ScalarReplAggregates.cpp revision 145c532e68acdf70d40bab5bc2034f692848b8dc
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/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 hasDT, char &ID) 58 : FunctionPass(ID), HasDomTree(hasDT) { 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 HasDomTree; 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 /// The alloca to promote. 83 AllocaInst *AI; 84 85 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite 86 /// looping and avoid redundant work. 87 SmallPtrSet<PHINode*, 8> CheckedPHIs; 88 89 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 90 bool isUnsafe : 1; 91 92 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 93 bool isMemCpySrc : 1; 94 95 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 96 bool isMemCpyDst : 1; 97 98 /// hasSubelementAccess - This is true if a subelement of the alloca is 99 /// ever accessed, or false if the alloca is only accessed with mem 100 /// intrinsics or load/store that only access the entire alloca at once. 101 bool hasSubelementAccess : 1; 102 103 /// hasALoadOrStore - This is true if there are any loads or stores to it. 104 /// The alloca may just be accessed with memcpy, for example, which would 105 /// not set this. 106 bool hasALoadOrStore : 1; 107 108 explicit AllocaInfo(AllocaInst *ai) 109 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 110 hasSubelementAccess(false), hasALoadOrStore(false) {} 111 }; 112 113 unsigned SRThreshold; 114 115 void MarkUnsafe(AllocaInfo &I, Instruction *User) { 116 I.isUnsafe = true; 117 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); 118 } 119 120 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 121 122 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info); 123 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset, 124 AllocaInfo &Info); 125 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info); 126 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 127 const Type *MemOpType, bool isStore, AllocaInfo &Info, 128 Instruction *TheAccess, bool AllowWholeAccess); 129 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 130 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 131 const Type *&IdxTy); 132 133 void DoScalarReplacement(AllocaInst *AI, 134 std::vector<AllocaInst*> &WorkList); 135 void DeleteDeadInstructions(); 136 137 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 138 SmallVector<AllocaInst*, 32> &NewElts); 139 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 140 SmallVector<AllocaInst*, 32> &NewElts); 141 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 142 SmallVector<AllocaInst*, 32> &NewElts); 143 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 144 AllocaInst *AI, 145 SmallVector<AllocaInst*, 32> &NewElts); 146 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 147 SmallVector<AllocaInst*, 32> &NewElts); 148 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 149 SmallVector<AllocaInst*, 32> &NewElts); 150 151 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 152 }; 153 154 // SROA_DT - SROA that uses DominatorTree. 155 struct SROA_DT : public SROA { 156 static char ID; 157 public: 158 SROA_DT(int T = -1) : SROA(T, true, ID) { 159 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 160 } 161 162 // getAnalysisUsage - This pass does not require any passes, but we know it 163 // will not alter the CFG, so say so. 164 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 165 AU.addRequired<DominatorTree>(); 166 AU.setPreservesCFG(); 167 } 168 }; 169 170 // SROA_SSAUp - SROA that uses SSAUpdater. 171 struct SROA_SSAUp : public SROA { 172 static char ID; 173 public: 174 SROA_SSAUp(int T = -1) : SROA(T, false, ID) { 175 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 176 } 177 178 // getAnalysisUsage - This pass does not require any passes, but we know it 179 // will not alter the CFG, so say so. 180 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 181 AU.setPreservesCFG(); 182 } 183 }; 184 185} 186 187char SROA_DT::ID = 0; 188char SROA_SSAUp::ID = 0; 189 190INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 191 "Scalar Replacement of Aggregates (DT)", false, false) 192INITIALIZE_PASS_DEPENDENCY(DominatorTree) 193INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 194 "Scalar Replacement of Aggregates (DT)", false, false) 195 196INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 197 "Scalar Replacement of Aggregates (SSAUp)", false, false) 198INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 199 "Scalar Replacement of Aggregates (SSAUp)", false, false) 200 201// Public interface to the ScalarReplAggregates pass 202FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 203 bool UseDomTree) { 204 if (UseDomTree) 205 return new SROA_DT(Threshold); 206 return new SROA_SSAUp(Threshold); 207} 208 209 210//===----------------------------------------------------------------------===// 211// Convert To Scalar Optimization. 212//===----------------------------------------------------------------------===// 213 214namespace { 215/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 216/// optimization, which scans the uses of an alloca and determines if it can 217/// rewrite it in terms of a single new alloca that can be mem2reg'd. 218class ConvertToScalarInfo { 219 /// AllocaSize - The size of the alloca being considered. 220 unsigned AllocaSize; 221 const TargetData &TD; 222 223 /// IsNotTrivial - This is set to true if there is some access to the object 224 /// which means that mem2reg can't promote it. 225 bool IsNotTrivial; 226 227 /// VectorTy - This tracks the type that we should promote the vector to if 228 /// it is possible to turn it into a vector. This starts out null, and if it 229 /// isn't possible to turn into a vector type, it gets set to VoidTy. 230 const Type *VectorTy; 231 232 /// HadAVector - True if there is at least one vector access to the alloca. 233 /// We don't want to turn random arrays into vectors and use vector element 234 /// insert/extract, but if there are element accesses to something that is 235 /// also declared as a vector, we do want to promote to a vector. 236 bool HadAVector; 237 238public: 239 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 240 : AllocaSize(Size), TD(td) { 241 IsNotTrivial = false; 242 VectorTy = 0; 243 HadAVector = false; 244 } 245 246 AllocaInst *TryConvert(AllocaInst *AI); 247 248private: 249 bool CanConvertToScalar(Value *V, uint64_t Offset); 250 void MergeInType(const Type *In, uint64_t Offset); 251 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 252 253 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 254 uint64_t Offset, IRBuilder<> &Builder); 255 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 256 uint64_t Offset, IRBuilder<> &Builder); 257}; 258} // end anonymous namespace. 259 260 261/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 262/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 263/// alloca if possible or null if not. 264AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 265 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 266 // out. 267 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 268 return 0; 269 270 // If we were able to find a vector type that can handle this with 271 // insert/extract elements, and if there was at least one use that had 272 // a vector type, promote this to a vector. We don't want to promote 273 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 274 // we just get a lot of insert/extracts. If at least one vector is 275 // involved, then we probably really do have a union of vector/array. 276 const Type *NewTy; 277 if (VectorTy && VectorTy->isVectorTy() && HadAVector) { 278 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 279 << *VectorTy << '\n'); 280 NewTy = VectorTy; // Use the vector type. 281 } else { 282 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 283 // Create and insert the integer alloca. 284 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 285 } 286 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 287 ConvertUsesToScalar(AI, NewAI, 0); 288 return NewAI; 289} 290 291/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy) 292/// so far at the offset specified by Offset (which is specified in bytes). 293/// 294/// There are two cases we handle here: 295/// 1) A union of vector types of the same size and potentially its elements. 296/// Here we turn element accesses into insert/extract element operations. 297/// This promotes a <4 x float> with a store of float to the third element 298/// into a <4 x float> that uses insert element. 299/// 2) A fully general blob of memory, which we turn into some (potentially 300/// large) integer type with extract and insert operations where the loads 301/// and stores would mutate the memory. We mark this by setting VectorTy 302/// to VoidTy. 303void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) { 304 // If we already decided to turn this into a blob of integer memory, there is 305 // nothing to be done. 306 if (VectorTy && VectorTy->isVoidTy()) 307 return; 308 309 // If this could be contributing to a vector, analyze it. 310 311 // If the In type is a vector that is the same size as the alloca, see if it 312 // matches the existing VecTy. 313 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 314 // Remember if we saw a vector type. 315 HadAVector = true; 316 317 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 318 // If we're storing/loading a vector of the right size, allow it as a 319 // vector. If this the first vector we see, remember the type so that 320 // we know the element size. If this is a subsequent access, ignore it 321 // even if it is a differing type but the same size. Worst case we can 322 // bitcast the resultant vectors. 