ScalarReplAggregates.cpp revision 70728532799d751b8e0e97719dcb3344a2fc97de
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 AllocaInfo() 92 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false) {} 93 }; 94 95 unsigned SRThreshold; 96 97 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 98 99 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 100 101 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 102 AllocaInfo &Info); 103 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset, 104 AllocaInfo &Info); 105 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 106 const Type *MemOpType, bool isStore, AllocaInfo &Info); 107 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 108 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 109 const Type *&IdxTy); 110 111 void DoScalarReplacement(AllocaInst *AI, 112 std::vector<AllocaInst*> &WorkList); 113 void DeleteDeadInstructions(); 114 115 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 116 SmallVector<AllocaInst*, 32> &NewElts); 117 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 118 SmallVector<AllocaInst*, 32> &NewElts); 119 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 120 SmallVector<AllocaInst*, 32> &NewElts); 121 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 122 AllocaInst *AI, 123 SmallVector<AllocaInst*, 32> &NewElts); 124 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 125 SmallVector<AllocaInst*, 32> &NewElts); 126 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 127 SmallVector<AllocaInst*, 32> &NewElts); 128 129 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 130 }; 131 132 // SROA_DF - SROA that uses DominanceFrontier. 133 struct SROA_DF : public SROA { 134 static char ID; 135 public: 136 SROA_DF(int T = -1) : SROA(T, true, ID) { 137 initializeSROA_DFPass(*PassRegistry::getPassRegistry()); 138 } 139 140 // getAnalysisUsage - This pass does not require any passes, but we know it 141 // will not alter the CFG, so say so. 142 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 143 AU.addRequired<DominatorTree>(); 144 AU.addRequired<DominanceFrontier>(); 145 AU.setPreservesCFG(); 146 } 147 }; 148 149 // SROA_SSAUp - SROA that uses SSAUpdater. 150 struct SROA_SSAUp : public SROA { 151 static char ID; 152 public: 153 SROA_SSAUp(int T = -1) : SROA(T, false, ID) { 154 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 155 } 156 157 // getAnalysisUsage - This pass does not require any passes, but we know it 158 // will not alter the CFG, so say so. 159 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 160 AU.setPreservesCFG(); 161 } 162 }; 163 164} 165 166char SROA_DF::ID = 0; 167char SROA_SSAUp::ID = 0; 168 169INITIALIZE_PASS_BEGIN(SROA_DF, "scalarrepl", 170 "Scalar Replacement of Aggregates (DF)", false, false) 171INITIALIZE_PASS_DEPENDENCY(DominatorTree) 172INITIALIZE_PASS_DEPENDENCY(DominanceFrontier) 173INITIALIZE_PASS_END(SROA_DF, "scalarrepl", 174 "Scalar Replacement of Aggregates (DF)", false, false) 175 176INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 177 "Scalar Replacement of Aggregates (SSAUp)", false, false) 178INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 179 "Scalar Replacement of Aggregates (SSAUp)", false, false) 180 181// Public interface to the ScalarReplAggregates pass 182FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 183 bool UseDomFrontier) { 184 if (UseDomFrontier) 185 return new SROA_DF(Threshold); 186 return new SROA_SSAUp(Threshold); 187} 188 189 190//===----------------------------------------------------------------------===// 191// Convert To Scalar Optimization. 192//===----------------------------------------------------------------------===// 193 194namespace { 195/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 196/// optimization, which scans the uses of an alloca and determines if it can 197/// rewrite it in terms of a single new alloca that can be mem2reg'd. 198class ConvertToScalarInfo { 199 /// AllocaSize - The size of the alloca being considered. 200 unsigned AllocaSize; 201 const TargetData &TD; 202 203 /// IsNotTrivial - This is set to true if there is some access to the object 204 /// which means that mem2reg can't promote it. 205 bool IsNotTrivial; 206 207 /// VectorTy - This tracks the type that we should promote the vector to if 208 /// it is possible to turn it into a vector. This starts out null, and if it 209 /// isn't possible to turn into a vector type, it gets set to VoidTy. 210 const Type *VectorTy; 211 212 /// HadAVector - True if there is at least one vector access to the alloca. 213 /// We don't want to turn random arrays into vectors and use vector element 214 /// insert/extract, but if there are element accesses to something that is 215 /// also declared as a vector, we do want to promote to a vector. 216 bool HadAVector; 217 218public: 219 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 220 : AllocaSize(Size), TD(td) { 221 IsNotTrivial = false; 222 VectorTy = 0; 223 HadAVector = false; 224 } 225 226 AllocaInst *TryConvert(AllocaInst *AI); 227 228private: 229 bool CanConvertToScalar(Value *V, uint64_t Offset); 230 void MergeInType(const Type *In, uint64_t Offset); 231 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 232 233 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 234 uint64_t Offset, IRBuilder<> &Builder); 235 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 236 uint64_t Offset, IRBuilder<> &Builder); 237}; 238} // end anonymous namespace. 239 240 241/// IsVerbotenVectorType - Return true if this is a vector type ScalarRepl isn't 242/// allowed to form. We do this to avoid MMX types, which is a complete hack, 243/// but is required until the backend is fixed. 244static bool IsVerbotenVectorType(const VectorType *VTy, const Instruction *I) { 245 StringRef Triple(I->getParent()->getParent()->getParent()->getTargetTriple()); 246 if (!Triple.startswith("i386") && 247 !Triple.startswith("x86_64")) 248 return false; 249 250 // Reject all the MMX vector types. 251 switch (VTy->getNumElements()) { 252 default: return false; 253 case 1: return VTy->getElementType()->isIntegerTy(64); 254 case 2: return VTy->getElementType()->isIntegerTy(32); 255 case 4: return VTy->getElementType()->isIntegerTy(16); 256 case 8: return VTy->getElementType()->isIntegerTy(8); 257 } 258} 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 !IsVerbotenVectorType(cast<VectorType>(VectorTy), AI)) { 279 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 280 << *VectorTy << '\n'); 281 NewTy = VectorTy; // Use the vector type. 282 } else { 283 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 284 // Create and insert the integer alloca. 285 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 286 } 287 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 288 ConvertUsesToScalar(AI, NewAI, 0); 289 return NewAI; 290} 291 292/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy) 293/// so far at the offset specified by Offset (which is specified in bytes). 294/// 295/// There are two cases we handle here: 296/// 1) A union of vector types of the same size and potentially its elements. 297/// Here we turn element accesses into insert/extract element operations. 298/// This promotes a <4 x float> with a store of float to the third element 299/// into a <4 x float> that uses insert element. 300/// 2) A fully general blob of memory, which we turn into some (potentially 301/// large) integer type with extract and insert operations where the loads 302/// and stores would mutate the memory. We mark this by setting VectorTy 303/// to VoidTy. 304void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) { 305 // If we already decided to turn this into a blob of integer memory, there is 306 // nothing to be done. 307 if (VectorTy && VectorTy->isVoidTy()) 308 return; 309 310 // If this could be contributing to a vector, analyze it. 311 312 // If the In type is a vector that is the same size as the alloca, see if it 313 // matches the existing VecTy. 314 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 315 // Remember if we saw a vector type. 316 HadAVector = true; 317 318 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 319 // If we're storing/loading a vector of the right size, allow it as a 320 // vector. If this the first vector we see, remember the type so that 321 // we know the element size. If this is a subsequent access, ignore it 322 // even if it is a differing type but the same size. Worst case we can 323 // bitcast the resultant vectors. 