ScalarReplAggregates.cpp revision b352d6eb49927a7c707cbd9046cfc525b0c3f2d7
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} 842 843/// PromoteAlloca - Promote an alloca to registers, using SSAUpdater. 844static void PromoteAlloca(AllocaInst *AI, SSAUpdater &SSA) { 845 SSA.Initialize(AI->getType()->getElementType(), AI->getName()); 846 847 // First step: bucket up uses of the alloca by the block they occur in. 848 // This is important because we have to handle multiple defs/uses in a block 849 // ourselves: SSAUpdater is purely for cross-block references. 850 // FIXME: Want a TinyVector<Instruction*> since there is often 0/1 element. 851 DenseMap<BasicBlock*, std::vector<Instruction*> > UsesByBlock; 852 853 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 854 UI != E; ++UI) { 855 Instruction *User = cast<Instruction>(*UI); 856 UsesByBlock[User->getParent()].push_back(User); 857 } 858 859 // Okay, now we can iterate over all the blocks in the function with uses, 860 // processing them. Keep track of which loads are loading a live-in value. 861 // Walk the uses in the use-list order to be determinstic. 862 SmallVector<LoadInst*, 32> LiveInLoads; 863 DenseMap<Value*, Value*> ReplacedLoads; 864 865 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 866 UI != E; ++UI) { 867 Instruction *User = cast<Instruction>(*UI); 868 BasicBlock *BB = User->getParent(); 869 std::vector<Instruction*> &BlockUses = UsesByBlock[BB]; 870 871 // If this block has already been processed, ignore this repeat use. 872 if (BlockUses.empty()) continue; 873 874 // Okay, this is the first use in the block. If this block just has a 875 // single user in it, we can rewrite it trivially. 876 if (BlockUses.size() == 1) { 877 // If it is a store, it is a trivial def of the value in the block. 878 if (StoreInst *SI = dyn_cast<StoreInst>(User)) 879 SSA.AddAvailableValue(BB, SI->getOperand(0)); 880 else 881 // Otherwise it is a load, queue it to rewrite as a live-in load. 882 LiveInLoads.push_back(cast<LoadInst>(User)); 883 BlockUses.clear(); 884 continue; 885 } 886 887 // Otherwise, check to see if this block is all loads. 888 bool HasStore = false; 889 for (unsigned i = 0, e = BlockUses.size(); i != e; ++i) { 890 if (isa<StoreInst>(BlockUses[i])) { 891 HasStore = true; 892 break; 893 } 894 } 895 896 // If so, we can queue them all as live in loads. We don't have an 897 // efficient way to tell which on is first in the block and don't want to 898 // scan large blocks, so just add all loads as live ins. 899 if (!HasStore) { 900 for (unsigned i = 0, e = BlockUses.size(); i != e; ++i) 901 LiveInLoads.push_back(cast<LoadInst>(BlockUses[i])); 902 BlockUses.clear(); 903 continue; 904 } 905 906 // Otherwise, we have mixed loads and stores (or just a bunch of stores). 907 // Since SSAUpdater is purely for cross-block values, we need to determine 908 // the order of these instructions in the block. If the first use in the 909 // block is a load, then it uses the live in value. The last store defines 910 // the live out value. We handle this by doing a linear scan of the block. 911 Value *StoredValue = 0; 912 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) { 913 if (LoadInst *L = dyn_cast<LoadInst>(II)) { 914 // If this is a load from an unrelated pointer, ignore it. 915 if (L->getOperand(0) != AI) continue; 916 917 // If we haven't seen a store yet, this is a live in use, otherwise 918 // use the stored value. 919 if (StoredValue) { 920 L->replaceAllUsesWith(StoredValue); 921 ReplacedLoads[L] = StoredValue; 922 } else { 923 LiveInLoads.push_back(L); 924 } 925 continue; 926 } 927 928 if (StoreInst *S = dyn_cast<StoreInst>(II)) { 929 // If this is a store to an unrelated pointer, ignore it. 930 if (S->getPointerOperand() != AI) continue; 931 932 // Remember that this is the active value in the block. 933 StoredValue = S->getOperand(0); 934 } 935 } 936 937 // The last stored value that happened is the live-out for the block. 938 assert(StoredValue && "Already checked that there is a store in block"); 939 SSA.AddAvailableValue(BB, StoredValue); 940 BlockUses.clear(); 941 } 942 943 // Okay, now we rewrite all loads that use live-in values in the loop, 944 // inserting PHI nodes as necessary. 945 for (unsigned i = 0, e = LiveInLoads.size(); i != e; ++i) { 946 LoadInst *ALoad = LiveInLoads[i]; 947 Value *NewVal = SSA.GetValueInMiddleOfBlock(ALoad->getParent()); 948 ALoad->replaceAllUsesWith(NewVal); 949 ReplacedLoads[ALoad] = NewVal; 950 } 951 952 // Now that everything is rewritten, delete the old instructions from the 953 // function. They should all be dead now. 954 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) { 955 Instruction *User = cast<Instruction>(*UI++); 956 957 // If this is a load that still has uses, then the load must have been added 958 // as a live value in the SSAUpdate data structure for a block (e.g. because 959 // the loaded value was stored later). In this case, we need to recursively 960 // propagate the updates until we get to the real value. 961 if (!User->use_empty()) { 962 Value *NewVal = ReplacedLoads[User]; 963 assert(NewVal && "not a replaced load?"); 964 965 // Propagate down to the ultimate replacee. The intermediately loads 966 // could theoretically already have been deleted, so we don't want to 967 // dereference the Value*'s. 968 DenseMap<Value*, Value*>::iterator RLI = ReplacedLoads.find(NewVal); 969 while (RLI != ReplacedLoads.end()) { 970 NewVal = RLI->second; 971 RLI = ReplacedLoads.find(NewVal); 972 } 973 974 User->replaceAllUsesWith(NewVal); 975 } 976 977 User->eraseFromParent(); 978 } 979} 980 981 982bool SROA::performPromotion(Function &F) { 983 std::vector<AllocaInst*> Allocas; 984 DominatorTree *DT = 0; 985 DominanceFrontier *DF = 0; 986 if (HasDomFrontiers) { 987 DT = &getAnalysis<DominatorTree>(); 988 DF = &getAnalysis<DominanceFrontier>(); 989 } 990 991 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 