 const StructLayout &Layout = *TD.getStructLayout(ST); 746 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 747 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 748 Old = ConvertScalar_InsertValue(Elt, Old, 749 Offset+Layout.getElementOffsetInBits(i), 750 Builder); 751 } 752 return Old; 753 } 754 755 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 756 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 757 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 758 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 759 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 760 } 761 return Old; 762 } 763 764 // If SV is a float, convert it to the appropriate integer type. 765 // If it is a pointer, do the same. 766 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 767 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 768 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 769 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 770 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 771 SV = Builder.CreateBitCast(SV, 772 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 773 else if (SV->getType()->isPointerTy()) 774 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 775 776 // Zero extend or truncate the value if needed. 777 if (SV->getType() != AllocaType) { 778 if (SV->getType()->getPrimitiveSizeInBits() < 779 AllocaType->getPrimitiveSizeInBits()) 780 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 781 else { 782 // Truncation may be needed if storing more than the alloca can hold 783 // (undefined behavior). 784 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 785 SrcWidth = DestWidth; 786 SrcStoreWidth = DestStoreWidth; 787 } 788 } 789 790 // If this is a big-endian system and the store is narrower than the 791 // full alloca type, we need to do a shift to get the right bits. 792 int ShAmt = 0; 793 if (TD.isBigEndian()) { 794 // On big-endian machines, the lowest bit is stored at the bit offset 795 // from the pointer given by getTypeStoreSizeInBits. This matters for 796 // integers with a bitwidth that is not a multiple of 8. 797 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 798 } else { 799 ShAmt = Offset; 800 } 801 802 // Note: we support negative bitwidths (with shr) which are not defined. 803 // We do this to support (f.e.) stores off the end of a structure where 804 // only some bits in the structure are set. 805 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 806 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 807 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 808 ShAmt), "tmp"); 809 Mask <<= ShAmt; 810 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 811 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 812 -ShAmt), "tmp"); 813 Mask = Mask.lshr(-ShAmt); 814 } 815 816 // Mask out the bits we are about to insert from the old value, and or 817 // in the new bits. 818 if (SrcWidth != DestWidth) { 819 assert(DestWidth > SrcWidth); 820 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 821 SV = Builder.CreateOr(Old, SV, "ins"); 822 } 823 return SV; 824} 825 826 827//===----------------------------------------------------------------------===// 828// SRoA Driver 829//===----------------------------------------------------------------------===// 830 831 832bool SROA::runOnFunction(Function &F) { 833 TD = getAnalysisIfAvailable<TargetData>(); 834 835 bool Changed = performPromotion(F); 836 837 // FIXME: ScalarRepl currently depends on TargetData more than it 838 // theoretically needs to. It should be refactored in order to support 839 // target-independent IR. Until this is done, just skip the actual 840 // scalar-replacement portion of this pass. 841 if (!TD) return Changed; 842 843 while (1) { 844 bool LocalChange = performScalarRepl(F); 845 if (!LocalChange) break; // No need to repromote if no scalarrepl 846 Changed = true; 847 LocalChange = performPromotion(F); 848 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 849 } 850 851 return Changed; 852} 853 854namespace { 855class AllocaPromoter : public LoadAndStorePromoter { 856 AllocaInst *AI; 857public: 858 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S) 859 : LoadAndStorePromoter(Insts, S), AI(0) {} 860 861 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 862 // Remember which alloca we're promoting (for isInstInList). 863 this->AI = AI; 864 LoadAndStorePromoter::run(Insts); 865 AI->eraseFromParent(); 866 } 867 868 virtual bool isInstInList(Instruction *I, 869 const SmallVectorImpl<Instruction*> &Insts) const { 870 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 871 return LI->getOperand(0) == AI; 872 return cast<StoreInst>(I)->getPointerOperand() == AI; 873 } 874}; 875} // end anon namespace 876 877bool SROA::performPromotion(Function &F) { 878 std::vector<AllocaInst*> Allocas; 879 DominatorTree *DT = 0; 880 DominanceFrontier *DF = 0; 881 if (HasDomFrontiers) { 882 DT = &getAnalysis<DominatorTree>(); 883 DF = &getAnalysis<DominanceFrontier>(); 884 } 885 886 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 887 888 bool Changed = false; 889 SmallVector<Instruction*, 64> Insts; 890 while (1) { 891 Allocas.clear(); 892 893 // Find allocas that are safe to promote, by looking at all instructions in 894 // the entry node 895 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 896 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 897 if (isAllocaPromotable(AI)) 898 Allocas.push_back(AI); 899 900 if (Allocas.empty()) break; 901 902 if (HasDomFrontiers) 903 PromoteMemToReg(Allocas, *DT, *DF); 904 else { 905 SSAUpdater SSA; 906 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 907 AllocaInst *AI = Allocas[i]; 908 909 // Build list of instructions to promote. 