324 if (VectorTy == 0) 325 VectorTy = VInTy; 326 return; 327 } 328 } else if (In->isFloatTy() || In->isDoubleTy() || 329 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 330 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 331 // If we're accessing something that could be an element of a vector, see 332 // if the implied vector agrees with what we already have and if Offset is 333 // compatible with it. 334 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 335 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 336 (VectorTy == 0 || 337 cast<VectorType>(VectorTy)->getElementType() 338 ->getPrimitiveSizeInBits()/8 == EltSize)) { 339 if (VectorTy == 0) 340 VectorTy = VectorType::get(In, AllocaSize/EltSize); 341 return; 342 } 343 } 344 345 // Otherwise, we have a case that we can't handle with an optimized vector 346 // form. We can still turn this into a large integer. 347 VectorTy = Type::getVoidTy(In->getContext()); 348} 349 350/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 351/// its accesses to a single vector type, return true and set VecTy to 352/// the new type. If we could convert the alloca into a single promotable 353/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 354/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 355/// is the current offset from the base of the alloca being analyzed. 356/// 357/// If we see at least one access to the value that is as a vector type, set the 358/// SawVec flag. 359bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 360 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 361 Instruction *User = cast<Instruction>(*UI); 362 363 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 364 // Don't break volatile loads. 365 if (LI->isVolatile()) 366 return false; 367 // Don't touch MMX operations. 368 if (LI->getType()->isX86_MMXTy()) 369 return false; 370 MergeInType(LI->getType(), Offset); 371 continue; 372 } 373 374 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 375 // Storing the pointer, not into the value? 376 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 377 // Don't touch MMX operations. 378 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 379 return false; 380 MergeInType(SI->getOperand(0)->getType(), Offset); 381 continue; 382 } 383 384 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 385 IsNotTrivial = true; // Can't be mem2reg'd. 386 if (!CanConvertToScalar(BCI, Offset)) 387 return false; 388 continue; 389 } 390 391 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 392 // If this is a GEP with a variable indices, we can't handle it. 393 if (!GEP->hasAllConstantIndices()) 394 return false; 395 396 // Compute the offset that this GEP adds to the pointer. 397 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 398 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 399 &Indices[0], Indices.size()); 400 // See if all uses can be converted. 401 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 402 return false; 403 IsNotTrivial = true; // Can't be mem2reg'd. 404 continue; 405 } 406 407 // If this is a constant sized memset of a constant value (e.g. 0) we can 408 // handle it. 409 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 410 // Store of constant value and constant size. 411 if (!isa<ConstantInt>(MSI->getValue()) || 412 !isa<ConstantInt>(MSI->getLength())) 413 return false; 414 IsNotTrivial = true; // Can't be mem2reg'd. 415 continue; 416 } 417 418 // If this is a memcpy or memmove into or out of the whole allocation, we 419 // can handle it like a load or store of the scalar type. 420 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 421 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 422 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 423 return false; 424 425 IsNotTrivial = true; // Can't be mem2reg'd. 426 continue; 427 } 428 429 // Otherwise, we cannot handle this! 430 return false; 431 } 432 433 return true; 434} 435 436/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 437/// directly. This happens when we are converting an "integer union" to a 438/// single integer scalar, or when we are converting a "vector union" to a 439/// vector with insert/extractelement instructions. 440/// 441/// Offset is an offset from the original alloca, in bits that need to be 442/// shifted to the right. By the end of this, there should be no uses of Ptr. 443void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 444 uint64_t Offset) { 445 while (!Ptr->use_empty()) { 446 Instruction *User = cast<Instruction>(Ptr->use_back()); 447 448 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 449 ConvertUsesToScalar(CI, NewAI, Offset); 450 CI->eraseFromParent(); 451 continue; 452 } 453 454 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 455 // Compute the offset that this GEP adds to the pointer. 456 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 457 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 458 &Indices[0], Indices.size()); 459 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 460 GEP->eraseFromParent(); 461 continue; 462 } 463 464 IRBuilder<> Builder(User); 465 466 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 467 // The load is a bit extract from NewAI shifted right by Offset bits. 468 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 469 Value *NewLoadVal 470 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 471 LI->replaceAllUsesWith(NewLoadVal); 472 LI->eraseFromParent(); 473 continue; 474 } 475 476 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 477 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 478 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 479 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 480 Builder); 481 Builder.CreateStore(New, NewAI); 482 SI->eraseFromParent(); 483 484 // If the load we just inserted is now dead, then the inserted store 485 // overwrote the entire thing. 486 if (Old->use_empty()) 487 Old->eraseFromParent(); 488 continue; 489 } 490 491 // If this is a constant sized memset of a constant value (e.g. 0) we can 492 // transform it into a store of the expanded constant value. 493 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 494 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 495 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 496 if (NumBytes != 0) { 497 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 498 499 // Compute the value replicated the right number of times. 500 APInt APVal(NumBytes*8, Val); 501 502 // Splat the value if non-zero. 503 if (Val) 504 for (unsigned i = 1; i != NumBytes; ++i) 505 APVal |= APVal << 8; 506 507 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 508 Value *New = ConvertScalar_InsertValue( 509 ConstantInt::get(User->getContext(), APVal), 510 Old, Offset, Builder); 511 Builder.CreateStore(New, NewAI); 512 513 // If the load we just inserted is now dead, then the memset overwrote 514 // the entire thing. 515 if (Old->use_empty()) 516 Old->eraseFromParent(); 517 } 518 MSI->eraseFromParent(); 519 continue; 520 } 521 522 // If this is a memcpy or memmove into or out of the whole allocation, we 523 // can handle it like a load or store of the scalar type. 524 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 525 assert(Offset == 0 && "must be store to start of alloca"); 526 527 // If the source and destination are both to the same alloca, then this is 528 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 529 // as appropriate. 530 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, 0)); 531 532 if (GetUnderlyingObject(MTI->getSource(), 0) != OrigAI) { 533 // Dest must be OrigAI, change this to be a load from the original 534 // pointer (bitcasted), then a store to our new alloca. 535 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 536 Value *SrcPtr = MTI->getSource(); 537 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 538 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 539 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 540 AIPTy = PointerType::get(AIPTy->getElementType(), 541 SPTy->getAddressSpace()); 542 } 543 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 544 545 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 546 SrcVal->setAlignment(MTI->getAlignment()); 547 Builder.CreateStore(SrcVal, NewAI); 548 } else if (GetUnderlyingObject(MTI->getDest(), 0) != OrigAI) { 549 // Src must be OrigAI, change this to be a load from NewAI then a store 550 // through the original dest pointer (bitcasted). 551 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 552 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 553 554 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 555 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 556 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 557 AIPTy = PointerType::get(AIPTy->getElementType(), 558 DPTy->getAddressSpace()); 559 } 560 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 561 562 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 563 NewStore->setAlignment(MTI->getAlignment()); 564 } else { 565 // Noop transfer. Src == Dst 566 } 567 568 MTI->eraseFromParent(); 569 continue; 570 } 571 572 llvm_unreachable("Unsupported operation!"); 573 } 574} 575 576/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 577/// or vector value FromVal, extracting the bits from the offset specified by 578/// Offset. This returns the value, which is of type ToType. 