323 if (VectorTy == 0) 324 VectorTy = VInTy; 325 return; 326 } 327 } else if (In->isFloatTy() || In->isDoubleTy() || 328 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 329 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 330 // If we're accessing something that could be an element of a vector, see 331 // if the implied vector agrees with what we already have and if Offset is 332 // compatible with it. 333 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 334 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 335 (VectorTy == 0 || 336 cast<VectorType>(VectorTy)->getElementType() 337 ->getPrimitiveSizeInBits()/8 == EltSize)) { 338 if (VectorTy == 0) 339 VectorTy = VectorType::get(In, AllocaSize/EltSize); 340 return; 341 } 342 } 343 344 // Otherwise, we have a case that we can't handle with an optimized vector 345 // form. We can still turn this into a large integer. 346 VectorTy = Type::getVoidTy(In->getContext()); 347} 348 349/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 350/// its accesses to a single vector type, return true and set VecTy to 351/// the new type. If we could convert the alloca into a single promotable 352/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 353/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 354/// is the current offset from the base of the alloca being analyzed. 355/// 356/// If we see at least one access to the value that is as a vector type, set the 357/// SawVec flag. 358bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 359 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 360 Instruction *User = cast<Instruction>(*UI); 361 362 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 363 // Don't break volatile loads. 364 if (LI->isVolatile()) 365 return false; 366 // Don't touch MMX operations. 367 if (LI->getType()->isX86_MMXTy()) 368 return false; 369 MergeInType(LI->getType(), Offset); 370 continue; 371 } 372 373 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 374 // Storing the pointer, not into the value? 375 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 376 // Don't touch MMX operations. 377 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 378 return false; 379 MergeInType(SI->getOperand(0)->getType(), Offset); 380 continue; 381 } 382 383 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 384 IsNotTrivial = true; // Can't be mem2reg'd. 385 if (!CanConvertToScalar(BCI, Offset)) 386 return false; 387 continue; 388 } 389 390 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 391 // If this is a GEP with a variable indices, we can't handle it. 392 if (!GEP->hasAllConstantIndices()) 393 return false; 394 395 // Compute the offset that this GEP adds to the pointer. 396 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 397 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 398 &Indices[0], Indices.size()); 399 // See if all uses can be converted. 400 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 401 return false; 402 IsNotTrivial = true; // Can't be mem2reg'd. 403 continue; 404 } 405 406 // If this is a constant sized memset of a constant value (e.g. 0) we can 407 // handle it. 408 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 409 // Store of constant value and constant size. 410 if (!isa<ConstantInt>(MSI->getValue()) || 411 !isa<ConstantInt>(MSI->getLength())) 412 return false; 413 IsNotTrivial = true; // Can't be mem2reg'd. 414 continue; 415 } 416 417 // If this is a memcpy or memmove into or out of the whole allocation, we 418 // can handle it like a load or store of the scalar type. 419 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 420 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 421 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 422 return false; 423 424 IsNotTrivial = true; // Can't be mem2reg'd. 425 continue; 426 } 427 428 // Otherwise, we cannot handle this! 429 return false; 430 } 431 432 return true; 433} 434 435/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 436/// directly. This happens when we are converting an "integer union" to a 437/// single integer scalar, or when we are converting a "vector union" to a 438/// vector with insert/extractelement instructions. 439/// 440/// Offset is an offset from the original alloca, in bits that need to be 441/// shifted to the right. By the end of this, there should be no uses of Ptr. 442void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 443 uint64_t Offset) { 444 while (!Ptr->use_empty()) { 445 Instruction *User = cast<Instruction>(Ptr->use_back()); 446 447 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 448 ConvertUsesToScalar(CI, NewAI, Offset); 449 CI->eraseFromParent(); 450 continue; 451 } 452 453 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 454 // Compute the offset that this GEP adds to the pointer. 455 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 456 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 457 &Indices[0], Indices.size()); 458 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 459 GEP->eraseFromParent(); 460 continue; 461 } 462 463 IRBuilder<> Builder(User); 464 465 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 466 // The load is a bit extract from NewAI shifted right by Offset bits. 467 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 468 Value *NewLoadVal 469 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 470 LI->replaceAllUsesWith(NewLoadVal); 471 LI->eraseFromParent(); 472 continue; 473 } 474 475 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 476 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 477 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 478 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 479 Builder); 480 Builder.CreateStore(New, NewAI); 481 SI->eraseFromParent(); 482 483 // If the load we just inserted is now dead, then the inserted store 484 // overwrote the entire thing. 485 if (Old->use_empty()) 486 Old->eraseFromParent(); 487 continue; 488 } 489 490 // If this is a constant sized memset of a constant value (e.g. 0) we can 491 // transform it into a store of the expanded constant value. 492 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 493 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 494 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 495 if (NumBytes != 0) { 496 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 497 498 // Compute the value replicated the right number of times. 499 APInt APVal(NumBytes*8, Val); 500 501 // Splat the value if non-zero. 502 if (Val) 503 for (unsigned i = 1; i != NumBytes; ++i) 504 APVal |= APVal << 8; 505 506 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 507 Value *New = ConvertScalar_InsertValue( 508 ConstantInt::get(User->getContext(), APVal), 509 Old, Offset, Builder); 510 Builder.CreateStore(New, NewAI); 511 512 // If the load we just inserted is now dead, then the memset overwrote 513 // the entire thing. 514 if (Old->use_empty()) 515 Old->eraseFromParent(); 516 } 517 MSI->eraseFromParent(); 518 continue; 519 } 520 521 // If this is a memcpy or memmove into or out of the whole allocation, we 522 // can handle it like a load or store of the scalar type. 523 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 524 assert(Offset == 0 && "must be store to start of alloca"); 525 526 // If the source and destination are both to the same alloca, then this is 527 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 528 // as appropriate. 529 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, 0)); 530 531 if (GetUnderlyingObject(MTI->getSource(), 0) != OrigAI) { 532 // Dest must be OrigAI, change this to be a load from the original 533 // pointer (bitcasted), then a store to our new alloca. 534 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 535 Value *SrcPtr = MTI->getSource(); 536 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 537 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 538 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 539 AIPTy = PointerType::get(AIPTy->getElementType(), 540 SPTy->getAddressSpace()); 541 } 542 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 543 544 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 545 SrcVal->setAlignment(MTI->getAlignment()); 546 Builder.CreateStore(SrcVal, NewAI); 547 } else if (GetUnderlyingObject(MTI->getDest(), 0) != OrigAI) { 548 // Src must be OrigAI, change this to be a load from NewAI then a store 549 // through the original dest pointer (bitcasted). 550 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 551 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 552 553 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 554 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 555 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 556 AIPTy = PointerType::get(AIPTy->getElementType(), 557 DPTy->getAddressSpace()); 558 } 559 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 560 561 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 562 NewStore->setAlignment(MTI->getAlignment()); 563 } else { 564 // Noop transfer. Src == Dst 565 } 566 567 MTI->eraseFromParent(); 568 continue; 569 } 570 571 llvm_unreachable("Unsupported operation!"); 572 } 573} 574 575/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 576/// or vector value FromVal, extracting the bits from the offset specified by 577/// Offset. This returns the value, which is of type ToType. 