910 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 911 UI != E; ++UI) 912 Insts.push_back(cast<Instruction>(*UI)); 913 914 AllocaPromoter(Insts, SSA).run(AI, Insts); 915 Insts.clear(); 916 } 917 } 918 NumPromoted += Allocas.size(); 919 Changed = true; 920 } 921 922 return Changed; 923} 924 925 926/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 927/// SROA. It must be a struct or array type with a small number of elements. 928static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 929 const Type *T = AI->getAllocatedType(); 930 // Do not promote any struct into more than 32 separate vars. 931 if (const StructType *ST = dyn_cast<StructType>(T)) 932 return ST->getNumElements() <= 32; 933 // Arrays are much less likely to be safe for SROA; only consider 934 // them if they are very small. 935 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 936 return AT->getNumElements() <= 8; 937 return false; 938} 939 940 941// performScalarRepl - This algorithm is a simple worklist driven algorithm, 942// which runs on all of the malloc/alloca instructions in the function, removing 943// them if they are only used by getelementptr instructions. 944// 945bool SROA::performScalarRepl(Function &F) { 946 std::vector<AllocaInst*> WorkList; 947 948 // Scan the entry basic block, adding allocas to the worklist. 949 BasicBlock &BB = F.getEntryBlock(); 950 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 951 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 952 WorkList.push_back(A); 953 954 // Process the worklist 955 bool Changed = false; 956 while (!WorkList.empty()) { 957 AllocaInst *AI = WorkList.back(); 958 WorkList.pop_back(); 959 960 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 961 // with unused elements. 962 if (AI->use_empty()) { 963 AI->eraseFromParent(); 964 Changed = true; 965 continue; 966 } 967 968 // If this alloca is impossible for us to promote, reject it early. 969 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 970 continue; 971 972 // Check to see if this allocation is only modified by a memcpy/memmove from 973 // a constant global. If this is the case, we can change all users to use 974 // the constant global instead. This is commonly produced by the CFE by 975 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 976 // is only subsequently read. 977 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 978 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 979 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 980 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 981 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 982 TheCopy->eraseFromParent(); // Don't mutate the global. 983 AI->eraseFromParent(); 984 ++NumGlobals; 985 Changed = true; 986 continue; 987 } 988 989 // Check to see if we can perform the core SROA transformation. We cannot 990 // transform the allocation instruction if it is an array allocation 991 // (allocations OF arrays are ok though), and an allocation of a scalar 992 // value cannot be decomposed at all. 993 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 994 995 // Do not promote [0 x %struct]. 996 if (AllocaSize == 0) continue; 997 998 // Do not promote any struct whose size is too big. 999 if (AllocaSize > SRThreshold) continue; 1000 1001 // If the alloca looks like a good candidate for scalar replacement, and if 1002 // all its users can be transformed, then split up the aggregate into its 1003 // separate elements. 1004 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1005 DoScalarReplacement(AI, WorkList); 1006 Changed = true; 1007 continue; 1008 } 1009 1010 // If we can turn this aggregate value (potentially with casts) into a 1011 // simple scalar value that can be mem2reg'd into a register value. 1012 // IsNotTrivial tracks whether this is something that mem2reg could have 1013 // promoted itself. If so, we don't want to transform it needlessly. Note 1014 // that we can't just check based on the type: the alloca may be of an i32 1015 // but that has pointer arithmetic to set byte 3 of it or something. 1016 if (AllocaInst *NewAI = 1017 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1018 NewAI->takeName(AI); 1019 AI->eraseFromParent(); 1020 ++NumConverted; 1021 Changed = true; 1022 continue; 1023 } 1024 1025 // Otherwise, couldn't process this alloca. 1026 } 1027 1028 return Changed; 1029} 1030 1031/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1032/// predicate, do SROA now. 1033void SROA::DoScalarReplacement(AllocaInst *AI, 1034 std::vector<AllocaInst*> &WorkList) { 1035 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1036 SmallVector<AllocaInst*, 32> ElementAllocas; 1037 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1038 ElementAllocas.reserve(ST->getNumContainedTypes()); 1039 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1040 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1041 AI->getAlignment(), 1042 AI->getName() + "." + Twine(i), AI); 1043 ElementAllocas.push_back(NA); 1044 WorkList.push_back(NA); // Add to worklist for recursive processing 1045 } 1046 } else { 1047 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1048 ElementAllocas.reserve(AT->getNumElements()); 1049 const Type *ElTy = AT->getElementType(); 1050 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1051 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1052 AI->getName() + "." + Twine(i), AI); 1053 ElementAllocas.push_back(NA); 1054 WorkList.push_back(NA); // Add to worklist for recursive processing 1055 } 1056 } 1057 1058 // Now that we have created the new alloca instructions, rewrite all the 1059 // uses of the old alloca. 1060 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1061 1062 // Now erase any instructions that were made dead while rewriting the alloca. 