579/// 580/// This happens when we are converting an "integer union" to a single 581/// integer scalar, or when we are converting a "vector union" to a vector with 582/// insert/extractelement instructions. 583/// 584/// Offset is an offset from the original alloca, in bits that need to be 585/// shifted to the right. 586Value *ConvertToScalarInfo:: 587ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 588 uint64_t Offset, IRBuilder<> &Builder) { 589 // If the load is of the whole new alloca, no conversion is needed. 590 if (FromVal->getType() == ToType && Offset == 0) 591 return FromVal; 592 593 // If the result alloca is a vector type, this is either an element 594 // access or a bitcast to another vector type of the same size. 595 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 596 if (ToType->isVectorTy()) 597 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 598 599 // Otherwise it must be an element access. 600 unsigned Elt = 0; 601 if (Offset) { 602 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 603 Elt = Offset/EltSize; 604 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 605 } 606 // Return the element extracted out of it. 607 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 608 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 609 if (V->getType() != ToType) 610 V = Builder.CreateBitCast(V, ToType, "tmp"); 611 return V; 612 } 613 614 // If ToType is a first class aggregate, extract out each of the pieces and 615 // use insertvalue's to form the FCA. 616 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 617 const StructLayout &Layout = *TD.getStructLayout(ST); 618 Value *Res = UndefValue::get(ST); 619 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 620 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 621 Offset+Layout.getElementOffsetInBits(i), 622 Builder); 623 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 624 } 625 return Res; 626 } 627 628 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 629 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 630 Value *Res = UndefValue::get(AT); 631 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 632 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 633 Offset+i*EltSize, Builder); 634 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 635 } 636 return Res; 637 } 638 639 // Otherwise, this must be a union that was converted to an integer value. 640 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 641 642 // If this is a big-endian system and the load is narrower than the 643 // full alloca type, we need to do a shift to get the right bits. 644 int ShAmt = 0; 645 if (TD.isBigEndian()) { 646 // On big-endian machines, the lowest bit is stored at the bit offset 647 // from the pointer given by getTypeStoreSizeInBits. This matters for 648 // integers with a bitwidth that is not a multiple of 8. 649 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 650 TD.getTypeStoreSizeInBits(ToType) - Offset; 651 } else { 652 ShAmt = Offset; 653 } 654 655 // Note: we support negative bitwidths (with shl) which are not defined. 656 // We do this to support (f.e.) loads off the end of a structure where 657 // only some bits are used. 658 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 659 FromVal = Builder.CreateLShr(FromVal, 660 ConstantInt::get(FromVal->getType(), 661 ShAmt), "tmp"); 662 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 663 FromVal = Builder.CreateShl(FromVal, 664 ConstantInt::get(FromVal->getType(), 665 -ShAmt), "tmp"); 666 667 // Finally, unconditionally truncate the integer to the right width. 668 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 669 if (LIBitWidth < NTy->getBitWidth()) 670 FromVal = 671 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 672 LIBitWidth), "tmp"); 673 else if (LIBitWidth > NTy->getBitWidth()) 674 FromVal = 675 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 676 LIBitWidth), "tmp"); 677 678 // If the result is an integer, this is a trunc or bitcast. 679 if (ToType->isIntegerTy()) { 680 // Should be done. 681 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 682 // Just do a bitcast, we know the sizes match up. 683 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 684 } else { 685 // Otherwise must be a pointer. 686 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 687 } 688 assert(FromVal->getType() == ToType && "Didn't convert right?"); 689 return FromVal; 690} 691 692/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 693/// or vector value "Old" at the offset specified by Offset. 694/// 695/// This happens when we are converting an "integer union" to a 696/// single integer scalar, or when we are converting a "vector union" to a 697/// vector with insert/extractelement instructions. 698/// 699/// Offset is an offset from the original alloca, in bits that need to be 700/// shifted to the right. 701Value *ConvertToScalarInfo:: 702ConvertScalar_InsertValue(Value *SV, Value *Old, 703 uint64_t Offset, IRBuilder<> &Builder) { 704 // Convert the stored type to the actual type, shift it left to insert 705 // then 'or' into place. 706 const Type *AllocaType = Old->getType(); 707 LLVMContext &Context = Old->getContext(); 708 709 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 710 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 711 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 712 713 // Changing the whole vector with memset or with an access of a different 714 // vector type? 715 if (ValSize == VecSize) 716 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 717 718 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 719 720 // Must be an element insertion. 721 unsigned Elt = Offset/EltSize; 722 723 if (SV->getType() != VTy->getElementType()) 724 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 725 726 SV = Builder.CreateInsertElement(Old, SV, 727 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 728 "tmp"); 729 return SV; 730 } 731 732 // If SV is a first-class aggregate value, insert each value recursively. 733 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 734 const StructLayout &Layout = *TD.getStructLayout(ST); 735 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 736 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 737 Old = ConvertScalar_InsertValue(Elt, Old, 738 Offset+Layout.getElementOffsetInBits(i), 739 Builder); 740 } 741 return Old; 742 } 743 744 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 745 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 746 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 747 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 748 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 749 } 750 return Old; 751 } 752 753 // If SV is a float, convert it to the appropriate integer type. 754 // If it is a pointer, do the same. 755 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 756 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 757 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 758 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 759 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 760 SV = Builder.CreateBitCast(SV, 761 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 762 else if (SV->getType()->isPointerTy()) 763 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 764 765 // Zero extend or truncate the value if needed. 766 if (SV->getType() != AllocaType) { 767 if (SV->getType()->getPrimitiveSizeInBits() < 768 AllocaType->getPrimitiveSizeInBits()) 769 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 770 else { 771 // Truncation may be needed if storing more than the alloca can hold 772 // (undefined behavior). 773 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 774 SrcWidth = DestWidth; 775 SrcStoreWidth = DestStoreWidth; 776 } 777 } 778 779 // If this is a big-endian system and the store is narrower than the 780 // full alloca type, we need to do a shift to get the right bits. 781 int ShAmt = 0; 782 if (TD.isBigEndian()) { 783 // On big-endian machines, the lowest bit is stored at the bit offset 784 // from the pointer given by getTypeStoreSizeInBits. This matters for 785 // integers with a bitwidth that is not a multiple of 8. 786 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 787 } else { 788 ShAmt = Offset; 789 } 790 791 // Note: we support negative bitwidths (with shr) which are not defined. 