ScalarReplAggregates.cpp revision be0761c8202405cdd33f1103d262c0aa97895a8e
1//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This transformation implements the well known scalar replacement of 11// aggregates transformation. This xform breaks up alloca instructions of 12// aggregate type (structure or array) into individual alloca instructions for 13// each member (if possible). Then, if possible, it transforms the individual 14// alloca instructions into nice clean scalar SSA form. 15// 16// This combines a simple SRoA algorithm with the Mem2Reg algorithm because 17// often interact, especially for C++ programs. As such, iterating between 18// SRoA, then Mem2Reg until we run out of things to promote works well. 19// 20//===----------------------------------------------------------------------===// 21 22#define DEBUG_TYPE "scalarrepl" 23#include "llvm/Transforms/Scalar.h" 24#include "llvm/Constants.h" 25#include "llvm/DerivedTypes.h" 26#include "llvm/Function.h" 27#include "llvm/GlobalVariable.h" 28#include "llvm/Instructions.h" 29#include "llvm/IntrinsicInst.h" 30#include "llvm/LLVMContext.h" 31#include "llvm/Module.h" 32#include "llvm/Pass.h" 33#include "llvm/Analysis/Dominators.h" 34#include "llvm/Analysis/Loads.h" 35#include "llvm/Analysis/ValueTracking.h" 36#include "llvm/Target/TargetData.h" 37#include "llvm/Transforms/Utils/PromoteMemToReg.h" 38#include "llvm/Transforms/Utils/Local.h" 39#include "llvm/Transforms/Utils/SSAUpdater.h" 40#include "llvm/Support/CallSite.h" 41#include "llvm/Support/Debug.h" 42#include "llvm/Support/ErrorHandling.h" 43#include "llvm/Support/GetElementPtrTypeIterator.h" 44#include "llvm/Support/IRBuilder.h" 45#include "llvm/Support/MathExtras.h" 46#include "llvm/Support/raw_ostream.h" 47#include "llvm/ADT/SetVector.h" 48#include "llvm/ADT/SmallVector.h" 49#include "llvm/ADT/Statistic.h" 50using namespace llvm; 51 52STATISTIC(NumReplaced, "Number of allocas broken up"); 53STATISTIC(NumPromoted, "Number of allocas promoted"); 54STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); 55STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 56STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 57 58namespace { 59 struct SROA : public FunctionPass { 60 SROA(int T, bool hasDT, char &ID) 61 : FunctionPass(ID), HasDomTree(hasDT) { 62 if (T == -1) 63 SRThreshold = 128; 64 else 65 SRThreshold = T; 66 } 67 68 bool runOnFunction(Function &F); 69 70 bool performScalarRepl(Function &F); 71 bool performPromotion(Function &F); 72 73 private: 74 bool HasDomTree; 75 TargetData *TD; 76 77 /// DeadInsts - Keep track of instructions we have made dead, so that 78 /// we can remove them after we are done working. 79 SmallVector<Value*, 32> DeadInsts; 80 81 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 82 /// information about the uses. All these fields are initialized to false 83 /// and set to true when something is learned. 84 struct AllocaInfo { 85 /// The alloca to promote. 86 AllocaInst *AI; 87 88 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite 89 /// looping and avoid redundant work. 90 SmallPtrSet<PHINode*, 8> CheckedPHIs; 91 92 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 93 bool isUnsafe : 1; 94 95 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 96 bool isMemCpySrc : 1; 97 98 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 99 bool isMemCpyDst : 1; 100 101 /// hasSubelementAccess - This is true if a subelement of the alloca is 102 /// ever accessed, or false if the alloca is only accessed with mem 103 /// intrinsics or load/store that only access the entire alloca at once. 104 bool hasSubelementAccess : 1; 105 106 /// hasALoadOrStore - This is true if there are any loads or stores to it. 107 /// The alloca may just be accessed with memcpy, for example, which would 108 /// not set this. 109 bool hasALoadOrStore : 1; 110 111 explicit AllocaInfo(AllocaInst *ai) 112 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 113 hasSubelementAccess(false), hasALoadOrStore(false) {} 114 }; 115 116 unsigned SRThreshold; 117 118 void MarkUnsafe(AllocaInfo &I, Instruction *User) { 119 I.isUnsafe = true; 120 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); 121 } 122 123 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 124 125 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info); 126 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset, 127 AllocaInfo &Info); 128 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info); 129 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 130 const Type *MemOpType, bool isStore, AllocaInfo &Info, 131 Instruction *TheAccess, bool AllowWholeAccess); 132 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 133 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 134 const Type *&IdxTy); 135 136 void DoScalarReplacement(AllocaInst *AI, 137 std::vector<AllocaInst*> &WorkList); 138 void DeleteDeadInstructions(); 139 140 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 141 SmallVector<AllocaInst*, 32> &NewElts); 142 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 143 SmallVector<AllocaInst*, 32> &NewElts); 144 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 145 SmallVector<AllocaInst*, 32> &NewElts); 146 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 147 AllocaInst *AI, 148 SmallVector<AllocaInst*, 32> &NewElts); 149 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 150 SmallVector<AllocaInst*, 32> &NewElts); 151 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 152 SmallVector<AllocaInst*, 32> &NewElts); 153 154 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 155 }; 156 157 // SROA_DT - SROA that uses DominatorTree. 