578/// 579/// This happens when we are converting an "integer union" to a single 580/// integer scalar, or when we are converting a "vector union" to a vector with 581/// insert/extractelement instructions. 582/// 583/// Offset is an offset from the original alloca, in bits that need to be 584/// shifted to the right. 585Value *ConvertToScalarInfo:: 586ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 587 uint64_t Offset, IRBuilder<> &Builder) { 588 // If the load is of the whole new alloca, no conversion is needed. 589 if (FromVal->getType() == ToType && Offset == 0) 590 return FromVal; 591 592 // If the result alloca is a vector type, this is either an element 593 // access or a bitcast to another vector type of the same size. 594 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 595 if (ToType->isVectorTy()) 596 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 597 598 // Otherwise it must be an element access. 599 unsigned Elt = 0; 600 if (Offset) { 601 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 602 Elt = Offset/EltSize; 603 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 604 } 605 // Return the element extracted out of it. 606 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 607 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 608 if (V->getType() != ToType) 609 V = Builder.CreateBitCast(V, ToType, "tmp"); 610 return V; 611 } 612 613 // If ToType is a first class aggregate, extract out each of the pieces and 614 // use insertvalue's to form the FCA. 615 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 616 const StructLayout &Layout = *TD.getStructLayout(ST); 617 Value *Res = UndefValue::get(ST); 618 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 619 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 620 Offset+Layout.getElementOffsetInBits(i), 621 Builder); 622 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 623 } 624 return Res; 625 } 626 627 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 628 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 629 Value *Res = UndefValue::get(AT); 630 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 631 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 632 Offset+i*EltSize, Builder); 633 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 634 } 635 return Res; 636 } 637 638 // Otherwise, this must be a union that was converted to an integer value. 639 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 640 641 // If this is a big-endian system and the load is narrower than the 642 // full alloca type, we need to do a shift to get the right bits. 643 int ShAmt = 0; 644 if (TD.isBigEndian()) { 645 // On big-endian machines, the lowest bit is stored at the bit offset 646 // from the pointer given by getTypeStoreSizeInBits. This matters for 647 // integers with a bitwidth that is not a multiple of 8. 648 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 649 TD.getTypeStoreSizeInBits(ToType) - Offset; 650 } else { 651 ShAmt = Offset; 652 } 653 654 // Note: we support negative bitwidths (with shl) which are not defined. 655 // We do this to support (f.e.) loads off the end of a structure where 656 // only some bits are used. 657 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 658 FromVal = Builder.CreateLShr(FromVal, 659 ConstantInt::get(FromVal->getType(), 660 ShAmt), "tmp"); 661 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 662 FromVal = Builder.CreateShl(FromVal, 663 ConstantInt::get(FromVal->getType(), 664 -ShAmt), "tmp"); 665 666 // Finally, unconditionally truncate the integer to the right width. 667 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 668 if (LIBitWidth < NTy->getBitWidth()) 669 FromVal = 670 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 671 LIBitWidth), "tmp"); 672 else if (LIBitWidth > NTy->getBitWidth()) 673 FromVal = 674 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 675 LIBitWidth), "tmp"); 676 677 // If the result is an integer, this is a trunc or bitcast. 678 if (ToType->isIntegerTy()) { 679 // Should be done. 680 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 681 // Just do a bitcast, we know the sizes match up. 682 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 683 } else { 684 // Otherwise must be a pointer. 685 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 686 } 687 assert(FromVal->getType() == ToType && "Didn't convert right?"); 688 return FromVal; 689} 690 691/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 692/// or vector value "Old" at the offset specified by Offset. 693/// 694/// This happens when we are converting an "integer union" to a 695/// single integer scalar, or when we are converting a "vector union" to a 696/// vector with insert/extractelement instructions. 697/// 698/// Offset is an offset from the original alloca, in bits that need to be 699/// shifted to the right. 700Value *ConvertToScalarInfo:: 701ConvertScalar_InsertValue(Value *SV, Value *Old, 702 uint64_t Offset, IRBuilder<> &Builder) { 703 // Convert the stored type to the actual type, shift it left to insert 704 // then 'or' into place. 705 const Type *AllocaType = Old->getType(); 706 LLVMContext &Context = Old->getContext(); 707 708 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 709 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 710 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 711 712 // Changing the whole vector with memset or with an access of a different 713 // vector type? 714 if (ValSize == VecSize) 715 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 716 717 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 718 719 // Must be an element insertion. 720 unsigned Elt = Offset/EltSize; 721 722 if (SV->getType() != VTy->getElementType()) 723 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 724 725 SV = Builder.CreateInsertElement(Old, SV, 726 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 727 "tmp"); 728 return SV; 729 } 730 731 // If SV is a first-class aggregate value, insert each value recursively. 732 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 733 const StructLayout &Layout = *TD.getStructLayout(ST); 734 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 735 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 736 Old = ConvertScalar_InsertValue(Elt, Old, 737 Offset+Layout.getElementOffsetInBits(i), 738 Builder); 739 } 740 return Old; 741 } 742 743 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 744 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 745 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 746 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 747 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 748 } 749 return Old; 750 } 751 752 // If SV is a float, convert it to the appropriate integer type. 753 // If it is a pointer, do the same. 754 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 755 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 756 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 757 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 758 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 759 SV = Builder.CreateBitCast(SV, 760 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 761 else if (SV->getType()->isPointerTy()) 762 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 763 764 // Zero extend or truncate the value if needed. 765 if (SV->getType() != AllocaType) { 766 if (SV->getType()->getPrimitiveSizeInBits() < 767 AllocaType->getPrimitiveSizeInBits()) 768 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 769 else { 770 // Truncation may be needed if storing more than the alloca can hold 771 // (undefined behavior). 772 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 773 SrcWidth = DestWidth; 774 SrcStoreWidth = DestStoreWidth; 775 } 776 } 777 778 // If this is a big-endian system and the store is narrower than the 779 // full alloca type, we need to do a shift to get the right bits. 780 int ShAmt = 0; 781 if (TD.isBigEndian()) { 782 // On big-endian machines, the lowest bit is stored at the bit offset 783 // from the pointer given by getTypeStoreSizeInBits. This matters for 784 // integers with a bitwidth that is not a multiple of 8. 785 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 786 } else { 787 ShAmt = Offset; 788 } 789 790 // Note: we support negative bitwidths (with shr) which are not defined. 