792 // We do this to support (f.e.) stores off the end of a structure where 793 // only some bits in the structure are set. 794 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 795 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 796 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 797 ShAmt), "tmp"); 798 Mask <<= ShAmt; 799 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 800 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 801 -ShAmt), "tmp"); 802 Mask = Mask.lshr(-ShAmt); 803 } 804 805 // Mask out the bits we are about to insert from the old value, and or 806 // in the new bits. 807 if (SrcWidth != DestWidth) { 808 assert(DestWidth > SrcWidth); 809 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 810 SV = Builder.CreateOr(Old, SV, "ins"); 811 } 812 return SV; 813} 814 815 816//===----------------------------------------------------------------------===// 817// SRoA Driver 818//===----------------------------------------------------------------------===// 819 820 821bool SROA::runOnFunction(Function &F) { 822 TD = getAnalysisIfAvailable<TargetData>(); 823 824 bool Changed = performPromotion(F); 825 826 // FIXME: ScalarRepl currently depends on TargetData more than it 827 // theoretically needs to. It should be refactored in order to support 828 // target-independent IR. Until this is done, just skip the actual 829 // scalar-replacement portion of this pass. 830 if (!TD) return Changed; 831 832 while (1) { 833 bool LocalChange = performScalarRepl(F); 834 if (!LocalChange) break; // No need to repromote if no scalarrepl 835 Changed = true; 836 LocalChange = performPromotion(F); 837 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 838 } 839 840 return Changed; 841} 842 843namespace { 844class AllocaPromoter : public LoadAndStorePromoter { 845 AllocaInst *AI; 846public: 847 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S) 848 : LoadAndStorePromoter(Insts, S), AI(0) {} 849 850 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 851 // Remember which alloca we're promoting (for isInstInList). 852 this->AI = AI; 853 LoadAndStorePromoter::run(Insts); 854 AI->eraseFromParent(); 855 } 856 857 virtual bool isInstInList(Instruction *I, 858 const SmallVectorImpl<Instruction*> &Insts) const { 859 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 860 return LI->getOperand(0) == AI; 861 return cast<StoreInst>(I)->getPointerOperand() == AI; 862 } 863}; 864} // end anon namespace 865 866bool SROA::performPromotion(Function &F) { 867 std::vector<AllocaInst*> Allocas; 868 DominatorTree *DT = 0; 869 DominanceFrontier *DF = 0; 870 if (HasDomFrontiers) { 871 DT = &getAnalysis<DominatorTree>(); 872 DF = &getAnalysis<DominanceFrontier>(); 873 } 874 875 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 876 877 bool Changed = false; 878 SmallVector<Instruction*, 64> Insts; 879 while (1) { 880 Allocas.clear(); 881 882 // Find allocas that are safe to promote, by looking at all instructions in 883 // the entry node 884 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 885 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 886 if (isAllocaPromotable(AI)) 887 Allocas.push_back(AI); 888 889 if (Allocas.empty()) break; 890 891 if (HasDomFrontiers) 892 PromoteMemToReg(Allocas, *DT, *DF); 893 else { 894 SSAUpdater SSA; 895 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 896 AllocaInst *AI = Allocas[i]; 897 898 // Build list of instructions to promote. 899 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 900 UI != E; ++UI) 901 Insts.push_back(cast<Instruction>(*UI)); 902 903 AllocaPromoter(Insts, SSA).run(AI, Insts); 904 Insts.clear(); 905 } 906 } 907 NumPromoted += Allocas.size(); 908 Changed = true; 909 } 910 911 return Changed; 912} 913 914 915/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 916/// SROA. It must be a struct or array type with a small number of elements. 917static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 918 const Type *T = AI->getAllocatedType(); 919 // Do not promote any struct into more than 32 separate vars. 920 if (const StructType *ST = dyn_cast<StructType>(T)) 921 return ST->getNumElements() <= 32; 922 // Arrays are much less likely to be safe for SROA; only consider 923 // them if they are very small. 924 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 925 return AT->getNumElements() <= 8; 926 return false; 927} 928 929 930// performScalarRepl - This algorithm is a simple worklist driven algorithm, 931// which runs on all of the malloc/alloca instructions in the function, removing 932// them if they are only used by getelementptr instructions. 933// 934bool SROA::performScalarRepl(Function &F) { 935 std::vector<AllocaInst*> WorkList; 936 937 // Scan the entry basic block, adding allocas to the worklist. 938 BasicBlock &BB = F.getEntryBlock(); 939 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 940 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 941 WorkList.push_back(A); 942 943 // Process the worklist 944 bool Changed = false; 945 while (!WorkList.empty()) { 946 AllocaInst *AI = WorkList.back(); 947 WorkList.pop_back(); 948 949 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 950 // with unused elements. 951 if (AI->use_empty()) { 952 AI->eraseFromParent(); 953 Changed = true; 954 continue; 955 } 956 957 // If this alloca is impossible for us to promote, reject it early. 958 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 959 continue; 960 961 // Check to see if this allocation is only modified by a memcpy/memmove from 962 // a constant global. If this is the case, we can change all users to use 963 // the constant global instead. This is commonly produced by the CFE by 964 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 965 // is only subsequently read. 966 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 967 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 968 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 969 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 970 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 971 TheCopy->eraseFromParent(); // Don't mutate the global. 972 AI->eraseFromParent(); 973 ++NumGlobals; 974 Changed = true; 975 continue; 976 } 977 978 // Check to see if we can perform the core SROA transformation. We cannot 979 // transform the allocation instruction if it is an array allocation 980 // (allocations OF arrays are ok though), and an allocation of a scalar 981 // value cannot be decomposed at all. 982 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 983 984 // Do not promote [0 x %struct]. 985 if (AllocaSize == 0) continue; 986 987 // Do not promote any struct whose size is too big. 988 if (AllocaSize > SRThreshold) continue; 989 990 // If the alloca looks like a good candidate for scalar replacement, and if 991 // all its users can be transformed, then split up the aggregate into its 992 // separate elements. 993 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 994 DoScalarReplacement(AI, WorkList); 995 Changed = true; 996 continue; 997 } 998 999 // If we can turn this aggregate value (potentially with casts) into a 1000 // simple scalar value that can be mem2reg'd into a register value. 1001 // IsNotTrivial tracks whether this is something that mem2reg could have 1002 // promoted itself. If so, we don't want to transform it needlessly. Note 1003 // that we can't just check based on the type: the alloca may be of an i32 1004 // but that has pointer arithmetic to set byte 3 of it or something. 1005 if (AllocaInst *NewAI = 1006 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1007 NewAI->takeName(AI); 1008 AI->eraseFromParent(); 1009 ++NumConverted; 1010 Changed = true; 1011 continue; 1012 } 1013 1014 // Otherwise, couldn't process this alloca. 1015 } 1016 1017 return Changed; 1018} 1019 1020/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1021/// predicate, do SROA now. 1022void SROA::DoScalarReplacement(AllocaInst *AI, 1023 std::vector<AllocaInst*> &WorkList) { 1024 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1025 SmallVector<AllocaInst*, 32> ElementAllocas; 1026 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1027 ElementAllocas.reserve(ST->getNumContainedTypes()); 1028 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1029 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1030 AI->getAlignment(), 1031 AI->getName() + "." + Twine(i), AI); 1032 ElementAllocas.push_back(NA); 1033 WorkList.push_back(NA); // Add to worklist for recursive processing 1034 } 1035 } else { 1036 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1037 ElementAllocas.reserve(AT->getNumElements()); 1038 const Type *ElTy = AT->getElementType(); 1039 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1040 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1041 AI->getName() + "." + Twine(i), AI); 1042 ElementAllocas.push_back(NA); 1043 WorkList.push_back(NA); // Add to worklist for recursive processing 1044 } 1045 } 1046 1047 // Now that we have created the new alloca instructions, rewrite all the 1048 // uses of the old alloca. 