992 993 bool Changed = false; 994 995 while (1) { 996 Allocas.clear(); 997 998 // Find allocas that are safe to promote, by looking at all instructions in 999 // the entry node 1000 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1001 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1002 if (isAllocaPromotable(AI)) 1003 Allocas.push_back(AI); 1004 1005 if (Allocas.empty()) break; 1006 1007 if (HasDomFrontiers) 1008 PromoteMemToReg(Allocas, *DT, *DF); 1009 else { 1010 SSAUpdater SSA; 1011 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1012 PromoteAlloca(Allocas[i], SSA); 1013 Allocas[i]->eraseFromParent(); 1014 } 1015 } 1016 NumPromoted += Allocas.size(); 1017 Changed = true; 1018 } 1019 1020 return Changed; 1021} 1022 1023 1024/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1025/// SROA. It must be a struct or array type with a small number of elements. 1026static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1027 const Type *T = AI->getAllocatedType(); 1028 // Do not promote any struct into more than 32 separate vars. 1029 if (const StructType *ST = dyn_cast<StructType>(T)) 1030 return ST->getNumElements() <= 32; 1031 // Arrays are much less likely to be safe for SROA; only consider 1032 // them if they are very small. 1033 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 1034 return AT->getNumElements() <= 8; 1035 return false; 1036} 1037 1038 1039// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1040// which runs on all of the malloc/alloca instructions in the function, removing 1041// them if they are only used by getelementptr instructions. 1042// 1043bool SROA::performScalarRepl(Function &F) { 1044 std::vector<AllocaInst*> WorkList; 1045 1046 // Scan the entry basic block, adding allocas to the worklist. 1047 BasicBlock &BB = F.getEntryBlock(); 1048 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1049 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1050 WorkList.push_back(A); 1051 1052 // Process the worklist 1053 bool Changed = false; 1054 while (!WorkList.empty()) { 1055 AllocaInst *AI = WorkList.back(); 1056 WorkList.pop_back(); 1057 1058 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1059 // with unused elements. 1060 if (AI->use_empty()) { 1061 AI->eraseFromParent(); 1062 Changed = true; 1063 continue; 1064 } 1065 1066 // If this alloca is impossible for us to promote, reject it early. 1067 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1068 continue; 1069 1070 // Check to see if this allocation is only modified by a memcpy/memmove from 1071 // a constant global. If this is the case, we can change all users to use 1072 // the constant global instead. This is commonly produced by the CFE by 1073 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1074 // is only subsequently read. 1075 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 1076 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1077 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 1078 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 1079 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1080 TheCopy->eraseFromParent(); // Don't mutate the global. 1081 AI->eraseFromParent(); 1082 ++NumGlobals; 1083 Changed = true; 1084 continue; 1085 } 1086 1087 // Check to see if we can perform the core SROA transformation. We cannot 1088 // transform the allocation instruction if it is an array allocation 1089 // (allocations OF arrays are ok though), and an allocation of a scalar 1090 // value cannot be decomposed at all. 1091 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1092 1093 // Do not promote [0 x %struct]. 1094 if (AllocaSize == 0) continue; 1095 1096 // Do not promote any struct whose size is too big. 1097 if (AllocaSize > SRThreshold) continue; 1098 1099 // If the alloca looks like a good candidate for scalar replacement, and if 1100 // all its users can be transformed, then split up the aggregate into its 1101 // separate elements. 1102 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1103 DoScalarReplacement(AI, WorkList); 1104 Changed = true; 1105 continue; 1106 } 1107 1108 // If we can turn this aggregate value (potentially with casts) into a 1109 // simple scalar value that can be mem2reg'd into a register value. 1110 // IsNotTrivial tracks whether this is something that mem2reg could have 1111 // promoted itself. If so, we don't want to transform it needlessly. Note 1112 // that we can't just check based on the type: the alloca may be of an i32 1113 // but that has pointer arithmetic to set byte 3 of it or something. 1114 if (AllocaInst *NewAI = 1115 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1116 NewAI->takeName(AI); 1117 AI->eraseFromParent(); 1118 ++NumConverted; 1119 Changed = true; 1120 continue; 1121 } 1122 1123 // Otherwise, couldn't process this alloca. 1124 } 1125 1126 return Changed; 1127} 1128 1129/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1130/// predicate, do SROA now. 1131void SROA::DoScalarReplacement(AllocaInst *AI, 1132 std::vector<AllocaInst*> &WorkList) { 1133 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1134 SmallVector<AllocaInst*, 32> ElementAllocas; 1135 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1136 ElementAllocas.reserve(ST->getNumContainedTypes()); 1137 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1138 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1139 AI->getAlignment(), 1140 AI->getName() + "." + Twine(i), AI); 1141 ElementAllocas.push_back(NA); 1142 WorkList.push_back(NA); // Add to worklist for recursive processing 1143 } 1144 } else { 1145 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1146 ElementAllocas.reserve(AT->getNumElements()); 1147 const Type *ElTy = AT->getElementType(); 1148 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1149 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1150 AI->getName() + "." + Twine(i), AI); 1151 ElementAllocas.push_back(NA); 1152 WorkList.push_back(NA); // Add to worklist for recursive processing 1153 } 1154 } 1155 1156 // Now that we have created the new alloca instructions, rewrite all the 1157 // uses of the old alloca. 