791 // We do this to support (f.e.) stores off the end of a structure where 792 // only some bits in the structure are set. 793 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 794 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 795 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 796 ShAmt), "tmp"); 797 Mask <<= ShAmt; 798 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 799 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 800 -ShAmt), "tmp"); 801 Mask = Mask.lshr(-ShAmt); 802 } 803 804 // Mask out the bits we are about to insert from the old value, and or 805 // in the new bits. 806 if (SrcWidth != DestWidth) { 807 assert(DestWidth > SrcWidth); 808 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 809 SV = Builder.CreateOr(Old, SV, "ins"); 810 } 811 return SV; 812} 813 814 815//===----------------------------------------------------------------------===// 816// SRoA Driver 817//===----------------------------------------------------------------------===// 818 819 820bool SROA::runOnFunction(Function &F) { 821 TD = getAnalysisIfAvailable<TargetData>(); 822 823 bool Changed = performPromotion(F); 824 825 // FIXME: ScalarRepl currently depends on TargetData more than it 826 // theoretically needs to. It should be refactored in order to support 827 // target-independent IR. Until this is done, just skip the actual 828 // scalar-replacement portion of this pass. 829 if (!TD) return Changed; 830 831 while (1) { 832 bool LocalChange = performScalarRepl(F); 833 if (!LocalChange) break; // No need to repromote if no scalarrepl 834 Changed = true; 835 LocalChange = performPromotion(F); 836 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 837 } 838 839 return Changed; 840} 841 842namespace { 843class AllocaPromoter : public LoadAndStorePromoter { 844 AllocaInst *AI; 845public: 846 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S) 847 : LoadAndStorePromoter(Insts, S), AI(0) {} 848 849 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 850 // Remember which alloca we're promoting (for isInstInList). 851 this->AI = AI; 852 LoadAndStorePromoter::run(Insts); 853 AI->eraseFromParent(); 854 } 855 856 virtual bool isInstInList(Instruction *I, 857 const SmallVectorImpl<Instruction*> &Insts) const { 858 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 859 return LI->getOperand(0) == AI; 860 return cast<StoreInst>(I)->getPointerOperand() == AI; 861 } 862}; 863} // end anon namespace 864 865bool SROA::performPromotion(Function &F) { 866 std::vector<AllocaInst*> Allocas; 867 DominatorTree *DT = 0; 868 if (HasDomTree) 869 DT = &getAnalysis<DominatorTree>(); 870 871 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 872 873 bool Changed = false; 874 SmallVector<Instruction*, 64> Insts; 875 while (1) { 876 Allocas.clear(); 877 878 // Find allocas that are safe to promote, by looking at all instructions in 879 // the entry node 880 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 881 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 882 if (isAllocaPromotable(AI)) 883 Allocas.push_back(AI); 884 885 if (Allocas.empty()) break; 886 887 if (HasDomTree) 888 PromoteMemToReg(Allocas, *DT); 889 else { 890 SSAUpdater SSA; 891 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 892 AllocaInst *AI = Allocas[i]; 893 894 // Build list of instructions to promote. 895 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 896 UI != E; ++UI) 897 Insts.push_back(cast<Instruction>(*UI)); 898 899 AllocaPromoter(Insts, SSA).run(AI, Insts); 900 Insts.clear(); 901 } 902 } 903 NumPromoted += Allocas.size(); 904 Changed = true; 905 } 906 907 return Changed; 908} 909 910 911/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 912/// SROA. It must be a struct or array type with a small number of elements. 913static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 914 const Type *T = AI->getAllocatedType(); 915 // Do not promote any struct into more than 32 separate vars. 916 if (const StructType *ST = dyn_cast<StructType>(T)) 917 return ST->getNumElements() <= 32; 918 // Arrays are much less likely to be safe for SROA; only consider 919 // them if they are very small. 920 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 921 return AT->getNumElements() <= 8; 922 return false; 923} 924 925 926// performScalarRepl - This algorithm is a simple worklist driven algorithm, 927// which runs on all of the malloc/alloca instructions in the function, removing 928// them if they are only used by getelementptr instructions. 929// 930bool SROA::performScalarRepl(Function &F) { 931 std::vector<AllocaInst*> WorkList; 932 933 // Scan the entry basic block, adding allocas to the worklist. 934 BasicBlock &BB = F.getEntryBlock(); 935 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 936 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 937 WorkList.push_back(A); 938 939 // Process the worklist 940 bool Changed = false; 941 while (!WorkList.empty()) { 942 AllocaInst *AI = WorkList.back(); 943 WorkList.pop_back(); 944 945 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 946 // with unused elements. 947 if (AI->use_empty()) { 948 AI->eraseFromParent(); 949 Changed = true; 950 continue; 951 } 952 953 // If this alloca is impossible for us to promote, reject it early. 954 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 955 continue; 956 957 // Check to see if this allocation is only modified by a memcpy/memmove from 958 // a constant global. If this is the case, we can change all users to use 959 // the constant global instead. This is commonly produced by the CFE by 960 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 961 // is only subsequently read. 962 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 963 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 964 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 965 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 966 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 967 TheCopy->eraseFromParent(); // Don't mutate the global. 968 AI->eraseFromParent(); 969 ++NumGlobals; 970 Changed = true; 971 continue; 972 } 973 974 // Check to see if we can perform the core SROA transformation. We cannot 975 // transform the allocation instruction if it is an array allocation 976 // (allocations OF arrays are ok though), and an allocation of a scalar 977 // value cannot be decomposed at all. 978 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 979 980 // Do not promote [0 x %struct]. 981 if (AllocaSize == 0) continue; 982 983 // Do not promote any struct whose size is too big. 984 if (AllocaSize > SRThreshold) continue; 985 986 // If the alloca looks like a good candidate for scalar replacement, and if 987 // all its users can be transformed, then split up the aggregate into its 988 // separate elements. 989 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 990 DoScalarReplacement(AI, WorkList); 991 Changed = true; 992 continue; 993 } 994 995 // If we can turn this aggregate value (potentially with casts) into a 996 // simple scalar value that can be mem2reg'd into a register value. 997 // IsNotTrivial tracks whether this is something that mem2reg could have 998 // promoted itself. If so, we don't want to transform it needlessly. Note 999 // that we can't just check based on the type: the alloca may be of an i32 1000 // but that has pointer arithmetic to set byte 3 of it or something. 1001 if (AllocaInst *NewAI = 1002 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1003 NewAI->takeName(AI); 1004 AI->eraseFromParent(); 1005 ++NumConverted; 1006 Changed = true; 1007 continue; 1008 } 1009 1010 // Otherwise, couldn't process this alloca. 1011 } 1012 1013 return Changed; 1014} 1015 1016/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1017/// predicate, do SROA now. 1018void SROA::DoScalarReplacement(AllocaInst *AI, 1019 std::vector<AllocaInst*> &WorkList) { 1020 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1021 SmallVector<AllocaInst*, 32> ElementAllocas; 1022 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1023 ElementAllocas.reserve(ST->getNumContainedTypes()); 1024 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1025 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1026 AI->getAlignment(), 1027 AI->getName() + "." + Twine(i), AI); 1028 ElementAllocas.push_back(NA); 1029 WorkList.push_back(NA); // Add to worklist for recursive processing 1030 } 1031 } else { 1032 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1033 ElementAllocas.reserve(AT->getNumElements()); 1034 const Type *ElTy = AT->getElementType(); 1035 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1036 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1037 AI->getName() + "." + Twine(i), AI); 1038 ElementAllocas.push_back(NA); 1039 WorkList.push_back(NA); // Add to worklist for recursive processing 1040 } 1041 } 1042 1043 // Now that we have created the new alloca instructions, rewrite all the 1044 // uses of the old alloca. 1045 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1046 1047 // Now erase any instructions that were made dead while rewriting the alloca. 1048 DeleteDeadInstructions(); 1049 AI->eraseFromParent(); 1050 1051 ++NumReplaced; 1052} 1053 1054/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1055/// recursively including all their operands that become trivially dead. 1056void SROA::DeleteDeadInstructions() { 1057 while (!DeadInsts.empty()) { 1058 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1059 1060 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1061 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1062 // Zero out the operand and see if it becomes trivially dead. 1063 // (But, don't add allocas to the dead instruction list -- they are 1064 // already on the worklist and will be deleted separately.) 1065 *OI = 0; 1066 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1067 DeadInsts.push_back(U); 1068 } 1069 1070 I->eraseFromParent(); 1071 } 1072} 1073 1074/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1075/// performing scalar replacement of alloca AI. The results are flagged in 1076/// the Info parameter. Offset indicates the position within AI that is 1077/// referenced by this instruction. 