1049 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1050 1051 // Now erase any instructions that were made dead while rewriting the alloca. 1052 DeleteDeadInstructions(); 1053 AI->eraseFromParent(); 1054 1055 ++NumReplaced; 1056} 1057 1058/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1059/// recursively including all their operands that become trivially dead. 1060void SROA::DeleteDeadInstructions() { 1061 while (!DeadInsts.empty()) { 1062 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1063 1064 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1065 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1066 // Zero out the operand and see if it becomes trivially dead. 1067 // (But, don't add allocas to the dead instruction list -- they are 1068 // already on the worklist and will be deleted separately.) 1069 *OI = 0; 1070 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1071 DeadInsts.push_back(U); 1072 } 1073 1074 I->eraseFromParent(); 1075 } 1076} 1077 1078/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1079/// performing scalar replacement of alloca AI. The results are flagged in 1080/// the Info parameter. Offset indicates the position within AI that is 1081/// referenced by this instruction. 1082void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1083 AllocaInfo &Info) { 1084 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1085 Instruction *User = cast<Instruction>(*UI); 1086 1087 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1088 isSafeForScalarRepl(BC, AI, Offset, Info); 1089 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1090 uint64_t GEPOffset = Offset; 1091 isSafeGEP(GEPI, AI, GEPOffset, Info); 1092 if (!Info.isUnsafe) 1093 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 1094 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1095 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1096 if (Length) 1097 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 1098 UI.getOperandNo() == 0, Info); 1099 else 1100 MarkUnsafe(Info); 1101 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1102 if (!LI->isVolatile()) { 1103 const Type *LIType = LI->getType(); 1104 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 1105 LIType, false, Info); 1106 } else 1107 MarkUnsafe(Info); 1108 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1109 // Store is ok if storing INTO the pointer, not storing the pointer 1110 if (!SI->isVolatile() && SI->getOperand(0) != I) { 1111 const Type *SIType = SI->getOperand(0)->getType(); 1112 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 1113 SIType, true, Info); 1114 } else 1115 MarkUnsafe(Info); 1116 } else { 1117 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 1118 MarkUnsafe(Info); 1119 } 1120 if (Info.isUnsafe) return; 1121 } 1122} 1123 1124/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1125/// replacement. It is safe when all the indices are constant, in-bounds 1126/// references, and when the resulting offset corresponds to an element within 1127/// the alloca type. The results are flagged in the Info parameter. Upon 1128/// return, Offset is adjusted as specified by the GEP indices. 1129void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 1130 uint64_t &Offset, AllocaInfo &Info) { 1131 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1132 if (GEPIt == E) 1133 return; 1134 1135 // Walk through the GEP type indices, checking the types that this indexes 1136 // into. 1137 for (; GEPIt != E; ++GEPIt) { 1138 // Ignore struct elements, no extra checking needed for these. 1139 if ((*GEPIt)->isStructTy()) 1140 continue; 1141 1142 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1143 if (!IdxVal) 1144 return MarkUnsafe(Info); 1145 } 1146 1147 // Compute the offset due to this GEP and check if the alloca has a 1148 // component element at that offset. 1149 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1150 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1151 &Indices[0], Indices.size()); 1152 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 1153 MarkUnsafe(Info); 1154} 1155 1156/// isHomogeneousAggregate - Check if type T is a struct or array containing 1157/// elements of the same type (which is always true for arrays). If so, 1158/// return true with NumElts and EltTy set to the number of elements and the 1159/// element type, respectively. 1160static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1161 const Type *&EltTy) { 1162 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1163 NumElts = AT->getNumElements(); 1164 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1165 return true; 1166 } 1167 if (const StructType *ST = dyn_cast<StructType>(T)) { 1168 NumElts = ST->getNumContainedTypes(); 1169 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1170 for (unsigned n = 1; n < NumElts; ++n) { 1171 if (ST->getContainedType(n) != EltTy) 1172 return false; 1173 } 1174 return true; 1175 } 1176 return false; 1177} 1178 1179/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1180/// "homogeneous" aggregates with the same element type and number of elements. 1181static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1182 if (T1 == T2) 1183 return true; 1184 1185 unsigned NumElts1, NumElts2; 1186 const Type *EltTy1, *EltTy2; 1187 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1188 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1189 NumElts1 == NumElts2 && 1190 EltTy1 == EltTy2) 1191 return true; 1192 1193 return false; 1194} 1195 1196/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1197/// alloca or has an offset and size that corresponds to a component element 1198/// within it. The offset checked here may have been formed from a GEP with a 1199/// pointer bitcasted to a different type. 1200void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 1201 const Type *MemOpType, bool isStore, 1202 AllocaInfo &Info) { 1203 // Check if this is a load/store of the entire alloca. 1204 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 1205 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1206 // loads/stores (which are essentially the same as the MemIntrinsics with 1207 // regard to copying padding between elements). But, if an alloca is 1208 // flagged as both a source and destination of such operations, we'll need 1209 // to check later for padding between elements. 1210 if (!MemOpType || MemOpType->isIntegerTy()) { 1211 if (isStore) 1212 Info.isMemCpyDst = true; 1213 else 1214 Info.isMemCpySrc = true; 1215 return; 1216 } 1217 // This is also safe for references using a type that is compatible with 1218 // the type of the alloca, so that loads/stores can be rewritten using 1219 // insertvalue/extractvalue. 1220 if (isCompatibleAggregate(MemOpType, AI->getAllocatedType())) 1221 return; 1222 } 1223 // Check if the offset/size correspond to a component within the alloca type. 1224 const Type *T = AI->getAllocatedType(); 1225 if (TypeHasComponent(T, Offset, MemSize)) 1226 return; 1227 1228 return MarkUnsafe(Info); 1229} 1230 1231/// TypeHasComponent - Return true if T has a component type with the 1232/// specified offset and size. If Size is zero, do not check the size. 1233bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1234 const Type *EltTy; 1235 uint64_t EltSize; 1236 if (const StructType *ST = dyn_cast<StructType>(T)) { 1237 const StructLayout *Layout = TD->getStructLayout(ST); 1238 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1239 EltTy = ST->getContainedType(EltIdx); 1240 EltSize = TD->getTypeAllocSize(EltTy); 1241 Offset -= Layout->getElementOffset(EltIdx); 1242 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1243 EltTy = AT->getElementType(); 1244 EltSize = TD->getTypeAllocSize(EltTy); 1245 if (Offset >= AT->getNumElements() * EltSize) 1246 return false; 1247 Offset %= EltSize; 1248 } else { 1249 return false; 1250 } 1251 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1252 return true; 1253 // Check if the component spans multiple elements. 1254 if (Offset + Size > EltSize) 1255 return false; 1256 return TypeHasComponent(EltTy, Offset, Size); 1257} 1258 1259/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1260/// the instruction I, which references it, to use the separate elements. 1261/// Offset indicates the position within AI that is referenced by this 1262/// instruction. 