1158 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1159 1160 // Now erase any instructions that were made dead while rewriting the alloca. 1161 DeleteDeadInstructions(); 1162 AI->eraseFromParent(); 1163 1164 ++NumReplaced; 1165} 1166 1167/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1168/// recursively including all their operands that become trivially dead. 1169void SROA::DeleteDeadInstructions() { 1170 while (!DeadInsts.empty()) { 1171 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1172 1173 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1174 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1175 // Zero out the operand and see if it becomes trivially dead. 1176 // (But, don't add allocas to the dead instruction list -- they are 1177 // already on the worklist and will be deleted separately.) 1178 *OI = 0; 1179 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1180 DeadInsts.push_back(U); 1181 } 1182 1183 I->eraseFromParent(); 1184 } 1185} 1186 1187/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1188/// performing scalar replacement of alloca AI. The results are flagged in 1189/// the Info parameter. Offset indicates the position within AI that is 1190/// referenced by this instruction. 1191void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1192 AllocaInfo &Info) { 1193 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1194 Instruction *User = cast<Instruction>(*UI); 1195 1196 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1197 isSafeForScalarRepl(BC, AI, Offset, Info); 1198 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1199 uint64_t GEPOffset = Offset; 1200 isSafeGEP(GEPI, AI, GEPOffset, Info); 1201 if (!Info.isUnsafe) 1202 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 1203 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1204 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1205 if (Length) 1206 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 1207 UI.getOperandNo() == 0, Info); 1208 else 1209 MarkUnsafe(Info); 1210 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1211 if (!LI->isVolatile()) { 1212 const Type *LIType = LI->getType(); 1213 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 1214 LIType, false, Info); 1215 } else 1216 MarkUnsafe(Info); 1217 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1218 // Store is ok if storing INTO the pointer, not storing the pointer 1219 if (!SI->isVolatile() && SI->getOperand(0) != I) { 1220 const Type *SIType = SI->getOperand(0)->getType(); 1221 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 1222 SIType, true, Info); 1223 } else 1224 MarkUnsafe(Info); 1225 } else { 1226 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 1227 MarkUnsafe(Info); 1228 } 1229 if (Info.isUnsafe) return; 1230 } 1231} 1232 1233/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1234/// replacement. It is safe when all the indices are constant, in-bounds 1235/// references, and when the resulting offset corresponds to an element within 1236/// the alloca type. The results are flagged in the Info parameter. Upon 1237/// return, Offset is adjusted as specified by the GEP indices. 1238void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 1239 uint64_t &Offset, AllocaInfo &Info) { 1240 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1241 if (GEPIt == E) 1242 return; 1243 1244 // Walk through the GEP type indices, checking the types that this indexes 1245 // into. 1246 for (; GEPIt != E; ++GEPIt) { 1247 // Ignore struct elements, no extra checking needed for these. 1248 if ((*GEPIt)->isStructTy()) 1249 continue; 1250 1251 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1252 if (!IdxVal) 1253 return MarkUnsafe(Info); 1254 } 1255 1256 // Compute the offset due to this GEP and check if the alloca has a 1257 // component element at that offset. 1258 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1259 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1260 &Indices[0], Indices.size()); 1261 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 1262 MarkUnsafe(Info); 1263} 1264 1265/// isHomogeneousAggregate - Check if type T is a struct or array containing 1266/// elements of the same type (which is always true for arrays). If so, 1267/// return true with NumElts and EltTy set to the number of elements and the 1268/// element type, respectively. 1269static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1270 const Type *&EltTy) { 1271 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1272 NumElts = AT->getNumElements(); 1273 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1274 return true; 1275 } 1276 if (const StructType *ST = dyn_cast<StructType>(T)) { 1277 NumElts = ST->getNumContainedTypes(); 1278 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1279 for (unsigned n = 1; n < NumElts; ++n) { 1280 if (ST->getContainedType(n) != EltTy) 1281 return false; 1282 } 1283 return true; 1284 } 1285 return false; 1286} 1287 1288/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1289/// "homogeneous" aggregates with the same element type and number of elements. 1290static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1291 if (T1 == T2) 1292 return true; 1293 1294 unsigned NumElts1, NumElts2; 1295 const Type *EltTy1, *EltTy2; 1296 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1297 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1298 NumElts1 == NumElts2 && 1299 EltTy1 == EltTy2) 1300 return true; 1301 1302 return false; 1303} 1304 1305/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1306/// alloca or has an offset and size that corresponds to a component element 1307/// within it. The offset checked here may have been formed from a GEP with a 1308/// pointer bitcasted to a different type. 1309void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 1310 const Type *MemOpType, bool isStore, 1311 AllocaInfo &Info) { 1312 // Check if this is a load/store of the entire alloca. 