1063 DeleteDeadInstructions(); 1064 AI->eraseFromParent(); 1065 1066 ++NumReplaced; 1067} 1068 1069/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1070/// recursively including all their operands that become trivially dead. 1071void SROA::DeleteDeadInstructions() { 1072 while (!DeadInsts.empty()) { 1073 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1074 1075 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1076 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1077 // Zero out the operand and see if it becomes trivially dead. 1078 // (But, don't add allocas to the dead instruction list -- they are 1079 // already on the worklist and will be deleted separately.) 1080 *OI = 0; 1081 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1082 DeadInsts.push_back(U); 1083 } 1084 1085 I->eraseFromParent(); 1086 } 1087} 1088 1089/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1090/// performing scalar replacement of alloca AI. The results are flagged in 1091/// the Info parameter. Offset indicates the position within AI that is 1092/// referenced by this instruction. 1093void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1094 AllocaInfo &Info) { 1095 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1096 Instruction *User = cast<Instruction>(*UI); 1097 1098 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1099 isSafeForScalarRepl(BC, AI, Offset, Info); 1100 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1101 uint64_t GEPOffset = Offset; 1102 isSafeGEP(GEPI, AI, GEPOffset, Info); 1103 if (!Info.isUnsafe) 1104 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 1105 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1106 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1107 if (Length) 1108 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 1109 UI.getOperandNo() == 0, Info); 1110 else 1111 MarkUnsafe(Info); 1112 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1113 if (!LI->isVolatile()) { 1114 const Type *LIType = LI->getType(); 1115 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 1116 LIType, false, Info); 1117 Info.hasALoadOrStore = true; 1118 } else 1119 MarkUnsafe(Info); 1120 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1121 // Store is ok if storing INTO the pointer, not storing the pointer 1122 if (!SI->isVolatile() && SI->getOperand(0) != I) { 1123 const Type *SIType = SI->getOperand(0)->getType(); 1124 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 1125 SIType, true, Info); 1126 Info.hasALoadOrStore = true; 1127 } else 1128 MarkUnsafe(Info); 1129 } else { 1130 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 1131 MarkUnsafe(Info); 1132 } 1133 if (Info.isUnsafe) return; 1134 } 1135} 1136 1137/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1138/// replacement. It is safe when all the indices are constant, in-bounds 1139/// references, and when the resulting offset corresponds to an element within 1140/// the alloca type. The results are flagged in the Info parameter. Upon 1141/// return, Offset is adjusted as specified by the GEP indices. 1142void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 1143 uint64_t &Offset, AllocaInfo &Info) { 1144 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1145 if (GEPIt == E) 1146 return; 1147 1148 // Walk through the GEP type indices, checking the types that this indexes 1149 // into. 1150 for (; GEPIt != E; ++GEPIt) { 1151 // Ignore struct elements, no extra checking needed for these. 1152 if ((*GEPIt)->isStructTy()) 1153 continue; 1154 1155 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1156 if (!IdxVal) 1157 return MarkUnsafe(Info); 1158 } 1159 1160 // Compute the offset due to this GEP and check if the alloca has a 1161 // component element at that offset. 1162 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1163 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1164 &Indices[0], Indices.size()); 1165 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 1166 MarkUnsafe(Info); 1167} 1168 1169/// isHomogeneousAggregate - Check if type T is a struct or array containing 1170/// elements of the same type (which is always true for arrays). If so, 1171/// return true with NumElts and EltTy set to the number of elements and the 1172/// element type, respectively. 1173static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1174 const Type *&EltTy) { 1175 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1176 NumElts = AT->getNumElements(); 1177 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1178 return true; 1179 } 1180 if (const StructType *ST = dyn_cast<StructType>(T)) { 1181 NumElts = ST->getNumContainedTypes(); 1182 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1183 for (unsigned n = 1; n < NumElts; ++n) { 1184 if (ST->getContainedType(n) != EltTy) 1185 return false; 1186 } 1187 return true; 1188 } 1189 return false; 1190} 1191 1192/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1193/// "homogeneous" aggregates with the same element type and number of elements. 1194static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1195 if (T1 == T2) 1196 return true; 1197 1198 unsigned NumElts1, NumElts2; 1199 const Type *EltTy1, *EltTy2; 1200 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1201 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1202 NumElts1 == NumElts2 && 1203 EltTy1 == EltTy2) 1204 return true; 1205 1206 return false; 1207} 1208 1209/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1210/// alloca or has an offset and size that corresponds to a component element 1211/// within it. The offset checked here may have been formed from a GEP with a 1212/// pointer bitcasted to a different type. 1213void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 1214 const Type *MemOpType, bool isStore, 1215 AllocaInfo &Info) { 1216 // Check if this is a load/store of the entire alloca. 1217 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 1218 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1219 // loads/stores (which are essentially the same as the MemIntrinsics with 1220 // regard to copying padding between elements). But, if an alloca is 1221 // flagged as both a source and destination of such operations, we'll need 1222 // to check later for padding between elements. 