1263void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1264 SmallVector<AllocaInst*, 32> &NewElts) { 1265 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1266 Instruction *User = cast<Instruction>(*UI); 1267 1268 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1269 RewriteBitCast(BC, AI, Offset, NewElts); 1270 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1271 RewriteGEP(GEPI, AI, Offset, NewElts); 1272 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1273 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1274 uint64_t MemSize = Length->getZExtValue(); 1275 if (Offset == 0 && 1276 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1277 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1278 // Otherwise the intrinsic can only touch a single element and the 1279 // address operand will be updated, so nothing else needs to be done. 1280 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1281 const Type *LIType = LI->getType(); 1282 1283 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1284 // Replace: 1285 // %res = load { i32, i32 }* %alloc 1286 // with: 1287 // %load.0 = load i32* %alloc.0 1288 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1289 // %load.1 = load i32* %alloc.1 1290 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1291 // (Also works for arrays instead of structs) 1292 Value *Insert = UndefValue::get(LIType); 1293 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1294 Value *Load = new LoadInst(NewElts[i], "load", LI); 1295 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1296 } 1297 LI->replaceAllUsesWith(Insert); 1298 DeadInsts.push_back(LI); 1299 } else if (LIType->isIntegerTy() && 1300 TD->getTypeAllocSize(LIType) == 1301 TD->getTypeAllocSize(AI->getAllocatedType())) { 1302 // If this is a load of the entire alloca to an integer, rewrite it. 1303 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1304 } 1305 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1306 Value *Val = SI->getOperand(0); 1307 const Type *SIType = Val->getType(); 1308 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1309 // Replace: 1310 // store { i32, i32 } %val, { i32, i32 }* %alloc 1311 // with: 1312 // %val.0 = extractvalue { i32, i32 } %val, 0 1313 // store i32 %val.0, i32* %alloc.0 1314 // %val.1 = extractvalue { i32, i32 } %val, 1 1315 // store i32 %val.1, i32* %alloc.1 1316 // (Also works for arrays instead of structs) 1317 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1318 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1319 new StoreInst(Extract, NewElts[i], SI); 1320 } 1321 DeadInsts.push_back(SI); 1322 } else if (SIType->isIntegerTy() && 1323 TD->getTypeAllocSize(SIType) == 1324 TD->getTypeAllocSize(AI->getAllocatedType())) { 1325 // If this is a store of the entire alloca from an integer, rewrite it. 1326 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1327 } 1328 } 1329 } 1330} 1331 1332/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1333/// and recursively continue updating all of its uses. 1334void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1335 SmallVector<AllocaInst*, 32> &NewElts) { 1336 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1337 if (BC->getOperand(0) != AI) 1338 return; 1339 1340 // The bitcast references the original alloca. Replace its uses with 1341 // references to the first new element alloca. 1342 Instruction *Val = NewElts[0]; 1343 if (Val->getType() != BC->getDestTy()) { 1344 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1345 Val->takeName(BC); 1346 } 1347 BC->replaceAllUsesWith(Val); 1348 DeadInsts.push_back(BC); 1349} 1350 1351/// FindElementAndOffset - Return the index of the element containing Offset 1352/// within the specified type, which must be either a struct or an array. 1353/// Sets T to the type of the element and Offset to the offset within that 1354/// element. IdxTy is set to the type of the index result to be used in a 1355/// GEP instruction. 1356uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1357 const Type *&IdxTy) { 1358 uint64_t Idx = 0; 1359 if (const StructType *ST = dyn_cast<StructType>(T)) { 1360 const StructLayout *Layout = TD->getStructLayout(ST); 1361 Idx = Layout->getElementContainingOffset(Offset); 1362 T = ST->getContainedType(Idx); 1363 Offset -= Layout->getElementOffset(Idx); 1364 IdxTy = Type::getInt32Ty(T->getContext()); 1365 return Idx; 1366 } 1367 const ArrayType *AT = cast<ArrayType>(T); 1368 T = AT->getElementType(); 1369 uint64_t EltSize = TD->getTypeAllocSize(T); 1370 Idx = Offset / EltSize; 1371 Offset -= Idx * EltSize; 1372 IdxTy = Type::getInt64Ty(T->getContext()); 1373 return Idx; 1374} 1375 1376/// RewriteGEP - Check if this GEP instruction moves the pointer across 1377/// elements of the alloca that are being split apart, and if so, rewrite 1378/// the GEP to be relative to the new element. 1379void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1380 SmallVector<AllocaInst*, 32> &NewElts) { 1381 uint64_t OldOffset = Offset; 1382 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1383 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1384 &Indices[0], Indices.size()); 1385 1386 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1387 1388 const Type *T = AI->getAllocatedType(); 1389 const Type *IdxTy; 1390 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1391 if (GEPI->getOperand(0) == AI) 1392 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1393 1394 T = AI->getAllocatedType(); 1395 uint64_t EltOffset = Offset; 1396 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1397 1398 // If this GEP does not move the pointer across elements of the alloca 1399 // being split, then it does not needs to be rewritten. 1400 if (Idx == OldIdx) 1401 return; 1402 1403 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1404 SmallVector<Value*, 8> NewArgs; 1405 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1406 while (EltOffset != 0) { 1407 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1408 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1409 } 1410 Instruction *Val = NewElts[Idx]; 1411 if (NewArgs.size() > 1) { 1412 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1413 NewArgs.end(), "", GEPI); 1414 Val->takeName(GEPI); 1415 } 1416 if (Val->getType() != GEPI->getType()) 1417 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1418 GEPI->replaceAllUsesWith(Val); 1419 DeadInsts.push_back(GEPI); 1420} 1421 1422/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1423/// Rewrite it to copy or set the elements of the scalarized memory. 1424void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1425 AllocaInst *AI, 1426 SmallVector<AllocaInst*, 32> &NewElts) { 1427 // If this is a memcpy/memmove, construct the other pointer as the 1428 // appropriate type. The "Other" pointer is the pointer that goes to memory 1429 // that doesn't have anything to do with the alloca that we are promoting. For 1430 // memset, this Value* stays null. 1431 Value *OtherPtr = 0; 1432 unsigned MemAlignment = MI->getAlignment(); 1433 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1434 if (Inst == MTI->getRawDest()) 1435 OtherPtr = MTI->getRawSource(); 1436 else { 1437 assert(Inst == MTI->getRawSource()); 1438 OtherPtr = MTI->getRawDest(); 1439 } 1440 } 1441 1442 // If there is an other pointer, we want to convert it to the same pointer 1443 // type as AI has, so we can GEP through it safely. 1444 if (OtherPtr) { 1445 unsigned AddrSpace = 1446 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 1447 1448 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1449 // optimization, but it's also required to detect the corner case where 1450 // both pointer operands are referencing the same memory, and where 1451 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1452 // function is only called for mem intrinsics that access the whole 1453 // aggregate, so non-zero GEPs are not an issue here.) 1454 OtherPtr = OtherPtr->stripPointerCasts(); 1455 1456 // Copying the alloca to itself is a no-op: just delete it. 1457 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1458 // This code will run twice for a no-op memcpy -- once for each operand. 1459 // Put only one reference to MI on the DeadInsts list. 1460 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1461 E = DeadInsts.end(); I != E; ++I) 1462 if (*I == MI) return; 1463 DeadInsts.push_back(MI); 1464 return; 1465 } 1466 1467 // If the pointer is not the right type, insert a bitcast to the right 1468 // type. 1469 const Type *NewTy = 1470 PointerType::get(AI->getType()->getElementType(), AddrSpace); 1471 1472 if (OtherPtr->getType() != NewTy) 1473 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 1474 } 1475 1476 // Process each element of the aggregate. 1477 bool SROADest = MI->getRawDest() == Inst; 1478 1479 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1480 1481 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1482 // If this is a memcpy/memmove, emit a GEP of the other element address. 1483 Value *OtherElt = 0; 1484 unsigned OtherEltAlign = MemAlignment; 1485 1486 if (OtherPtr) { 1487 Value *Idx[2] = { Zero, 1488 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1489 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1490 OtherPtr->getName()+"."