1313 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 1314 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1315 // loads/stores (which are essentially the same as the MemIntrinsics with 1316 // regard to copying padding between elements). But, if an alloca is 1317 // flagged as both a source and destination of such operations, we'll need 1318 // to check later for padding between elements. 1319 if (!MemOpType || MemOpType->isIntegerTy()) { 1320 if (isStore) 1321 Info.isMemCpyDst = true; 1322 else 1323 Info.isMemCpySrc = true; 1324 return; 1325 } 1326 // This is also safe for references using a type that is compatible with 1327 // the type of the alloca, so that loads/stores can be rewritten using 1328 // insertvalue/extractvalue. 1329 if (isCompatibleAggregate(MemOpType, AI->getAllocatedType())) 1330 return; 1331 } 1332 // Check if the offset/size correspond to a component within the alloca type. 1333 const Type *T = AI->getAllocatedType(); 1334 if (TypeHasComponent(T, Offset, MemSize)) 1335 return; 1336 1337 return MarkUnsafe(Info); 1338} 1339 1340/// TypeHasComponent - Return true if T has a component type with the 1341/// specified offset and size. If Size is zero, do not check the size. 1342bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1343 const Type *EltTy; 1344 uint64_t EltSize; 1345 if (const StructType *ST = dyn_cast<StructType>(T)) { 1346 const StructLayout *Layout = TD->getStructLayout(ST); 1347 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1348 EltTy = ST->getContainedType(EltIdx); 1349 EltSize = TD->getTypeAllocSize(EltTy); 1350 Offset -= Layout->getElementOffset(EltIdx); 1351 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1352 EltTy = AT->getElementType(); 1353 EltSize = TD->getTypeAllocSize(EltTy); 1354 if (Offset >= AT->getNumElements() * EltSize) 1355 return false; 1356 Offset %= EltSize; 1357 } else { 1358 return false; 1359 } 1360 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1361 return true; 1362 // Check if the component spans multiple elements. 1363 if (Offset + Size > EltSize) 1364 return false; 1365 return TypeHasComponent(EltTy, Offset, Size); 1366} 1367 1368/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1369/// the instruction I, which references it, to use the separate elements. 1370/// Offset indicates the position within AI that is referenced by this 1371/// instruction. 1372void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1373 SmallVector<AllocaInst*, 32> &NewElts) { 1374 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1375 Instruction *User = cast<Instruction>(*UI); 1376 1377 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1378 RewriteBitCast(BC, AI, Offset, NewElts); 1379 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1380 RewriteGEP(GEPI, AI, Offset, NewElts); 1381 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1382 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1383 uint64_t MemSize = Length->getZExtValue(); 1384 if (Offset == 0 && 1385 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1386 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1387 // Otherwise the intrinsic can only touch a single element and the 1388 // address operand will be updated, so nothing else needs to be done. 1389 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1390 const Type *LIType = LI->getType(); 1391 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1392 // Replace: 1393 // %res = load { i32, i32 }* %alloc 1394 // with: 1395 // %load.0 = load i32* %alloc.0 1396 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1397 // %load.1 = load i32* %alloc.1 1398 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1399 // (Also works for arrays instead of structs) 1400 Value *Insert = UndefValue::get(LIType); 1401 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1402 Value *Load = new LoadInst(NewElts[i], "load", LI); 1403 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1404 } 1405 LI->replaceAllUsesWith(Insert); 1406 DeadInsts.push_back(LI); 1407 } else if (LIType->isIntegerTy() && 1408 TD->getTypeAllocSize(LIType) == 1409 TD->getTypeAllocSize(AI->getAllocatedType())) { 1410 // If this is a load of the entire alloca to an integer, rewrite it. 1411 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1412 } 1413 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1414 Value *Val = SI->getOperand(0); 1415 const Type *SIType = Val->getType(); 1416 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1417 // Replace: 1418 // store { i32, i32 } %val, { i32, i32 }* %alloc 1419 // with: 1420 // %val.0 = extractvalue { i32, i32 } %val, 0 1421 // store i32 %val.0, i32* %alloc.0 1422 // %val.1 = extractvalue { i32, i32 } %val, 1 1423 // store i32 %val.1, i32* %alloc.1 1424 // (Also works for arrays instead of structs) 1425 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1426 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1427 new StoreInst(Extract, NewElts[i], SI); 1428 } 1429 DeadInsts.push_back(SI); 1430 } else if (SIType->isIntegerTy() && 1431 TD->getTypeAllocSize(SIType) == 1432 TD->getTypeAllocSize(AI->getAllocatedType())) { 1433 // If this is a store of the entire alloca from an integer, rewrite it. 1434 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1435 } 1436 } 1437 } 1438} 1439 1440/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1441/// and recursively continue updating all of its uses. 1442void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1443 SmallVector<AllocaInst*, 32> &NewElts) { 1444 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1445 if (BC->getOperand(0) != AI) 1446 return; 1447 1448 // The bitcast references the original alloca. Replace its uses with 1449 // references to the first new element alloca. 