158 struct SROA_DT : public SROA { 159 static char ID; 160 public: 161 SROA_DT(int T = -1) : SROA(T, true, ID) { 162 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 163 } 164 165 // getAnalysisUsage - This pass does not require any passes, but we know it 166 // will not alter the CFG, so say so. 167 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 168 AU.addRequired<DominatorTree>(); 169 AU.setPreservesCFG(); 170 } 171 }; 172 173 // SROA_SSAUp - SROA that uses SSAUpdater. 174 struct SROA_SSAUp : public SROA { 175 static char ID; 176 public: 177 SROA_SSAUp(int T = -1) : SROA(T, false, ID) { 178 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 179 } 180 181 // getAnalysisUsage - This pass does not require any passes, but we know it 182 // will not alter the CFG, so say so. 183 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 184 AU.setPreservesCFG(); 185 } 186 }; 187 188} 189 190char SROA_DT::ID = 0; 191char SROA_SSAUp::ID = 0; 192 193INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 194 "Scalar Replacement of Aggregates (DT)", false, false) 195INITIALIZE_PASS_DEPENDENCY(DominatorTree) 196INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 197 "Scalar Replacement of Aggregates (DT)", false, false) 198 199INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 200 "Scalar Replacement of Aggregates (SSAUp)", false, false) 201INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 202 "Scalar Replacement of Aggregates (SSAUp)", false, false) 203 204// Public interface to the ScalarReplAggregates pass 205FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 206 bool UseDomTree) { 207 if (UseDomTree) 208 return new SROA_DT(Threshold); 209 return new SROA_SSAUp(Threshold); 210} 211 212 213//===----------------------------------------------------------------------===// 214// Convert To Scalar Optimization. 215//===----------------------------------------------------------------------===// 216 217namespace { 218/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 219/// optimization, which scans the uses of an alloca and determines if it can 220/// rewrite it in terms of a single new alloca that can be mem2reg'd. 221class ConvertToScalarInfo { 222 /// AllocaSize - The size of the alloca being considered in bytes. 223 unsigned AllocaSize; 224 const TargetData &TD; 225 226 /// IsNotTrivial - This is set to true if there is some access to the object 227 /// which means that mem2reg can't promote it. 228 bool IsNotTrivial; 229 230 /// VectorTy - This tracks the type that we should promote the vector to if 231 /// it is possible to turn it into a vector. This starts out null, and if it 232 /// isn't possible to turn into a vector type, it gets set to VoidTy. 233 const Type *VectorTy; 234 235 /// HadAVector - True if there is at least one vector access to the alloca. 236 /// We don't want to turn random arrays into vectors and use vector element 237 /// insert/extract, but if there are element accesses to something that is 238 /// also declared as a vector, we do want to promote to a vector. 239 bool HadAVector; 240 241 /// HadNonMemTransferAccess - True if there is at least one access to the 242 /// alloca that is not a MemTransferInst. We don't want to turn structs into 243 /// large integers unless there is some potential for optimization. 244 bool HadNonMemTransferAccess; 245 246public: 247 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 248 : AllocaSize(Size), TD(td), IsNotTrivial(false), VectorTy(0), 249 HadAVector(false), HadNonMemTransferAccess(false) { } 250 251 AllocaInst *TryConvert(AllocaInst *AI); 252 253private: 254 bool CanConvertToScalar(Value *V, uint64_t Offset); 255 void MergeInType(const Type *In, uint64_t Offset, bool IsLoadOrStore); 256 bool MergeInVectorType(const VectorType *VInTy, uint64_t Offset); 257 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 258 259 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 260 uint64_t Offset, IRBuilder<> &Builder); 261 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 262 uint64_t Offset, IRBuilder<> &Builder); 263}; 264} // end anonymous namespace. 265 266 267/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 268/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 269/// alloca if possible or null if not. 270AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 271 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 272 // out. 273 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 274 return 0; 275 276 // If we were able to find a vector type that can handle this with 277 // insert/extract elements, and if there was at least one use that had 278 // a vector type, promote this to a vector. We don't want to promote 279 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 280 // we just get a lot of insert/extracts. If at least one vector is 281 // involved, then we probably really do have a union of vector/array. 282 const Type *NewTy; 283 if (VectorTy && VectorTy->isVectorTy() && HadAVector) { 284 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 285 << *VectorTy << '\n'); 286 NewTy = VectorTy; // Use the vector type. 287 } else { 288 unsigned BitWidth = AllocaSize * 8; 289 if (!HadAVector && !HadNonMemTransferAccess && 290 !TD.fitsInLegalInteger(BitWidth)) 291 return 0; 292 293 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 294 // Create and insert the integer alloca. 295 NewTy = IntegerType::get(AI->getContext(), BitWidth); 296 } 297 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 298 ConvertUsesToScalar(AI, NewAI, 0); 299 return NewAI; 300} 301 302/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy) 303/// so far at the offset specified by Offset (which is specified in bytes). 