+Twine(i), 1491 MI); 1492 uint64_t EltOffset; 1493 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1494 const Type *OtherTy = OtherPtrTy->getElementType(); 1495 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1496 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1497 } else { 1498 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 1499 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1500 } 1501 1502 // The alignment of the other pointer is the guaranteed alignment of the 1503 // element, which is affected by both the known alignment of the whole 1504 // mem intrinsic and the alignment of the element. If the alignment of 1505 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 1506 // known alignment is just 4 bytes. 1507 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 1508 } 1509 1510 Value *EltPtr = NewElts[i]; 1511 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 1512 1513 // If we got down to a scalar, insert a load or store as appropriate. 1514 if (EltTy->isSingleValueType()) { 1515 if (isa<MemTransferInst>(MI)) { 1516 if (SROADest) { 1517 // From Other to Alloca. 1518 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 1519 new StoreInst(Elt, EltPtr, MI); 1520 } else { 1521 // From Alloca to Other. 1522 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 1523 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 1524 } 1525 continue; 1526 } 1527 assert(isa<MemSetInst>(MI)); 1528 1529 // If the stored element is zero (common case), just store a null 1530 // constant. 1531 Constant *StoreVal; 1532 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 1533 if (CI->isZero()) { 1534 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 1535 } else { 1536 // If EltTy is a vector type, get the element type. 1537 const Type *ValTy = EltTy->getScalarType(); 1538 1539 // Construct an integer with the right value. 1540 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 1541 APInt OneVal(EltSize, CI->getZExtValue()); 1542 APInt TotalVal(OneVal); 1543 // Set each byte. 1544 for (unsigned i = 0; 8*i < EltSize; ++i) { 1545 TotalVal = TotalVal.shl(8); 1546 TotalVal |= OneVal; 1547 } 1548 1549 // Convert the integer value to the appropriate type. 1550 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 1551 if (ValTy->isPointerTy()) 1552 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 1553 else if (ValTy->isFloatingPointTy()) 1554 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 1555 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 1556 1557 // If the requested value was a vector constant, create it. 1558 if (EltTy != ValTy) { 1559 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 1560 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 1561 StoreVal = ConstantVector::get(&Elts[0], NumElts); 1562 } 1563 } 1564 new StoreInst(StoreVal, EltPtr, MI); 1565 continue; 1566 } 1567 // Otherwise, if we're storing a byte variable, use a memset call for 1568 // this element. 1569 } 1570 1571 unsigned EltSize = TD->getTypeAllocSize(EltTy); 1572 1573 IRBuilder<> Builder(MI); 1574 1575 // Finally, insert the meminst for this element. 1576 if (isa<MemSetInst>(MI)) { 1577 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 1578 MI->isVolatile()); 1579 } else { 1580 assert(isa<MemTransferInst>(MI)); 1581 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 1582 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 1583 1584 if (isa<MemCpyInst>(MI)) 1585 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 1586 else 1587 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 1588 } 1589 } 1590 DeadInsts.push_back(MI); 1591} 1592 1593/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 1594/// overwrites the entire allocation. Extract out the pieces of the stored 1595/// integer and store them individually. 1596void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 1597 SmallVector<AllocaInst*, 32> &NewElts){ 1598 // Extract each element out of the integer according to its structure offset 1599 // and store the element value to the individual alloca. 1600 Value *SrcVal = SI->getOperand(0); 1601 const Type *AllocaEltTy = AI->getAllocatedType(); 1602 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1603 1604 IRBuilder<> Builder(SI); 1605 1606 // Handle tail padding by extending the operand 1607 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 1608 SrcVal = Builder.CreateZExt(SrcVal, 1609 IntegerType::get(SI->getContext(), AllocaSizeBits)); 1610 1611 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 1612 << '\n'); 1613 1614 // There are two forms here: AI could be an array or struct. Both cases 1615 // have different ways to compute the element offset. 1616 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1617 const StructLayout *Layout = TD->getStructLayout(EltSTy); 1618 1619 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1620 // Get the number of bits to shift SrcVal to get the value. 1621 const Type *FieldTy = EltSTy->getElementType(i); 1622 uint64_t Shift = Layout->getElementOffsetInBits(i); 1623 1624 if (TD->isBigEndian()) 1625 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 1626 1627 Value *EltVal = SrcVal; 1628 if (Shift) { 1629 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1630 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 1631 } 1632 1633 // Truncate down to an integer of the right size. 1634 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1635 1636 // Ignore zero sized fields like {}, they obviously contain no data. 1637 if (FieldSizeBits == 0) continue; 1638 1639 if (FieldSizeBits != AllocaSizeBits) 1640 EltVal = Builder.CreateTrunc(EltVal, 1641 IntegerType::get(SI->getContext(), FieldSizeBits)); 1642 Value *DestField = NewElts[i]; 1643 if (EltVal->getType() == FieldTy) { 1644 // Storing to an integer field of this size, just do it. 1645 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 1646 // Bitcast to the right element type (for fp/vector values). 1647 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 1648 } else { 1649 // Otherwise, bitcast the dest pointer (for aggregates). 1650 DestField = Builder.CreateBitCast(DestField, 1651 PointerType::getUnqual(EltVal->getType())); 1652 } 1653 new StoreInst(EltVal, DestField, SI); 1654 } 1655 1656 } else { 1657 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1658 const Type *ArrayEltTy = ATy->getElementType(); 1659 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1660 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1661 1662 uint64_t Shift; 1663 1664 if (TD->isBigEndian()) 1665 Shift = AllocaSizeBits-ElementOffset; 1666 else 1667 Shift = 0; 1668 1669 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1670 // Ignore zero sized fields like {}, they obviously contain no data. 1671 if (ElementSizeBits == 0) continue; 1672 1673 Value *EltVal = SrcVal; 1674 if (Shift) { 1675 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1676 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 1677 } 1678 1679 // Truncate down to an integer of the right size. 1680 if (ElementSizeBits != AllocaSizeBits) 1681 EltVal = Builder.CreateTrunc(EltVal, 1682 IntegerType::get(SI->getContext(), 1683 ElementSizeBits)); 1684 Value *DestField = NewElts[i]; 1685 if (EltVal->getType() == ArrayEltTy) { 1686 // Storing to an integer field of this size, just do it. 1687 } else if (ArrayEltTy->isFloatingPointTy() || 1688 ArrayEltTy->isVectorTy()) { 1689 // Bitcast to the right element type (for fp/vector values). 1690 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 1691 } else { 1692 // Otherwise, bitcast the dest pointer (for aggregates). 1693 DestField = Builder.CreateBitCast(DestField, 1694 PointerType::getUnqual(EltVal->getType())); 1695 } 1696 new StoreInst(EltVal, DestField, SI); 1697 1698 if (TD->isBigEndian()) 1699 Shift -= ElementOffset; 1700 else 1701 Shift += ElementOffset; 1702 } 1703 } 1704 1705 DeadInsts.push_back(SI); 1706} 1707 1708/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1709/// an integer. Load the individual pieces to form the aggregate value. 1710void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1711 SmallVector<AllocaInst*, 32> &NewElts) { 1712 // Extract each element out of the NewElts according to its structure offset 1713 // and form the result value. 1714 const Type *AllocaEltTy = AI->getAllocatedType(); 1715 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1716 1717 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1718 << '\n'); 1719 1720 // There are two forms here: AI could be an array or struct. Both cases 1721 // have different ways to compute the element offset. 