1450 Instruction *Val = NewElts[0]; 1451 if (Val->getType() != BC->getDestTy()) { 1452 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1453 Val->takeName(BC); 1454 } 1455 BC->replaceAllUsesWith(Val); 1456 DeadInsts.push_back(BC); 1457} 1458 1459/// FindElementAndOffset - Return the index of the element containing Offset 1460/// within the specified type, which must be either a struct or an array. 1461/// Sets T to the type of the element and Offset to the offset within that 1462/// element. IdxTy is set to the type of the index result to be used in a 1463/// GEP instruction. 1464uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1465 const Type *&IdxTy) { 1466 uint64_t Idx = 0; 1467 if (const StructType *ST = dyn_cast<StructType>(T)) { 1468 const StructLayout *Layout = TD->getStructLayout(ST); 1469 Idx = Layout->getElementContainingOffset(Offset); 1470 T = ST->getContainedType(Idx); 1471 Offset -= Layout->getElementOffset(Idx); 1472 IdxTy = Type::getInt32Ty(T->getContext()); 1473 return Idx; 1474 } 1475 const ArrayType *AT = cast<ArrayType>(T); 1476 T = AT->getElementType(); 1477 uint64_t EltSize = TD->getTypeAllocSize(T); 1478 Idx = Offset / EltSize; 1479 Offset -= Idx * EltSize; 1480 IdxTy = Type::getInt64Ty(T->getContext()); 1481 return Idx; 1482} 1483 1484/// RewriteGEP - Check if this GEP instruction moves the pointer across 1485/// elements of the alloca that are being split apart, and if so, rewrite 1486/// the GEP to be relative to the new element. 1487void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1488 SmallVector<AllocaInst*, 32> &NewElts) { 1489 uint64_t OldOffset = Offset; 1490 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1491 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1492 &Indices[0], Indices.size()); 1493 1494 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1495 1496 const Type *T = AI->getAllocatedType(); 1497 const Type *IdxTy; 1498 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1499 if (GEPI->getOperand(0) == AI) 1500 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1501 1502 T = AI->getAllocatedType(); 1503 uint64_t EltOffset = Offset; 1504 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1505 1506 // If this GEP does not move the pointer across elements of the alloca 1507 // being split, then it does not needs to be rewritten. 1508 if (Idx == OldIdx) 1509 return; 1510 1511 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1512 SmallVector<Value*, 8> NewArgs; 1513 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1514 while (EltOffset != 0) { 1515 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1516 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1517 } 1518 Instruction *Val = NewElts[Idx]; 1519 if (NewArgs.size() > 1) { 1520 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1521 NewArgs.end(), "", GEPI); 1522 Val->takeName(GEPI); 1523 } 1524 if (Val->getType() != GEPI->getType()) 1525 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1526 GEPI->replaceAllUsesWith(Val); 1527 DeadInsts.push_back(GEPI); 1528} 1529 1530/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1531/// Rewrite it to copy or set the elements of the scalarized memory. 1532void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1533 AllocaInst *AI, 1534 SmallVector<AllocaInst*, 32> &NewElts) { 1535 // If this is a memcpy/memmove, construct the other pointer as the 1536 // appropriate type. The "Other" pointer is the pointer that goes to memory 1537 // that doesn't have anything to do with the alloca that we are promoting. For 1538 // memset, this Value* stays null. 1539 Value *OtherPtr = 0; 1540 unsigned MemAlignment = MI->getAlignment(); 1541 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1542 if (Inst == MTI->getRawDest()) 1543 OtherPtr = MTI->getRawSource(); 1544 else { 1545 assert(Inst == MTI->getRawSource()); 1546 OtherPtr = MTI->getRawDest(); 1547 } 1548 } 1549 1550 // If there is an other pointer, we want to convert it to the same pointer 1551 // type as AI has, so we can GEP through it safely. 1552 if (OtherPtr) { 1553 unsigned AddrSpace = 1554 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 1555 1556 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1557 // optimization, but it's also required to detect the corner case where 1558 // both pointer operands are referencing the same memory, and where 1559 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1560 // function is only called for mem intrinsics that access the whole 1561 // aggregate, so non-zero GEPs are not an issue here.) 1562 OtherPtr = OtherPtr->stripPointerCasts(); 1563 1564 // Copying the alloca to itself is a no-op: just delete it. 1565 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1566 // This code will run twice for a no-op memcpy -- once for each operand. 1567 // Put only one reference to MI on the DeadInsts list. 1568 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1569 E = DeadInsts.end(); I != E; ++I) 1570 if (*I == MI) return; 1571 DeadInsts.push_back(MI); 1572 return; 1573 } 1574 1575 // If the pointer is not the right type, insert a bitcast to the right 1576 // type. 1577 const Type *NewTy = 1578 PointerType::get(AI->getType()->getElementType(), AddrSpace); 1579 1580 if (OtherPtr->getType() != NewTy) 1581 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 1582 } 1583 1584 // Process each element of the aggregate. 1585 bool SROADest = MI->getRawDest() == Inst; 1586 1587 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1588 1589 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1590 // If this is a memcpy/memmove, emit a GEP of the other element address. 1591 Value *OtherElt = 0; 1592 unsigned OtherEltAlign = MemAlignment; 1593 1594 if (OtherPtr) { 1595 Value *Idx[2] = { Zero, 1596 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1597 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1598 OtherPtr->getName()+"."