1223 if (!MemOpType || MemOpType->isIntegerTy()) { 1224 if (isStore) 1225 Info.isMemCpyDst = true; 1226 else 1227 Info.isMemCpySrc = true; 1228 return; 1229 } 1230 // This is also safe for references using a type that is compatible with 1231 // the type of the alloca, so that loads/stores can be rewritten using 1232 // insertvalue/extractvalue. 1233 if (isCompatibleAggregate(MemOpType, AI->getAllocatedType())) { 1234 Info.hasSubelementAccess = true; 1235 return; 1236 } 1237 } 1238 // Check if the offset/size correspond to a component within the alloca type. 1239 const Type *T = AI->getAllocatedType(); 1240 if (TypeHasComponent(T, Offset, MemSize)) { 1241 Info.hasSubelementAccess = true; 1242 return; 1243 } 1244 1245 return MarkUnsafe(Info); 1246} 1247 1248/// TypeHasComponent - Return true if T has a component type with the 1249/// specified offset and size. If Size is zero, do not check the size. 1250bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1251 const Type *EltTy; 1252 uint64_t EltSize; 1253 if (const StructType *ST = dyn_cast<StructType>(T)) { 1254 const StructLayout *Layout = TD->getStructLayout(ST); 1255 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1256 EltTy = ST->getContainedType(EltIdx); 1257 EltSize = TD->getTypeAllocSize(EltTy); 1258 Offset -= Layout->getElementOffset(EltIdx); 1259 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1260 EltTy = AT->getElementType(); 1261 EltSize = TD->getTypeAllocSize(EltTy); 1262 if (Offset >= AT->getNumElements() * EltSize) 1263 return false; 1264 Offset %= EltSize; 1265 } else { 1266 return false; 1267 } 1268 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1269 return true; 1270 // Check if the component spans multiple elements. 1271 if (Offset + Size > EltSize) 1272 return false; 1273 return TypeHasComponent(EltTy, Offset, Size); 1274} 1275 1276/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1277/// the instruction I, which references it, to use the separate elements. 1278/// Offset indicates the position within AI that is referenced by this 1279/// instruction. 1280void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1281 SmallVector<AllocaInst*, 32> &NewElts) { 1282 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1283 Instruction *User = cast<Instruction>(*UI); 1284 1285 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1286 RewriteBitCast(BC, AI, Offset, NewElts); 1287 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1288 RewriteGEP(GEPI, AI, Offset, NewElts); 1289 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1290 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1291 uint64_t MemSize = Length->getZExtValue(); 1292 if (Offset == 0 && 1293 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1294 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1295 // Otherwise the intrinsic can only touch a single element and the 1296 // address operand will be updated, so nothing else needs to be done. 1297 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1298 const Type *LIType = LI->getType(); 1299 1300 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1301 // Replace: 1302 // %res = load { i32, i32 }* %alloc 1303 // with: 1304 // %load.0 = load i32* %alloc.0 1305 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1306 // %load.1 = load i32* %alloc.1 1307 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1308 // (Also works for arrays instead of structs) 1309 Value *Insert = UndefValue::get(LIType); 1310 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1311 Value *Load = new LoadInst(NewElts[i], "load", LI); 1312 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1313 } 1314 LI->replaceAllUsesWith(Insert); 1315 DeadInsts.push_back(LI); 1316 } else if (LIType->isIntegerTy() && 1317 TD->getTypeAllocSize(LIType) == 1318 TD->getTypeAllocSize(AI->getAllocatedType())) { 1319 // If this is a load of the entire alloca to an integer, rewrite it. 1320 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1321 } 1322 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1323 Value *Val = SI->getOperand(0); 1324 const Type *SIType = Val->getType(); 1325 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1326 // Replace: 1327 // store { i32, i32 } %val, { i32, i32 }* %alloc 1328 // with: 1329 // %val.0 = extractvalue { i32, i32 } %val, 0 1330 // store i32 %val.0, i32* %alloc.0 1331 // %val.1 = extractvalue { i32, i32 } %val, 1 1332 // store i32 %val.1, i32* %alloc.1 1333 // (Also works for arrays instead of structs) 1334 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1335 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1336 new StoreInst(Extract, NewElts[i], SI); 1337 } 1338 DeadInsts.push_back(SI); 1339 } else if (SIType->isIntegerTy() && 1340 TD->getTypeAllocSize(SIType) == 1341 TD->getTypeAllocSize(AI->getAllocatedType())) { 1342 // If this is a store of the entire alloca from an integer, rewrite it. 1343 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1344 } 1345 } 1346 } 1347} 1348 1349/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1350/// and recursively continue updating all of its uses. 1351void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1352 SmallVector<AllocaInst*, 32> &NewElts) { 1353 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1354 if (BC->getOperand(0) != AI) 1355 return; 1356 1357 // The bitcast references the original alloca. Replace its uses with 1358 // references to the first new element alloca. 1359 Instruction *Val = NewElts[0]; 1360 if (Val->getType() != BC->getDestTy()) { 1361 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1362 Val->takeName(BC); 1363 } 1364 BC->replaceAllUsesWith(Val); 1365 DeadInsts.push_back(BC); 1366} 1367 1368/// FindElementAndOffset - Return the index of the element containing Offset 1369/// within the specified type, which must be either a struct or an array. 