1078void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1079 AllocaInfo &Info) { 1080 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1081 Instruction *User = cast<Instruction>(*UI); 1082 1083 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1084 isSafeForScalarRepl(BC, Offset, Info); 1085 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1086 uint64_t GEPOffset = Offset; 1087 isSafeGEP(GEPI, GEPOffset, Info); 1088 if (!Info.isUnsafe) 1089 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1090 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1091 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1092 if (Length == 0) 1093 return MarkUnsafe(Info, User); 1094 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1095 UI.getOperandNo() == 0, Info, MI, 1096 true /*AllowWholeAccess*/); 1097 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1098 if (LI->isVolatile()) 1099 return MarkUnsafe(Info, User); 1100 const Type *LIType = LI->getType(); 1101 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1102 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1103 Info.hasALoadOrStore = true; 1104 1105 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1106 // Store is ok if storing INTO the pointer, not storing the pointer 1107 if (SI->isVolatile() || SI->getOperand(0) == I) 1108 return MarkUnsafe(Info, User); 1109 1110 const Type *SIType = SI->getOperand(0)->getType(); 1111 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1112 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1113 Info.hasALoadOrStore = true; 1114 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1115 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1116 } else { 1117 return MarkUnsafe(Info, User); 1118 } 1119 if (Info.isUnsafe) return; 1120 } 1121} 1122 1123 1124/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1125/// derived from the alloca, we can often still split the alloca into elements. 1126/// This is useful if we have a large alloca where one element is phi'd 1127/// together somewhere: we can SRoA and promote all the other elements even if 1128/// we end up not being able to promote this one. 1129/// 1130/// All we require is that the uses of the PHI do not index into other parts of 1131/// the alloca. The most important use case for this is single load and stores 1132/// that are PHI'd together, which can happen due to code sinking. 1133void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1134 AllocaInfo &Info) { 1135 // If we've already checked this PHI, don't do it again. 1136 if (PHINode *PN = dyn_cast<PHINode>(I)) 1137 if (!Info.CheckedPHIs.insert(PN)) 1138 return; 1139 1140 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1141 Instruction *User = cast<Instruction>(*UI); 1142 1143 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1144 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1145 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1146 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1147 // but would have to prove that we're staying inside of an element being 1148 // promoted. 1149 if (!GEPI->hasAllZeroIndices()) 1150 return MarkUnsafe(Info, User); 1151 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1152 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1153 if (LI->isVolatile()) 1154 return MarkUnsafe(Info, User); 1155 const Type *LIType = LI->getType(); 1156 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1157 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1158 Info.hasALoadOrStore = true; 1159 1160 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1161 // Store is ok if storing INTO the pointer, not storing the pointer 1162 if (SI->isVolatile() || SI->getOperand(0) == I) 1163 return MarkUnsafe(Info, User); 1164 1165 const Type *SIType = SI->getOperand(0)->getType(); 1166 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1167 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1168 Info.hasALoadOrStore = true; 1169 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1170 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1171 } else { 1172 return MarkUnsafe(Info, User); 1173 } 1174 if (Info.isUnsafe) return; 1175 } 1176} 1177 1178/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1179/// replacement. It is safe when all the indices are constant, in-bounds 1180/// references, and when the resulting offset corresponds to an element within 1181/// the alloca type. The results are flagged in the Info parameter. Upon 1182/// return, Offset is adjusted as specified by the GEP indices. 1183void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1184 uint64_t &Offset, AllocaInfo &Info) { 1185 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1186 if (GEPIt == E) 1187 return; 1188 1189 // Walk through the GEP type indices, checking the types that this indexes 1190 // into. 1191 for (; GEPIt != E; ++GEPIt) { 1192 // Ignore struct elements, no extra checking needed for these. 1193 if ((*GEPIt)->isStructTy()) 1194 continue; 1195 1196 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1197 if (!IdxVal) 1198 return MarkUnsafe(Info, GEPI); 1199 } 1200 1201 // Compute the offset due to this GEP and check if the alloca has a 1202 // component element at that offset. 1203 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1204 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1205 &Indices[0], Indices.size()); 1206 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1207 MarkUnsafe(Info, GEPI); 1208} 1209 1210/// isHomogeneousAggregate - Check if type T is a struct or array containing 1211/// elements of the same type (which is always true for arrays). If so, 1212/// return true with NumElts and EltTy set to the number of elements and the 1213/// element type, respectively. 1214static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1215 const Type *&EltTy) { 1216 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1217 NumElts = AT->getNumElements(); 1218 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1219 return true; 1220 } 1221 if (const StructType *ST = dyn_cast<StructType>(T)) { 1222 NumElts = ST->getNumContainedTypes(); 1223 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1224 for (unsigned n = 1; n < NumElts; ++n) { 1225 if (ST->getContainedType(n) != EltTy) 1226 return false; 1227 } 1228 return true; 1229 } 1230 return false; 1231} 1232 1233/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1234/// "homogeneous" aggregates with the same element type and number of elements. 1235static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1236 if (T1 == T2) 1237 return true; 1238 1239 unsigned NumElts1, NumElts2; 1240 const Type *EltTy1, *EltTy2; 1241 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1242 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1243 NumElts1 == NumElts2 && 1244 EltTy1 == EltTy2) 1245 return true; 1246 1247 return false; 1248} 1249 1250/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1251/// alloca or has an offset and size that corresponds to a component element 1252/// within it. The offset checked here may have been formed from a GEP with a 1253/// pointer bitcasted to a different type. 1254/// 1255/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1256/// unit. If false, it only allows accesses known to be in a single element. 1257void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1258 const Type *MemOpType, bool isStore, 1259 AllocaInfo &Info, Instruction *TheAccess, 1260 bool AllowWholeAccess) { 1261 // Check if this is a load/store of the entire alloca. 1262 if (Offset == 0 && AllowWholeAccess && 1263 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1264 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1265 // loads/stores (which are essentially the same as the MemIntrinsics with 1266 // regard to copying padding between elements). But, if an alloca is 1267 // flagged as both a source and destination of such operations, we'll need 1268 // to check later for padding between elements. 1269 if (!MemOpType || MemOpType->isIntegerTy()) { 1270 if (isStore) 1271 Info.isMemCpyDst = true; 1272 else 1273 Info.isMemCpySrc = true; 1274 return; 1275 } 1276 // This is also safe for references using a type that is compatible with 1277 // the type of the alloca, so that loads/stores can be rewritten using 1278 // insertvalue/extractvalue. 1279 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1280 Info.hasSubelementAccess = true; 1281 return; 1282 } 1283 } 1284 // Check if the offset/size correspond to a component within the alloca type. 1285 const Type *T = Info.AI->getAllocatedType(); 1286 if (TypeHasComponent(T, Offset, MemSize)) { 1287 Info.hasSubelementAccess = true; 1288 return; 1289 } 1290 1291 return MarkUnsafe(Info, TheAccess); 1292} 1293 1294/// TypeHasComponent - Return true if T has a component type with the 1295/// specified offset and size. If Size is zero, do not check the size. 1296bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1297 const Type *EltTy; 1298 uint64_t EltSize; 1299 if (const StructType *ST = dyn_cast<StructType>(T)) { 1300 const StructLayout *Layout = TD->getStructLayout(ST); 1301 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1302 EltTy = ST->getContainedType(EltIdx); 1303 EltSize = TD->getTypeAllocSize(EltTy); 1304 Offset -= Layout->getElementOffset(EltIdx); 1305 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1306 EltTy = AT->getElementType(); 1307 EltSize = TD->getTypeAllocSize(EltTy); 1308 if (Offset >= AT->getNumElements() * EltSize) 1309 return false; 1310 Offset %= EltSize; 1311 } else { 1312 return false; 1313 } 1314 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1315 return true; 1316 // Check if the component spans multiple elements. 1317 if (Offset + Size > EltSize) 1318 return false; 1319 return TypeHasComponent(EltTy, Offset, Size); 1320} 1321 1322/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1323/// the instruction I, which references it, to use the separate elements. 