+Twine(i), 1599 MI); 1600 uint64_t EltOffset; 1601 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1602 const Type *OtherTy = OtherPtrTy->getElementType(); 1603 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1604 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1605 } else { 1606 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 1607 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1608 } 1609 1610 // The alignment of the other pointer is the guaranteed alignment of the 1611 // element, which is affected by both the known alignment of the whole 1612 // mem intrinsic and the alignment of the element. If the alignment of 1613 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 1614 // known alignment is just 4 bytes. 1615 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 1616 } 1617 1618 Value *EltPtr = NewElts[i]; 1619 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 1620 1621 // If we got down to a scalar, insert a load or store as appropriate. 1622 if (EltTy->isSingleValueType()) { 1623 if (isa<MemTransferInst>(MI)) { 1624 if (SROADest) { 1625 // From Other to Alloca. 1626 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 1627 new StoreInst(Elt, EltPtr, MI); 1628 } else { 1629 // From Alloca to Other. 1630 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 1631 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 1632 } 1633 continue; 1634 } 1635 assert(isa<MemSetInst>(MI)); 1636 1637 // If the stored element is zero (common case), just store a null 1638 // constant. 1639 Constant *StoreVal; 1640 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 1641 if (CI->isZero()) { 1642 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 1643 } else { 1644 // If EltTy is a vector type, get the element type. 1645 const Type *ValTy = EltTy->getScalarType(); 1646 1647 // Construct an integer with the right value. 1648 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 1649 APInt OneVal(EltSize, CI->getZExtValue()); 1650 APInt TotalVal(OneVal); 1651 // Set each byte. 1652 for (unsigned i = 0; 8*i < EltSize; ++i) { 1653 TotalVal = TotalVal.shl(8); 1654 TotalVal |= OneVal; 1655 } 1656 1657 // Convert the integer value to the appropriate type. 1658 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 1659 if (ValTy->isPointerTy()) 1660 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 1661 else if (ValTy->isFloatingPointTy()) 1662 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 1663 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 1664 1665 // If the requested value was a vector constant, create it. 1666 if (EltTy != ValTy) { 1667 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 1668 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 1669 StoreVal = ConstantVector::get(&Elts[0], NumElts); 1670 } 1671 } 1672 new StoreInst(StoreVal, EltPtr, MI); 1673 continue; 1674 } 1675 // Otherwise, if we're storing a byte variable, use a memset call for 1676 // this element. 1677 } 1678 1679 unsigned EltSize = TD->getTypeAllocSize(EltTy); 1680 1681 IRBuilder<> Builder(MI); 1682 1683 // Finally, insert the meminst for this element. 1684 if (isa<MemSetInst>(MI)) { 1685 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 1686 MI->isVolatile()); 1687 } else { 1688 assert(isa<MemTransferInst>(MI)); 1689 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 1690 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 1691 1692 if (isa<MemCpyInst>(MI)) 1693 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 1694 else 1695 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 1696 } 1697 } 1698 DeadInsts.push_back(MI); 1699} 1700 1701/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 1702/// overwrites the entire allocation. Extract out the pieces of the stored 1703/// integer and store them individually. 1704void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 1705 SmallVector<AllocaInst*, 32> &NewElts){ 1706 // Extract each element out of the integer according to its structure offset 1707 // and store the element value to the individual alloca. 1708 Value *SrcVal = SI->getOperand(0); 1709 const Type *AllocaEltTy = AI->getAllocatedType(); 1710 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1711 1712 // Handle tail padding by extending the operand 1713 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 1714 SrcVal = new ZExtInst(SrcVal, 1715 IntegerType::get(SI->getContext(), AllocaSizeBits), 1716 "", SI); 1717 1718 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 1719 << '\n'); 1720 1721 // There are two forms here: AI could be an array or struct. Both cases 1722 // have different ways to compute the element offset. 1723 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1724 const StructLayout *Layout = TD->getStructLayout(EltSTy); 1725 1726 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1727 // Get the number of bits to shift SrcVal to get the value. 1728 const Type *FieldTy = EltSTy->getElementType(i); 1729 uint64_t Shift = Layout->getElementOffsetInBits(i); 1730 1731 if (TD->isBigEndian()) 1732 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 1733 1734 Value *EltVal = SrcVal; 1735 if (Shift) { 1736 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1737 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1738 "sroa.store.elt", SI); 1739 } 1740 1741 // Truncate down to an integer of the right size. 1742 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1743 1744 // Ignore zero sized fields like {}, they obviously contain no data. 1745 if (FieldSizeBits == 0) continue; 1746 1747 if (FieldSizeBits != AllocaSizeBits) 1748 EltVal = new TruncInst(EltVal, 1749 IntegerType::get(SI->getContext(), FieldSizeBits), 1750 "", SI); 1751 Value *DestField = NewElts[i]; 1752 if (EltVal->getType() == FieldTy) { 1753 // Storing to an integer field of this size, just do it. 1754 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 1755 // Bitcast to the right element type (for fp/vector values). 