1370/// Sets T to the type of the element and Offset to the offset within that 1371/// element. IdxTy is set to the type of the index result to be used in a 1372/// GEP instruction. 1373uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1374 const Type *&IdxTy) { 1375 uint64_t Idx = 0; 1376 if (const StructType *ST = dyn_cast<StructType>(T)) { 1377 const StructLayout *Layout = TD->getStructLayout(ST); 1378 Idx = Layout->getElementContainingOffset(Offset); 1379 T = ST->getContainedType(Idx); 1380 Offset -= Layout->getElementOffset(Idx); 1381 IdxTy = Type::getInt32Ty(T->getContext()); 1382 return Idx; 1383 } 1384 const ArrayType *AT = cast<ArrayType>(T); 1385 T = AT->getElementType(); 1386 uint64_t EltSize = TD->getTypeAllocSize(T); 1387 Idx = Offset / EltSize; 1388 Offset -= Idx * EltSize; 1389 IdxTy = Type::getInt64Ty(T->getContext()); 1390 return Idx; 1391} 1392 1393/// RewriteGEP - Check if this GEP instruction moves the pointer across 1394/// elements of the alloca that are being split apart, and if so, rewrite 1395/// the GEP to be relative to the new element. 1396void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1397 SmallVector<AllocaInst*, 32> &NewElts) { 1398 uint64_t OldOffset = Offset; 1399 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1400 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1401 &Indices[0], Indices.size()); 1402 1403 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1404 1405 const Type *T = AI->getAllocatedType(); 1406 const Type *IdxTy; 1407 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1408 if (GEPI->getOperand(0) == AI) 1409 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1410 1411 T = AI->getAllocatedType(); 1412 uint64_t EltOffset = Offset; 1413 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1414 1415 // If this GEP does not move the pointer across elements of the alloca 1416 // being split, then it does not needs to be rewritten. 1417 if (Idx == OldIdx) 1418 return; 1419 1420 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1421 SmallVector<Value*, 8> NewArgs; 1422 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1423 while (EltOffset != 0) { 1424 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1425 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1426 } 1427 Instruction *Val = NewElts[Idx]; 1428 if (NewArgs.size() > 1) { 1429 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1430 NewArgs.end(), "", GEPI); 1431 Val->takeName(GEPI); 1432 } 1433 if (Val->getType() != GEPI->getType()) 1434 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1435 GEPI->replaceAllUsesWith(Val); 1436 DeadInsts.push_back(GEPI); 1437} 1438 1439/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1440/// Rewrite it to copy or set the elements of the scalarized memory. 1441void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1442 AllocaInst *AI, 1443 SmallVector<AllocaInst*, 32> &NewElts) { 1444 // If this is a memcpy/memmove, construct the other pointer as the 1445 // appropriate type. The "Other" pointer is the pointer that goes to memory 1446 // that doesn't have anything to do with the alloca that we are promoting. For 1447 // memset, this Value* stays null. 1448 Value *OtherPtr = 0; 1449 unsigned MemAlignment = MI->getAlignment(); 1450 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1451 if (Inst == MTI->getRawDest()) 1452 OtherPtr = MTI->getRawSource(); 1453 else { 1454 assert(Inst == MTI->getRawSource()); 1455 OtherPtr = MTI->getRawDest(); 1456 } 1457 } 1458 1459 // If there is an other pointer, we want to convert it to the same pointer 1460 // type as AI has, so we can GEP through it safely. 1461 if (OtherPtr) { 1462 unsigned AddrSpace = 1463 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 1464 1465 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1466 // optimization, but it's also required to detect the corner case where 1467 // both pointer operands are referencing the same memory, and where 1468 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1469 // function is only called for mem intrinsics that access the whole 1470 // aggregate, so non-zero GEPs are not an issue here.) 1471 OtherPtr = OtherPtr->stripPointerCasts(); 1472 1473 // Copying the alloca to itself is a no-op: just delete it. 1474 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1475 // This code will run twice for a no-op memcpy -- once for each operand. 1476 // Put only one reference to MI on the DeadInsts list. 1477 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1478 E = DeadInsts.end(); I != E; ++I) 1479 if (*I == MI) return; 1480 DeadInsts.push_back(MI); 1481 return; 1482 } 1483 1484 // If the pointer is not the right type, insert a bitcast to the right 1485 // type. 1486 const Type *NewTy = 1487 PointerType::get(AI->getType()->getElementType(), AddrSpace); 1488 1489 if (OtherPtr->getType() != NewTy) 1490 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 1491 } 1492 1493 // Process each element of the aggregate. 1494 bool SROADest = MI->getRawDest() == Inst; 1495 1496 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1497 1498 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1499 // If this is a memcpy/memmove, emit a GEP of the other element address. 1500 Value *OtherElt = 0; 1501 unsigned OtherEltAlign = MemAlignment; 1502 1503 if (OtherPtr) { 1504 Value *Idx[2] = { Zero, 1505 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1506 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1507 OtherPtr->getName()+"."+Twine(i), 1508 MI); 1509 uint64_t EltOffset; 1510 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1511 const Type *OtherTy = OtherPtrTy->getElementType(); 1512 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1513 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1514 } else { 1515 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 1516 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1517 } 1518 1519 // The alignment of the other pointer is the guaranteed alignment of the 1520 // element, which is affected by both the known alignment of the whole 1521 // mem intrinsic and the alignment of the element. If the alignment of 1522 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 1523 // known alignment is just 4 bytes. 