1722 const StructLayout *Layout = 0; 1723 uint64_t ArrayEltBitOffset = 0; 1724 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1725 Layout = TD->getStructLayout(EltSTy); 1726 } else { 1727 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1728 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1729 } 1730 1731 Value *ResultVal = 1732 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1733 1734 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1735 // Load the value from the alloca. If the NewElt is an aggregate, cast 1736 // the pointer to an integer of the same size before doing the load. 1737 Value *SrcField = NewElts[i]; 1738 const Type *FieldTy = 1739 cast<PointerType>(SrcField->getType())->getElementType(); 1740 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1741 1742 // Ignore zero sized fields like {}, they obviously contain no data. 1743 if (FieldSizeBits == 0) continue; 1744 1745 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1746 FieldSizeBits); 1747 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1748 !FieldTy->isVectorTy()) 1749 SrcField = new BitCastInst(SrcField, 1750 PointerType::getUnqual(FieldIntTy), 1751 "", LI); 1752 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1753 1754 // If SrcField is a fp or vector of the right size but that isn't an 1755 // integer type, bitcast to an integer so we can shift it. 1756 if (SrcField->getType() != FieldIntTy) 1757 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1758 1759 // Zero extend the field to be the same size as the final alloca so that 1760 // we can shift and insert it. 1761 if (SrcField->getType() != ResultVal->getType()) 1762 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1763 1764 // Determine the number of bits to shift SrcField. 1765 uint64_t Shift; 1766 if (Layout) // Struct case. 1767 Shift = Layout->getElementOffsetInBits(i); 1768 else // Array case. 1769 Shift = i*ArrayEltBitOffset; 1770 1771 if (TD->isBigEndian()) 1772 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1773 1774 if (Shift) { 1775 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1776 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1777 } 1778 1779 // Don't create an 'or x, 0' on the first iteration. 1780 if (!isa<Constant>(ResultVal) || 1781 !cast<Constant>(ResultVal)->isNullValue()) 1782 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1783 else 1784 ResultVal = SrcField; 1785 } 1786 1787 // Handle tail padding by truncating the result 1788 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1789 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1790 1791 LI->replaceAllUsesWith(ResultVal); 1792 DeadInsts.push_back(LI); 1793} 1794 1795/// HasPadding - Return true if the specified type has any structure or 1796/// alignment padding in between the elements that would be split apart 1797/// by SROA; return false otherwise. 1798static bool HasPadding(const Type *Ty, const TargetData &TD) { 1799 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1800 Ty = ATy->getElementType(); 1801 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1802 } 1803 1804 // SROA currently handles only Arrays and Structs. 1805 const StructType *STy = cast<StructType>(Ty); 1806 const StructLayout *SL = TD.getStructLayout(STy); 1807 unsigned PrevFieldBitOffset = 0; 1808 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1809 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1810 1811 // Check to see if there is any padding between this element and the 1812 // previous one. 1813 if (i) { 1814 unsigned PrevFieldEnd = 1815 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1816 if (PrevFieldEnd < FieldBitOffset) 1817 return true; 1818 } 1819 PrevFieldBitOffset = FieldBitOffset; 1820 } 1821 // Check for tail padding. 1822 if (unsigned EltCount = STy->getNumElements()) { 1823 unsigned PrevFieldEnd = PrevFieldBitOffset + 1824 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1825 if (PrevFieldEnd < SL->getSizeInBits()) 1826 return true; 1827 } 1828 return false; 1829} 1830 1831/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1832/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1833/// or 1 if safe after canonicalization has been performed. 1834bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1835 // Loop over the use list of the alloca. We can only transform it if all of 1836 // the users are safe to transform. 1837 AllocaInfo Info; 1838 1839 isSafeForScalarRepl(AI, AI, 0, Info); 1840 if (Info.isUnsafe) { 1841 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1842 return false; 1843 } 1844 1845 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1846 // source and destination, we have to be careful. In particular, the memcpy 1847 // could be moving around elements that live in structure padding of the LLVM 1848 // types, but may actually be used. In these cases, we refuse to promote the 1849 // struct. 1850 if (Info.isMemCpySrc && Info.isMemCpyDst && 1851 HasPadding(AI->getAllocatedType(), *TD)) 1852 return false; 1853 1854 return true; 1855} 1856 1857 1858 1859/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1860/// some part of a constant global variable. This intentionally only accepts 1861/// constant expressions because we don't can't rewrite arbitrary instructions. 1862static bool PointsToConstantGlobal(Value *V) { 1863 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1864 return GV->isConstant(); 1865 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1866 if (CE->getOpcode() == Instruction::BitCast || 1867 CE->getOpcode() == Instruction::GetElementPtr) 1868 return PointsToConstantGlobal(CE->getOperand(0)); 1869 return false; 1870} 1871 1872/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1873/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1874/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1875/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1876/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1877/// the alloca, and if the source pointer is a pointer to a constant global, we 1878/// can optimize this. 1879static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 1880 bool isOffset) { 1881 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1882 User *U = cast<Instruction>(*UI); 1883 1884 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1885 // Ignore non-volatile loads, they are always ok. 1886 if (LI->isVolatile()) return false; 1887 continue; 1888 } 1889 1890 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1891 // If uses of the bitcast are ok, we are ok. 1892 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1893 return false; 1894 continue; 1895 } 1896 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 1897 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1898 // doesn't, it does. 1899 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1900 isOffset || !GEP->hasAllZeroIndices())) 1901 return false; 1902 continue; 1903 } 1904 1905 if (CallSite CS = U) { 1906 // If this is a readonly/readnone call site, then we know it is just a 1907 // load and we can ignore it. 1908 if (CS.onlyReadsMemory()) 1909 continue; 1910 1911 // If this is the function being called then we treat it like a load and 1912 // ignore it. 1913 if (CS.isCallee(UI)) 1914 continue; 1915 1916 // If this is being passed as a byval argument, the caller is making a 1917 // copy, so it is only a read of the alloca. 1918 unsigned ArgNo = CS.getArgumentNo(UI); 1919 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 1920 continue; 1921 } 1922 1923 // If this is isn't our memcpy/memmove, reject it as something we can't 1924 // handle. 1925 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 1926 if (MI == 0) 1927 return false; 1928 1929 // If the transfer is using the alloca as a source of the transfer, then 1930 // ignore it since it is a load (unless the transfer is volatile). 1931 if (UI.getOperandNo() == 1) { 1932 if (MI->isVolatile()) return false; 1933 continue; 1934 } 1935 1936 // If we already have seen a copy, reject the second one. 1937 if (TheCopy) return false; 1938 1939 // If the pointer has been offset from the start of the alloca, we can't 1940 // safely handle this. 1941 if (isOffset) return false; 1942 1943 // If the memintrinsic isn't using the alloca as the dest, reject it. 1944 if (UI.getOperandNo() != 0) return false; 1945 1946 // If the source of the memcpy/move is not a constant global, reject it. 1947 if (!PointsToConstantGlobal(MI->getSource())) 1948 return false; 1949 1950 // Otherwise, the transform is safe. Remember the copy instruction. 1951 TheCopy = MI; 1952 } 1953 return true; 1954} 1955 1956/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1957/// modified by a copy from a constant global. If we can prove this, we can 1958/// replace any uses of the alloca with uses of the global directly. 1959MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1960 MemTransferInst *TheCopy = 0; 1961 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1962 return TheCopy; 1963 return 0; 1964} 1965