1756 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 1757 } else { 1758 // Otherwise, bitcast the dest pointer (for aggregates). 1759 DestField = new BitCastInst(DestField, 1760 PointerType::getUnqual(EltVal->getType()), 1761 "", SI); 1762 } 1763 new StoreInst(EltVal, DestField, SI); 1764 } 1765 1766 } else { 1767 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1768 const Type *ArrayEltTy = ATy->getElementType(); 1769 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1770 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1771 1772 uint64_t Shift; 1773 1774 if (TD->isBigEndian()) 1775 Shift = AllocaSizeBits-ElementOffset; 1776 else 1777 Shift = 0; 1778 1779 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1780 // Ignore zero sized fields like {}, they obviously contain no data. 1781 if (ElementSizeBits == 0) continue; 1782 1783 Value *EltVal = SrcVal; 1784 if (Shift) { 1785 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1786 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1787 "sroa.store.elt", SI); 1788 } 1789 1790 // Truncate down to an integer of the right size. 1791 if (ElementSizeBits != AllocaSizeBits) 1792 EltVal = new TruncInst(EltVal, 1793 IntegerType::get(SI->getContext(), 1794 ElementSizeBits), "", SI); 1795 Value *DestField = NewElts[i]; 1796 if (EltVal->getType() == ArrayEltTy) { 1797 // Storing to an integer field of this size, just do it. 1798 } else if (ArrayEltTy->isFloatingPointTy() || 1799 ArrayEltTy->isVectorTy()) { 1800 // Bitcast to the right element type (for fp/vector values). 1801 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1802 } else { 1803 // Otherwise, bitcast the dest pointer (for aggregates). 1804 DestField = new BitCastInst(DestField, 1805 PointerType::getUnqual(EltVal->getType()), 1806 "", SI); 1807 } 1808 new StoreInst(EltVal, DestField, SI); 1809 1810 if (TD->isBigEndian()) 1811 Shift -= ElementOffset; 1812 else 1813 Shift += ElementOffset; 1814 } 1815 } 1816 1817 DeadInsts.push_back(SI); 1818} 1819 1820/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1821/// an integer. Load the individual pieces to form the aggregate value. 1822void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1823 SmallVector<AllocaInst*, 32> &NewElts) { 1824 // Extract each element out of the NewElts according to its structure offset 1825 // and form the result value. 1826 const Type *AllocaEltTy = AI->getAllocatedType(); 1827 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1828 1829 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1830 << '\n'); 1831 1832 // There are two forms here: AI could be an array or struct. Both cases 1833 // have different ways to compute the element offset. 1834 const StructLayout *Layout = 0; 1835 uint64_t ArrayEltBitOffset = 0; 1836 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1837 Layout = TD->getStructLayout(EltSTy); 1838 } else { 1839 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1840 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1841 } 1842 1843 Value *ResultVal = 1844 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1845 1846 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1847 // Load the value from the alloca. If the NewElt is an aggregate, cast 1848 // the pointer to an integer of the same size before doing the load. 1849 Value *SrcField = NewElts[i]; 1850 const Type *FieldTy = 1851 cast<PointerType>(SrcField->getType())->getElementType(); 1852 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1853 1854 // Ignore zero sized fields like {}, they obviously contain no data. 1855 if (FieldSizeBits == 0) continue; 1856 1857 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1858 FieldSizeBits); 1859 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1860 !FieldTy->isVectorTy()) 1861 SrcField = new BitCastInst(SrcField, 1862 PointerType::getUnqual(FieldIntTy), 1863 "", LI); 1864 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1865 1866 // If SrcField is a fp or vector of the right size but that isn't an 1867 // integer type, bitcast to an integer so we can shift it. 1868 if (SrcField->getType() != FieldIntTy) 1869 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1870 1871 // Zero extend the field to be the same size as the final alloca so that 1872 // we can shift and insert it. 1873 if (SrcField->getType() != ResultVal->getType()) 1874 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1875 1876 // Determine the number of bits to shift SrcField. 1877 uint64_t Shift; 1878 if (Layout) // Struct case. 1879 Shift = Layout->getElementOffsetInBits(i); 1880 else // Array case. 1881 Shift = i*ArrayEltBitOffset; 1882 1883 if (TD->isBigEndian()) 1884 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1885 1886 if (Shift) { 1887 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1888 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1889 } 1890 1891 // Don't create an 'or x, 0' on the first iteration. 1892 if (!isa<Constant>(ResultVal) || 1893 !cast<Constant>(ResultVal)->isNullValue()) 1894 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1895 else 1896 ResultVal = SrcField; 1897 } 1898 1899 // Handle tail padding by truncating the result 1900 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1901 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1902 1903 LI->replaceAllUsesWith(ResultVal); 1904 DeadInsts.push_back(LI); 1905} 1906 1907/// HasPadding - Return true if the specified type has any structure or 1908/// alignment padding in between the elements that would be split apart 1909/// by SROA; return false otherwise. 1910static bool HasPadding(const Type *Ty, const TargetData &TD) { 1911 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1912 Ty = ATy->getElementType(); 1913 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1914 } 1915 1916 // SROA currently handles only Arrays and Structs. 