1324/// Offset indicates the position within AI that is referenced by this 1325/// instruction. 1326void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1327 SmallVector<AllocaInst*, 32> &NewElts) { 1328 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1329 Use &TheUse = UI.getUse(); 1330 Instruction *User = cast<Instruction>(*UI++); 1331 1332 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1333 RewriteBitCast(BC, AI, Offset, NewElts); 1334 continue; 1335 } 1336 1337 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1338 RewriteGEP(GEPI, AI, Offset, NewElts); 1339 continue; 1340 } 1341 1342 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1343 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1344 uint64_t MemSize = Length->getZExtValue(); 1345 if (Offset == 0 && 1346 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1347 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1348 // Otherwise the intrinsic can only touch a single element and the 1349 // address operand will be updated, so nothing else needs to be done. 1350 continue; 1351 } 1352 1353 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1354 const Type *LIType = LI->getType(); 1355 1356 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1357 // Replace: 1358 // %res = load { i32, i32 }* %alloc 1359 // with: 1360 // %load.0 = load i32* %alloc.0 1361 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1362 // %load.1 = load i32* %alloc.1 1363 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1364 // (Also works for arrays instead of structs) 1365 Value *Insert = UndefValue::get(LIType); 1366 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1367 Value *Load = new LoadInst(NewElts[i], "load", LI); 1368 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1369 } 1370 LI->replaceAllUsesWith(Insert); 1371 DeadInsts.push_back(LI); 1372 } else if (LIType->isIntegerTy() && 1373 TD->getTypeAllocSize(LIType) == 1374 TD->getTypeAllocSize(AI->getAllocatedType())) { 1375 // If this is a load of the entire alloca to an integer, rewrite it. 1376 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1377 } 1378 continue; 1379 } 1380 1381 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1382 Value *Val = SI->getOperand(0); 1383 const Type *SIType = Val->getType(); 1384 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1385 // Replace: 1386 // store { i32, i32 } %val, { i32, i32 }* %alloc 1387 // with: 1388 // %val.0 = extractvalue { i32, i32 } %val, 0 1389 // store i32 %val.0, i32* %alloc.0 1390 // %val.1 = extractvalue { i32, i32 } %val, 1 1391 // store i32 %val.1, i32* %alloc.1 1392 // (Also works for arrays instead of structs) 1393 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1394 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1395 new StoreInst(Extract, NewElts[i], SI); 1396 } 1397 DeadInsts.push_back(SI); 1398 } else if (SIType->isIntegerTy() && 1399 TD->getTypeAllocSize(SIType) == 1400 TD->getTypeAllocSize(AI->getAllocatedType())) { 1401 // If this is a store of the entire alloca from an integer, rewrite it. 1402 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1403 } 1404 continue; 1405 } 1406 1407 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1408 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1409 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1410 // the new pointer. 1411 if (!isa<AllocaInst>(I)) continue; 1412 1413 assert(Offset == 0 && NewElts[0] && 1414 "Direct alloca use should have a zero offset"); 1415 1416 // If we have a use of the alloca, we know the derived uses will be 1417 // utilizing just the first element of the scalarized result. Insert a 1418 // bitcast of the first alloca before the user as required. 1419 AllocaInst *NewAI = NewElts[0]; 1420 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1421 NewAI->moveBefore(BCI); 1422 TheUse = BCI; 1423 continue; 1424 } 1425 } 1426} 1427 1428/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1429/// and recursively continue updating all of its uses. 1430void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1431 SmallVector<AllocaInst*, 32> &NewElts) { 1432 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1433 if (BC->getOperand(0) != AI) 1434 return; 1435 1436 // The bitcast references the original alloca. Replace its uses with 1437 // references to the first new element alloca. 1438 Instruction *Val = NewElts[0]; 1439 if (Val->getType() != BC->getDestTy()) { 1440 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1441 Val->takeName(BC); 1442 } 1443 BC->replaceAllUsesWith(Val); 1444 DeadInsts.push_back(BC); 1445} 1446 1447/// FindElementAndOffset - Return the index of the element containing Offset 1448/// within the specified type, which must be either a struct or an array. 1449/// Sets T to the type of the element and Offset to the offset within that 1450/// element. IdxTy is set to the type of the index result to be used in a 1451/// GEP instruction. 1452uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1453 const Type *&IdxTy) { 1454 uint64_t Idx = 0; 1455 if (const StructType *ST = dyn_cast<StructType>(T)) { 1456 const StructLayout *Layout = TD->getStructLayout(ST); 1457 Idx = Layout->getElementContainingOffset(Offset); 1458 T = ST->getContainedType(Idx); 1459 Offset -= Layout->getElementOffset(Idx); 1460 IdxTy = Type::getInt32Ty(T->getContext()); 1461 return Idx; 1462 } 1463 const ArrayType *AT = cast<ArrayType>(T); 1464 T = AT->getElementType(); 1465 uint64_t EltSize = TD->getTypeAllocSize(T); 1466 Idx = Offset / EltSize; 1467 Offset -= Idx * EltSize; 1468 IdxTy = Type::getInt64Ty(T->getContext()); 1469 return Idx; 1470} 1471 1472/// RewriteGEP - Check if this GEP instruction moves the pointer across 1473/// elements of the alloca that are being split apart, and if so, rewrite 1474/// the GEP to be relative to the new element. 1475void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1476 SmallVector<AllocaInst*, 32> &NewElts) { 1477 uint64_t OldOffset = Offset; 1478 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1479 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1480 &Indices[0], Indices.size()); 1481 1482 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1483 1484 const Type *T = AI->getAllocatedType(); 1485 const Type *IdxTy; 1486 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1487 if (GEPI->getOperand(0) == AI) 1488 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1489 1490 T = AI->getAllocatedType(); 1491 uint64_t EltOffset = Offset; 1492 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1493 1494 // If this GEP does not move the pointer across elements of the alloca 1495 // being split, then it does not needs to be rewritten. 1496 if (Idx == OldIdx) 1497 return; 1498 1499 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1500 SmallVector<Value*, 8> NewArgs; 1501 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1502 while (EltOffset != 0) { 1503 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1504 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1505 } 1506 Instruction *Val = NewElts[Idx]; 1507 if (NewArgs.size() > 1) { 1508 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1509 NewArgs.end(), "", GEPI); 1510 Val->takeName(GEPI); 1511 } 1512 if (Val->getType() != GEPI->getType()) 1513 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1514 GEPI->replaceAllUsesWith(Val); 1515 DeadInsts.push_back(GEPI); 1516} 1517 1518/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1519/// Rewrite it to copy or set the elements of the scalarized memory. 1520void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1521 AllocaInst *AI, 1522 SmallVector<AllocaInst*, 32> &NewElts) { 1523 // If this is a memcpy/memmove, construct the other pointer as the 1524 // appropriate type. The "Other" pointer is the pointer that goes to memory 1525 // that doesn't have anything to do with the alloca that we are promoting. For 1526 // memset, this Value* stays null. 1527 Value *OtherPtr = 0; 1528 unsigned MemAlignment = MI->getAlignment(); 1529 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1530 if (Inst == MTI->getRawDest()) 1531 OtherPtr = MTI->getRawSource(); 1532 else { 1533 assert(Inst == MTI->getRawSource()); 1534 OtherPtr = MTI->getRawDest(); 1535 } 1536 } 1537 1538 // If there is an other pointer, we want to convert it to the same pointer 1539 // type as AI has, so we can GEP through it safely. 1540 if (OtherPtr) { 1541 unsigned AddrSpace = 1542 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 1543 1544 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1545 // optimization, but it's also required to detect the corner case where 1546 // both pointer operands are referencing the same memory, and where 1547 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1548 // function is only called for mem intrinsics that access the whole 1549 // aggregate, so non-zero GEPs are not an issue here.) 1550 OtherPtr = OtherPtr->stripPointerCasts(); 1551 1552 // Copying the alloca to itself is a no-op: just delete it. 1553 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1554 // This code will run twice for a no-op memcpy -- once for each operand. 