1524 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 1525 } 1526 1527 Value *EltPtr = NewElts[i]; 1528 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 1529 1530 // If we got down to a scalar, insert a load or store as appropriate. 1531 if (EltTy->isSingleValueType()) { 1532 if (isa<MemTransferInst>(MI)) { 1533 if (SROADest) { 1534 // From Other to Alloca. 1535 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 1536 new StoreInst(Elt, EltPtr, MI); 1537 } else { 1538 // From Alloca to Other. 1539 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 1540 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 1541 } 1542 continue; 1543 } 1544 assert(isa<MemSetInst>(MI)); 1545 1546 // If the stored element is zero (common case), just store a null 1547 // constant. 1548 Constant *StoreVal; 1549 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 1550 if (CI->isZero()) { 1551 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 1552 } else { 1553 // If EltTy is a vector type, get the element type. 1554 const Type *ValTy = EltTy->getScalarType(); 1555 1556 // Construct an integer with the right value. 1557 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 1558 APInt OneVal(EltSize, CI->getZExtValue()); 1559 APInt TotalVal(OneVal); 1560 // Set each byte. 1561 for (unsigned i = 0; 8*i < EltSize; ++i) { 1562 TotalVal = TotalVal.shl(8); 1563 TotalVal |= OneVal; 1564 } 1565 1566 // Convert the integer value to the appropriate type. 1567 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 1568 if (ValTy->isPointerTy()) 1569 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 1570 else if (ValTy->isFloatingPointTy()) 1571 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 1572 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 1573 1574 // If the requested value was a vector constant, create it. 1575 if (EltTy != ValTy) { 1576 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 1577 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 1578 StoreVal = ConstantVector::get(&Elts[0], NumElts); 1579 } 1580 } 1581 new StoreInst(StoreVal, EltPtr, MI); 1582 continue; 1583 } 1584 // Otherwise, if we're storing a byte variable, use a memset call for 1585 // this element. 1586 } 1587 1588 unsigned EltSize = TD->getTypeAllocSize(EltTy); 1589 1590 IRBuilder<> Builder(MI); 1591 1592 // Finally, insert the meminst for this element. 1593 if (isa<MemSetInst>(MI)) { 1594 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 1595 MI->isVolatile()); 1596 } else { 1597 assert(isa<MemTransferInst>(MI)); 1598 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 1599 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 1600 1601 if (isa<MemCpyInst>(MI)) 1602 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 1603 else 1604 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 1605 } 1606 } 1607 DeadInsts.push_back(MI); 1608} 1609 1610/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 1611/// overwrites the entire allocation. Extract out the pieces of the stored 1612/// integer and store them individually. 1613void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 1614 SmallVector<AllocaInst*, 32> &NewElts){ 1615 // Extract each element out of the integer according to its structure offset 1616 // and store the element value to the individual alloca. 1617 Value *SrcVal = SI->getOperand(0); 1618 const Type *AllocaEltTy = AI->getAllocatedType(); 1619 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1620 1621 IRBuilder<> Builder(SI); 1622 1623 // Handle tail padding by extending the operand 1624 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 1625 SrcVal = Builder.CreateZExt(SrcVal, 1626 IntegerType::get(SI->getContext(), AllocaSizeBits)); 1627 1628 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 1629 << '\n'); 1630 1631 // There are two forms here: AI could be an array or struct. Both cases 1632 // have different ways to compute the element offset. 1633 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1634 const StructLayout *Layout = TD->getStructLayout(EltSTy); 1635 1636 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1637 // Get the number of bits to shift SrcVal to get the value. 1638 const Type *FieldTy = EltSTy->getElementType(i); 1639 uint64_t Shift = Layout->getElementOffsetInBits(i); 1640 1641 if (TD->isBigEndian()) 1642 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 1643 1644 Value *EltVal = SrcVal; 1645 if (Shift) { 1646 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1647 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 1648 } 1649 1650 // Truncate down to an integer of the right size. 1651 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1652 1653 // Ignore zero sized fields like {}, they obviously contain no data. 1654 if (FieldSizeBits == 0) continue; 1655 1656 if (FieldSizeBits != AllocaSizeBits) 1657 EltVal = Builder.CreateTrunc(EltVal, 1658 IntegerType::get(SI->getContext(), FieldSizeBits)); 1659 Value *DestField = NewElts[i]; 1660 if (EltVal->getType() == FieldTy) { 1661 // Storing to an integer field of this size, just do it. 1662 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 1663 // Bitcast to the right element type (for fp/vector values). 1664 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 1665 } else { 1666 // Otherwise, bitcast the dest pointer (for aggregates). 