1917 const StructType *STy = cast<StructType>(Ty); 1918 const StructLayout *SL = TD.getStructLayout(STy); 1919 unsigned PrevFieldBitOffset = 0; 1920 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1921 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1922 1923 // Check to see if there is any padding between this element and the 1924 // previous one. 1925 if (i) { 1926 unsigned PrevFieldEnd = 1927 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1928 if (PrevFieldEnd < FieldBitOffset) 1929 return true; 1930 } 1931 PrevFieldBitOffset = FieldBitOffset; 1932 } 1933 // Check for tail padding. 1934 if (unsigned EltCount = STy->getNumElements()) { 1935 unsigned PrevFieldEnd = PrevFieldBitOffset + 1936 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1937 if (PrevFieldEnd < SL->getSizeInBits()) 1938 return true; 1939 } 1940 return false; 1941} 1942 1943/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1944/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1945/// or 1 if safe after canonicalization has been performed. 1946bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1947 // Loop over the use list of the alloca. We can only transform it if all of 1948 // the users are safe to transform. 1949 AllocaInfo Info; 1950 1951 isSafeForScalarRepl(AI, AI, 0, Info); 1952 if (Info.isUnsafe) { 1953 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1954 return false; 1955 } 1956 1957 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1958 // source and destination, we have to be careful. In particular, the memcpy 1959 // could be moving around elements that live in structure padding of the LLVM 1960 // types, but may actually be used. In these cases, we refuse to promote the 1961 // struct. 1962 if (Info.isMemCpySrc && Info.isMemCpyDst && 1963 HasPadding(AI->getAllocatedType(), *TD)) 1964 return false; 1965 1966 return true; 1967} 1968 1969 1970 1971/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1972/// some part of a constant global variable. This intentionally only accepts 1973/// constant expressions because we don't can't rewrite arbitrary instructions. 1974static bool PointsToConstantGlobal(Value *V) { 1975 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1976 return GV->isConstant(); 1977 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1978 if (CE->getOpcode() == Instruction::BitCast || 1979 CE->getOpcode() == Instruction::GetElementPtr) 1980 return PointsToConstantGlobal(CE->getOperand(0)); 1981 return false; 1982} 1983 1984/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1985/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1986/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1987/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1988/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1989/// the alloca, and if the source pointer is a pointer to a constant global, we 1990/// can optimize this. 1991static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 1992 bool isOffset) { 1993 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1994 User *U = cast<Instruction>(*UI); 1995 1996 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1997 // Ignore non-volatile loads, they are always ok. 1998 if (LI->isVolatile()) return false; 1999 continue; 2000 } 2001 2002 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2003 // If uses of the bitcast are ok, we are ok. 2004 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 2005 return false; 2006 continue; 2007 } 2008 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2009 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2010 // doesn't, it does. 2011 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2012 isOffset || !GEP->hasAllZeroIndices())) 2013 return false; 2014 continue; 2015 } 2016 2017 if (CallSite CS = U) { 2018 // If this is a readonly/readnone call site, then we know it is just a 2019 // load and we can ignore it. 2020 if (CS.onlyReadsMemory()) 2021 continue; 2022 2023 // If this is the function being called then we treat it like a load and 2024 // ignore it. 2025 if (CS.isCallee(UI)) 2026 continue; 2027 2028 // If this is being passed as a byval argument, the caller is making a 2029 // copy, so it is only a read of the alloca. 2030 unsigned ArgNo = CS.getArgumentNo(UI); 2031 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2032 continue; 2033 } 2034 2035 // If this is isn't our memcpy/memmove, reject it as something we can't 2036 // handle. 2037 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2038 if (MI == 0) 2039 return false; 2040 2041 // If the transfer is using the alloca as a source of the transfer, then 2042 // ignore it since it is a load (unless the transfer is volatile). 2043 if (UI.getOperandNo() == 1) { 2044 if (MI->isVolatile()) return false; 2045 continue; 2046 } 2047 2048 // If we already have seen a copy, reject the second one. 2049 if (TheCopy) return false; 2050 2051 // If the pointer has been offset from the start of the alloca, we can't 2052 // safely handle this. 2053 if (isOffset) return false; 2054 2055 // If the memintrinsic isn't using the alloca as the dest, reject it. 2056 if (UI.getOperandNo() != 0) return false; 2057 2058 // If the source of the memcpy/move is not a constant global, reject it. 2059 if (!PointsToConstantGlobal(MI->getSource())) 2060 return false; 2061 2062 // Otherwise, the transform is safe. Remember the copy instruction. 2063 TheCopy = MI; 2064 } 2065 return true; 2066} 2067 2068/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2069/// modified by a copy from a constant global. If we can prove this, we can 2070/// replace any uses of the alloca with uses of the global directly. 2071MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 2072 MemTransferInst *TheCopy = 0; 2073 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 2074 return TheCopy; 2075 return 0; 2076} 2077