1555 // Put only one reference to MI on the DeadInsts list. 1556 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1557 E = DeadInsts.end(); I != E; ++I) 1558 if (*I == MI) return; 1559 DeadInsts.push_back(MI); 1560 return; 1561 } 1562 1563 // If the pointer is not the right type, insert a bitcast to the right 1564 // type. 1565 const Type *NewTy = 1566 PointerType::get(AI->getType()->getElementType(), AddrSpace); 1567 1568 if (OtherPtr->getType() != NewTy) 1569 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 1570 } 1571 1572 // Process each element of the aggregate. 1573 bool SROADest = MI->getRawDest() == Inst; 1574 1575 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1576 1577 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1578 // If this is a memcpy/memmove, emit a GEP of the other element address. 1579 Value *OtherElt = 0; 1580 unsigned OtherEltAlign = MemAlignment; 1581 1582 if (OtherPtr) { 1583 Value *Idx[2] = { Zero, 1584 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1585 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1586 OtherPtr->getName()+"."+Twine(i), 1587 MI); 1588 uint64_t EltOffset; 1589 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1590 const Type *OtherTy = OtherPtrTy->getElementType(); 1591 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1592 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1593 } else { 1594 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 1595 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1596 } 1597 1598 // The alignment of the other pointer is the guaranteed alignment of the 1599 // element, which is affected by both the known alignment of the whole 1600 // mem intrinsic and the alignment of the element. If the alignment of 1601 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 1602 // known alignment is just 4 bytes. 1603 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 1604 } 1605 1606 Value *EltPtr = NewElts[i]; 1607 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 1608 1609 // If we got down to a scalar, insert a load or store as appropriate. 1610 if (EltTy->isSingleValueType()) { 1611 if (isa<MemTransferInst>(MI)) { 1612 if (SROADest) { 1613 // From Other to Alloca. 1614 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 1615 new StoreInst(Elt, EltPtr, MI); 1616 } else { 1617 // From Alloca to Other. 1618 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 1619 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 1620 } 1621 continue; 1622 } 1623 assert(isa<MemSetInst>(MI)); 1624 1625 // If the stored element is zero (common case), just store a null 1626 // constant. 1627 Constant *StoreVal; 1628 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 1629 if (CI->isZero()) { 1630 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 1631 } else { 1632 // If EltTy is a vector type, get the element type. 1633 const Type *ValTy = EltTy->getScalarType(); 1634 1635 // Construct an integer with the right value. 1636 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 1637 APInt OneVal(EltSize, CI->getZExtValue()); 1638 APInt TotalVal(OneVal); 1639 // Set each byte. 1640 for (unsigned i = 0; 8*i < EltSize; ++i) { 1641 TotalVal = TotalVal.shl(8); 1642 TotalVal |= OneVal; 1643 } 1644 1645 // Convert the integer value to the appropriate type. 1646 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 1647 if (ValTy->isPointerTy()) 1648 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 1649 else if (ValTy->isFloatingPointTy()) 1650 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 1651 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 1652 1653 // If the requested value was a vector constant, create it. 1654 if (EltTy != ValTy) { 1655 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 1656 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 1657 StoreVal = ConstantVector::get(&Elts[0], NumElts); 1658 } 1659 } 1660 new StoreInst(StoreVal, EltPtr, MI); 1661 continue; 1662 } 1663 // Otherwise, if we're storing a byte variable, use a memset call for 1664 // this element. 1665 } 1666 1667 unsigned EltSize = TD->getTypeAllocSize(EltTy); 1668 1669 IRBuilder<> Builder(MI); 1670 1671 // Finally, insert the meminst for this element. 1672 if (isa<MemSetInst>(MI)) { 1673 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 1674 MI->isVolatile()); 1675 } else { 1676 assert(isa<MemTransferInst>(MI)); 1677 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 1678 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 1679 1680 if (isa<MemCpyInst>(MI)) 1681 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 1682 else 1683 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 1684 } 1685 } 1686 DeadInsts.push_back(MI); 1687} 1688 1689/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 1690/// overwrites the entire allocation. Extract out the pieces of the stored 1691/// integer and store them individually. 1692void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 1693 SmallVector<AllocaInst*, 32> &NewElts){ 1694 // Extract each element out of the integer according to its structure offset 1695 // and store the element value to the individual alloca. 1696 Value *SrcVal = SI->getOperand(0); 1697 const Type *AllocaEltTy = AI->getAllocatedType(); 1698 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1699 1700 IRBuilder<> Builder(SI); 1701 1702 // Handle tail padding by extending the operand 1703 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 1704 SrcVal = Builder.CreateZExt(SrcVal, 1705 IntegerType::get(SI->getContext(), AllocaSizeBits)); 1706 1707 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 1708 << '\n'); 1709 1710 // There are two forms here: AI could be an array or struct. Both cases 1711 // have different ways to compute the element offset. 1712 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1713 const StructLayout *Layout = TD->getStructLayout(EltSTy); 1714 1715 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1716 // Get the number of bits to shift SrcVal to get the value. 1717 const Type *FieldTy = EltSTy->getElementType(i); 1718 uint64_t Shift = Layout->getElementOffsetInBits(i); 1719 1720 if (TD->isBigEndian()) 1721 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 1722 1723 Value *EltVal = SrcVal; 1724 if (Shift) { 1725 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1726 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 1727 } 1728 1729 // Truncate down to an integer of the right size. 1730 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1731 1732 // Ignore zero sized fields like {}, they obviously contain no data. 1733 if (FieldSizeBits == 0) continue; 1734 1735 if (FieldSizeBits != AllocaSizeBits) 1736 EltVal = Builder.CreateTrunc(EltVal, 1737 IntegerType::get(SI->getContext(), FieldSizeBits)); 1738 Value *DestField = NewElts[i]; 1739 if (EltVal->getType() == FieldTy) { 1740 // Storing to an integer field of this size, just do it. 1741 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 1742 // Bitcast to the right element type (for fp/vector values). 1743 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 1744 } else { 1745 // Otherwise, bitcast the dest pointer (for aggregates). 1746 DestField = Builder.CreateBitCast(DestField, 1747 PointerType::getUnqual(EltVal->getType())); 1748 } 1749 new StoreInst(EltVal, DestField, SI); 1750 } 1751 1752 } else { 1753 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1754 const Type *ArrayEltTy = ATy->getElementType(); 1755 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1756 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1757 1758 uint64_t Shift; 1759 1760 if (TD->isBigEndian()) 1761 Shift = AllocaSizeBits-ElementOffset; 1762 else 1763 Shift = 0; 1764 1765 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1766 // Ignore zero sized fields like {}, they obviously contain no data. 1767 if (ElementSizeBits == 0) continue; 1768 1769 Value *EltVal = SrcVal; 1770 if (Shift) { 1771 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1772 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 1773 } 1774 1775 // Truncate down to an integer of the right size. 1776 if (ElementSizeBits != AllocaSizeBits) 1777 EltVal = Builder.CreateTrunc(EltVal, 1778 IntegerType::get(SI->getContext(), 1779 ElementSizeBits)); 1780 Value *DestField = NewElts[i]; 1781 if (EltVal->getType() == ArrayEltTy) { 1782 // Storing to an integer field of this size, just do it. 1783 } else if (ArrayEltTy->isFloatingPointTy() || 1784 ArrayEltTy->isVectorTy()) { 1785 // Bitcast to the right element type (for fp/vector values). 1786 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 1787 } else { 1788 // Otherwise, bitcast the dest pointer (for aggregates). 1789 DestField = Builder.CreateBitCast(DestField, 1790 PointerType::getUnqual(EltVal->getType())); 1791 } 1792 new StoreInst(EltVal, DestField, SI); 1793 1794 if (TD->isBigEndian()) 1795 Shift -= ElementOffset; 1796 else 1797 Shift += ElementOffset; 1798 } 1799 } 1800 1801 DeadInsts.push_back(SI); 1802} 1803 1804/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1805/// an integer. Load the individual pieces to form the aggregate value. 1806void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1807 SmallVector<AllocaInst*, 32> &NewElts) { 1808 // Extract each element out of the NewElts according to its structure offset 1809 // and form the result value. 1810 const Type *AllocaEltTy = AI->getAllocatedType(); 1811 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1812 1813 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1814 << '\n'); 1815 1816 // There are two forms here: AI could be an array or struct. Both cases 1817 // have different ways to compute the element offset. 