1667 DestField = Builder.CreateBitCast(DestField, 1668 PointerType::getUnqual(EltVal->getType())); 1669 } 1670 new StoreInst(EltVal, DestField, SI); 1671 } 1672 1673 } else { 1674 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1675 const Type *ArrayEltTy = ATy->getElementType(); 1676 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1677 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1678 1679 uint64_t Shift; 1680 1681 if (TD->isBigEndian()) 1682 Shift = AllocaSizeBits-ElementOffset; 1683 else 1684 Shift = 0; 1685 1686 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1687 // Ignore zero sized fields like {}, they obviously contain no data. 1688 if (ElementSizeBits == 0) continue; 1689 1690 Value *EltVal = SrcVal; 1691 if (Shift) { 1692 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1693 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 1694 } 1695 1696 // Truncate down to an integer of the right size. 1697 if (ElementSizeBits != AllocaSizeBits) 1698 EltVal = Builder.CreateTrunc(EltVal, 1699 IntegerType::get(SI->getContext(), 1700 ElementSizeBits)); 1701 Value *DestField = NewElts[i]; 1702 if (EltVal->getType() == ArrayEltTy) { 1703 // Storing to an integer field of this size, just do it. 1704 } else if (ArrayEltTy->isFloatingPointTy() || 1705 ArrayEltTy->isVectorTy()) { 1706 // Bitcast to the right element type (for fp/vector values). 1707 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 1708 } else { 1709 // Otherwise, bitcast the dest pointer (for aggregates). 1710 DestField = Builder.CreateBitCast(DestField, 1711 PointerType::getUnqual(EltVal->getType())); 1712 } 1713 new StoreInst(EltVal, DestField, SI); 1714 1715 if (TD->isBigEndian()) 1716 Shift -= ElementOffset; 1717 else 1718 Shift += ElementOffset; 1719 } 1720 } 1721 1722 DeadInsts.push_back(SI); 1723} 1724 1725/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1726/// an integer. Load the individual pieces to form the aggregate value. 1727void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1728 SmallVector<AllocaInst*, 32> &NewElts) { 1729 // Extract each element out of the NewElts according to its structure offset 1730 // and form the result value. 1731 const Type *AllocaEltTy = AI->getAllocatedType(); 1732 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1733 1734 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1735 << '\n'); 1736 1737 // There are two forms here: AI could be an array or struct. Both cases 1738 // have different ways to compute the element offset. 1739 const StructLayout *Layout = 0; 1740 uint64_t ArrayEltBitOffset = 0; 1741 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1742 Layout = TD->getStructLayout(EltSTy); 1743 } else { 1744 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1745 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1746 } 1747 1748 Value *ResultVal = 1749 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1750 1751 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1752 // Load the value from the alloca. If the NewElt is an aggregate, cast 1753 // the pointer to an integer of the same size before doing the load. 1754 Value *SrcField = NewElts[i]; 1755 const Type *FieldTy = 1756 cast<PointerType>(SrcField->getType())->getElementType(); 1757 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1758 1759 // Ignore zero sized fields like {}, they obviously contain no data. 1760 if (FieldSizeBits == 0) continue; 1761 1762 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1763 FieldSizeBits); 1764 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1765 !FieldTy->isVectorTy()) 1766 SrcField = new BitCastInst(SrcField, 1767 PointerType::getUnqual(FieldIntTy), 1768 "", LI); 1769 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1770 1771 // If SrcField is a fp or vector of the right size but that isn't an 1772 // integer type, bitcast to an integer so we can shift it. 1773 if (SrcField->getType() != FieldIntTy) 1774 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1775 1776 // Zero extend the field to be the same size as the final alloca so that 1777 // we can shift and insert it. 1778 if (SrcField->getType() != ResultVal->getType()) 1779 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1780 1781 // Determine the number of bits to shift SrcField. 1782 uint64_t Shift; 1783 if (Layout) // Struct case. 1784 Shift = Layout->getElementOffsetInBits(i); 1785 else // Array case. 1786 Shift = i*ArrayEltBitOffset; 1787 1788 if (TD->isBigEndian()) 1789 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1790 1791 if (Shift) { 1792 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1793 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1794 } 1795 1796 // Don't create an 'or x, 0' on the first iteration. 1797 if (!isa<Constant>(ResultVal) || 1798 !cast<Constant>(ResultVal)->isNullValue()) 1799 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1800 else 1801 ResultVal = SrcField; 1802 } 1803 1804 // Handle tail padding by truncating the result 1805 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1806 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1807 1808 LI->replaceAllUsesWith(ResultVal); 1809 DeadInsts.push_back(LI); 1810} 1811 1812/// HasPadding - Return true if the specified type has any structure or 1813/// alignment padding in between the elements that would be split apart 1814/// by SROA; return false otherwise. 1815static bool HasPadding(const Type *Ty, const TargetData &TD) { 1816 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1817 Ty = ATy->getElementType(); 1818 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1819 } 1820 1821 // SROA currently handles only Arrays and Structs. 1822 const StructType *STy = cast<StructType>(Ty); 1823 const StructLayout *SL = TD.getStructLayout(STy); 1824 unsigned PrevFieldBitOffset = 0; 1825 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1826 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1827 1828 // Check to see if there is any padding between this element and the 1829 // previous one. 1830 if (i) { 1831 unsigned PrevFieldEnd = 1832 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1833 if (PrevFieldEnd < FieldBitOffset) 1834 return true; 1835 } 1836 PrevFieldBitOffset = FieldBitOffset; 1837 } 1838 // Check for tail padding. 