1818 const StructLayout *Layout = 0; 1819 uint64_t ArrayEltBitOffset = 0; 1820 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1821 Layout = TD->getStructLayout(EltSTy); 1822 } else { 1823 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1824 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1825 } 1826 1827 Value *ResultVal = 1828 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1829 1830 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1831 // Load the value from the alloca. If the NewElt is an aggregate, cast 1832 // the pointer to an integer of the same size before doing the load. 1833 Value *SrcField = NewElts[i]; 1834 const Type *FieldTy = 1835 cast<PointerType>(SrcField->getType())->getElementType(); 1836 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1837 1838 // Ignore zero sized fields like {}, they obviously contain no data. 1839 if (FieldSizeBits == 0) continue; 1840 1841 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1842 FieldSizeBits); 1843 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1844 !FieldTy->isVectorTy()) 1845 SrcField = new BitCastInst(SrcField, 1846 PointerType::getUnqual(FieldIntTy), 1847 "", LI); 1848 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1849 1850 // If SrcField is a fp or vector of the right size but that isn't an 1851 // integer type, bitcast to an integer so we can shift it. 1852 if (SrcField->getType() != FieldIntTy) 1853 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1854 1855 // Zero extend the field to be the same size as the final alloca so that 1856 // we can shift and insert it. 1857 if (SrcField->getType() != ResultVal->getType()) 1858 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1859 1860 // Determine the number of bits to shift SrcField. 1861 uint64_t Shift; 1862 if (Layout) // Struct case. 1863 Shift = Layout->getElementOffsetInBits(i); 1864 else // Array case. 1865 Shift = i*ArrayEltBitOffset; 1866 1867 if (TD->isBigEndian()) 1868 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1869 1870 if (Shift) { 1871 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1872 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1873 } 1874 1875 // Don't create an 'or x, 0' on the first iteration. 1876 if (!isa<Constant>(ResultVal) || 1877 !cast<Constant>(ResultVal)->isNullValue()) 1878 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1879 else 1880 ResultVal = SrcField; 1881 } 1882 1883 // Handle tail padding by truncating the result 1884 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1885 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1886 1887 LI->replaceAllUsesWith(ResultVal); 1888 DeadInsts.push_back(LI); 1889} 1890 1891/// HasPadding - Return true if the specified type has any structure or 1892/// alignment padding in between the elements that would be split apart 1893/// by SROA; return false otherwise. 1894static bool HasPadding(const Type *Ty, const TargetData &TD) { 1895 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1896 Ty = ATy->getElementType(); 1897 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1898 } 1899 1900 // SROA currently handles only Arrays and Structs. 1901 const StructType *STy = cast<StructType>(Ty); 1902 const StructLayout *SL = TD.getStructLayout(STy); 1903 unsigned PrevFieldBitOffset = 0; 1904 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1905 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1906 1907 // Check to see if there is any padding between this element and the 1908 // previous one. 1909 if (i) { 1910 unsigned PrevFieldEnd = 1911 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1912 if (PrevFieldEnd < FieldBitOffset) 1913 return true; 1914 } 1915 PrevFieldBitOffset = FieldBitOffset; 1916 } 1917 // Check for tail padding. 1918 if (unsigned EltCount = STy->getNumElements()) { 1919 unsigned PrevFieldEnd = PrevFieldBitOffset + 1920 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1921 if (PrevFieldEnd < SL->getSizeInBits()) 1922 return true; 1923 } 1924 return false; 1925} 1926 1927/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1928/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1929/// or 1 if safe after canonicalization has been performed. 1930bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1931 // Loop over the use list of the alloca. We can only transform it if all of 1932 // the users are safe to transform. 1933 AllocaInfo Info(AI); 1934 1935 isSafeForScalarRepl(AI, 0, Info); 1936 if (Info.isUnsafe) { 1937 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1938 return false; 1939 } 1940 1941 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1942 // source and destination, we have to be careful. In particular, the memcpy 1943 // could be moving around elements that live in structure padding of the LLVM 1944 // types, but may actually be used. In these cases, we refuse to promote the 1945 // struct. 1946 if (Info.isMemCpySrc && Info.isMemCpyDst && 1947 HasPadding(AI->getAllocatedType(), *TD)) 1948 return false; 1949 1950 // If the alloca never has an access to just *part* of it, but is accessed 1951 // via loads and stores, then we should use ConvertToScalarInfo to promote 1952 // the alloca instead of promoting each piece at a time and inserting fission 1953 // and fusion code. 1954 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 1955 // If the struct/array just has one element, use basic SRoA. 1956 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1957 if (ST->getNumElements() > 1) return false; 1958 } else { 1959 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 1960 return false; 1961 } 1962 } 1963 1964 return true; 1965} 1966 1967 1968 1969/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1970/// some part of a constant global variable. This intentionally only accepts 1971/// constant expressions because we don't can't rewrite arbitrary instructions. 1972static bool PointsToConstantGlobal(Value *V) { 1973 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1974 return GV->isConstant(); 1975 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1976 if (CE->getOpcode() == Instruction::BitCast || 1977 CE->getOpcode() == Instruction::GetElementPtr) 1978 return PointsToConstantGlobal(CE->getOperand(0)); 1979 return false; 1980} 1981 1982/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1983/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1984/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1985/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1986/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1987/// the alloca, and if the source pointer is a pointer to a constant global, we 1988/// can optimize this. 1989static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 1990 bool isOffset) { 1991 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1992 User *U = cast<Instruction>(*UI); 1993 1994 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1995 // Ignore non-volatile loads, they are always ok. 1996 if (LI->isVolatile()) return false; 1997 continue; 1998 } 1999 2000 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2001 // If uses of the bitcast are ok, we are ok. 2002 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 2003 return false; 2004 continue; 2005 } 2006 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2007 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2008 // doesn't, it does. 2009 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2010 isOffset || !GEP->hasAllZeroIndices())) 2011 return false; 2012 continue; 2013 } 2014 2015 if (CallSite CS = U) { 2016 // If this is a readonly/readnone call site, then we know it is just a 2017 // load and we can ignore it. 2018 if (CS.onlyReadsMemory()) 2019 continue; 2020 2021 // If this is the function being called then we treat it like a load and 2022 // ignore it. 2023 if (CS.isCallee(UI)) 2024 continue; 2025 2026 // If this is being passed as a byval argument, the caller is making a 2027 // copy, so it is only a read of the alloca. 2028 unsigned ArgNo = CS.getArgumentNo(UI); 2029 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2030 continue; 2031 } 2032 2033 // If this is isn't our memcpy/memmove, reject it as something we can't 2034 // handle. 2035 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2036 if (MI == 0) 2037 return false; 2038 2039 // If the transfer is using the alloca as a source of the transfer, then 2040 // ignore it since it is a load (unless the transfer is volatile). 2041 if (UI.getOperandNo() == 1) { 2042 if (MI->isVolatile()) return false; 2043 continue; 2044 } 2045 2046 // If we already have seen a copy, reject the second one. 2047 if (TheCopy) return false; 2048 2049 // If the pointer has been offset from the start of the alloca, we can't 2050 // safely handle this. 2051 if (isOffset) return false; 2052 2053 // If the memintrinsic isn't using the alloca as the dest, reject it. 2054 if (UI.getOperandNo() != 0) return false; 2055 2056 // If the source of the memcpy/move is not a constant global, reject it. 2057 if (!PointsToConstantGlobal(MI->getSource())) 2058 return false; 2059 2060 // Otherwise, the transform is safe. Remember the copy instruction. 2061 TheCopy = MI; 2062 } 2063 return true; 2064} 2065 2066/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2067/// modified by a copy from a constant global. If we can prove this, we can 2068/// replace any uses of the alloca with uses of the global directly. 2069MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 2070 MemTransferInst *TheCopy = 0; 2071 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 2072 return TheCopy; 2073 return 0; 2074} 2075