1839 if (unsigned EltCount = STy->getNumElements()) { 1840 unsigned PrevFieldEnd = PrevFieldBitOffset + 1841 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1842 if (PrevFieldEnd < SL->getSizeInBits()) 1843 return true; 1844 } 1845 return false; 1846} 1847 1848/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1849/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1850/// or 1 if safe after canonicalization has been performed. 1851bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1852 // Loop over the use list of the alloca. We can only transform it if all of 1853 // the users are safe to transform. 1854 AllocaInfo Info; 1855 1856 isSafeForScalarRepl(AI, AI, 0, Info); 1857 if (Info.isUnsafe) { 1858 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1859 return false; 1860 } 1861 1862 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1863 // source and destination, we have to be careful. In particular, the memcpy 1864 // could be moving around elements that live in structure padding of the LLVM 1865 // types, but may actually be used. In these cases, we refuse to promote the 1866 // struct. 1867 if (Info.isMemCpySrc && Info.isMemCpyDst && 1868 HasPadding(AI->getAllocatedType(), *TD)) 1869 return false; 1870 1871 // If the alloca never has an access to just *part* of it, but is accessed 1872 // via loads and stores, then we should use ConvertToScalarInfo to promote 1873 // the alloca instead of promoting each piece at a time and inserting fission 1874 // and fusion code. 1875 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 1876 // If the struct/array just has one element, use basic SRoA. 1877 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1878 if (ST->getNumElements() > 1) return false; 1879 } else { 1880 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 1881 return false; 1882 } 1883 } 1884 return true; 1885} 1886 1887 1888 1889/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1890/// some part of a constant global variable. This intentionally only accepts 1891/// constant expressions because we don't can't rewrite arbitrary instructions. 1892static bool PointsToConstantGlobal(Value *V) { 1893 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1894 return GV->isConstant(); 1895 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1896 if (CE->getOpcode() == Instruction::BitCast || 1897 CE->getOpcode() == Instruction::GetElementPtr) 1898 return PointsToConstantGlobal(CE->getOperand(0)); 1899 return false; 1900} 1901 1902/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1903/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1904/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1905/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1906/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1907/// the alloca, and if the source pointer is a pointer to a constant global, we 1908/// can optimize this. 1909static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 1910 bool isOffset) { 1911 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1912 User *U = cast<Instruction>(*UI); 1913 1914 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1915 // Ignore non-volatile loads, they are always ok. 1916 if (LI->isVolatile()) return false; 1917 continue; 1918 } 1919 1920 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1921 // If uses of the bitcast are ok, we are ok. 1922 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1923 return false; 1924 continue; 1925 } 1926 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 1927 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1928 // doesn't, it does. 1929 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1930 isOffset || !GEP->hasAllZeroIndices())) 1931 return false; 1932 continue; 1933 } 1934 1935 if (CallSite CS = U) { 1936 // If this is a readonly/readnone call site, then we know it is just a 1937 // load and we can ignore it. 1938 if (CS.onlyReadsMemory()) 1939 continue; 1940 1941 // If this is the function being called then we treat it like a load and 1942 // ignore it. 1943 if (CS.isCallee(UI)) 1944 continue; 1945 1946 // If this is being passed as a byval argument, the caller is making a 1947 // copy, so it is only a read of the alloca. 1948 unsigned ArgNo = CS.getArgumentNo(UI); 1949 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 1950 continue; 1951 } 1952 1953 // If this is isn't our memcpy/memmove, reject it as something we can't 1954 // handle. 1955 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 1956 if (MI == 0) 1957 return false; 1958 1959 // If the transfer is using the alloca as a source of the transfer, then 1960 // ignore it since it is a load (unless the transfer is volatile). 1961 if (UI.getOperandNo() == 1) { 1962 if (MI->isVolatile()) return false; 1963 continue; 1964 } 1965 1966 // If we already have seen a copy, reject the second one. 1967 if (TheCopy) return false; 1968 1969 // If the pointer has been offset from the start of the alloca, we can't 1970 // safely handle this. 1971 if (isOffset) return false; 1972 1973 // If the memintrinsic isn't using the alloca as the dest, reject it. 1974 if (UI.getOperandNo() != 0) return false; 1975 1976 // If the source of the memcpy/move is not a constant global, reject it. 1977 if (!PointsToConstantGlobal(MI->getSource())) 1978 return false; 1979 1980 // Otherwise, the transform is safe. Remember the copy instruction. 1981 TheCopy = MI; 1982 } 1983 return true; 1984} 1985 1986/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1987/// modified by a copy from a constant global. If we can prove this, we can 1988/// replace any uses of the alloca with uses of the global directly. 1989MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1990 MemTransferInst *TheCopy = 0; 1991 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1992 return TheCopy; 1993 return 0; 1994} 1995