304/// 305/// There are three cases we handle here: 306/// 1) A union of vector types of the same size and potentially its elements. 307/// Here we turn element accesses into insert/extract element operations. 308/// This promotes a <4 x float> with a store of float to the third element 309/// into a <4 x float> that uses insert element. 310/// 2) A union of vector types with power-of-2 size differences, e.g. a float, 311/// <2 x float> and <4 x float>. Here we turn element accesses into insert 312/// and extract element operations, and <2 x float> accesses into a cast to 313/// <2 x double>, an extract, and a cast back to <2 x float>. 314/// 3) A fully general blob of memory, which we turn into some (potentially 315/// large) integer type with extract and insert operations where the loads 316/// and stores would mutate the memory. We mark this by setting VectorTy 317/// to VoidTy. 318void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset, 319 bool IsLoadOrStore) { 320 // If we already decided to turn this into a blob of integer memory, there is 321 // nothing to be done. 322 if (VectorTy && VectorTy->isVoidTy()) 323 return; 324 325 // If this could be contributing to a vector, analyze it. 326 327 // If the In type is a vector that is the same size as the alloca, see if it 328 // matches the existing VecTy. 329 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 330 if (MergeInVectorType(VInTy, Offset)) 331 return; 332 } else if (In->isFloatTy() || In->isDoubleTy() || 333 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 334 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 335 // Full width accesses can be ignored, because they can always be turned 336 // into bitcasts. 337 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 338 if (IsLoadOrStore && EltSize == AllocaSize) 339 return; 340 // If we're accessing something that could be an element of a vector, see 341 // if the implied vector agrees with what we already have and if Offset is 342 // compatible with it. 343 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 344 (VectorTy == 0 || 345 cast<VectorType>(VectorTy)->getElementType() 346 ->getPrimitiveSizeInBits()/8 == EltSize)) { 347 if (VectorTy == 0) 348 VectorTy = VectorType::get(In, AllocaSize/EltSize); 349 return; 350 } 351 } 352 353 // Otherwise, we have a case that we can't handle with an optimized vector 354 // form. We can still turn this into a large integer. 355 VectorTy = Type::getVoidTy(In->getContext()); 356} 357 358/// MergeInVectorType - Handles the vector case of MergeInType, returning true 359/// if the type was successfully merged and false otherwise. 360bool ConvertToScalarInfo::MergeInVectorType(const VectorType *VInTy, 361 uint64_t Offset) { 362 // Remember if we saw a vector type. 363 HadAVector = true; 364 365 // TODO: Support nonzero offsets? 366 if (Offset != 0) 367 return false; 368 369 // Only allow vectors that are a power-of-2 away from the size of the alloca. 370 if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8))) 371 return false; 372 373 // If this the first vector we see, remember the type so that we know the 374 // element size. 375 if (!VectorTy) { 376 VectorTy = VInTy; 377 return true; 378 } 379 380 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth(); 381 unsigned InBitWidth = VInTy->getBitWidth(); 382 383 // Vectors of the same size can be converted using a simple bitcast. 384 if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8)) 385 return true; 386 387 const Type *ElementTy = cast<VectorType>(VectorTy)->getElementType(); 388 const Type *InElementTy = cast<VectorType>(VInTy)->getElementType(); 389 390 // Do not allow mixed integer and floating-point accesses from vectors of 391 // different sizes. 392 if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy()) 393 return false; 394 395 if (ElementTy->isFloatingPointTy()) { 396 // Only allow floating-point vectors of different sizes if they have the 397 // same element type. 398 // TODO: This could be loosened a bit, but would anything benefit? 399 if (ElementTy != InElementTy) 400 return false; 401 402 // There are no arbitrary-precision floating-point types, which limits the 403 // number of legal vector types with larger element types that we can form 404 // to bitcast and extract a subvector. 405 // TODO: We could support some more cases with mixed fp128 and double here. 406 if (!(BitWidth == 64 || BitWidth == 128) || 407 !(InBitWidth == 64 || InBitWidth == 128)) 408 return false; 409 } else { 410 assert(ElementTy->isIntegerTy() && "Vector elements must be either integer " 411 "or floating-point."); 412 unsigned BitWidth = ElementTy->getPrimitiveSizeInBits(); 413 unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits(); 414 415 // Do not allow integer types smaller than a byte or types whose widths are 416 // not a multiple of a byte. 417 if (BitWidth < 8 || InBitWidth < 8 || 418 BitWidth % 8 != 0 || InBitWidth % 8 != 0) 419 return false; 420 } 421 422 // Pick the largest of the two vector types. 423 if (InBitWidth > BitWidth) 424 VectorTy = VInTy; 425 426 return true; 427} 428 429/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 430/// its accesses to a single vector type, return true and set VecTy to 431/// the new type. If we could convert the alloca into a single promotable 432/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 433/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 434/// is the current offset from the base of the alloca being analyzed. 435/// 436/// If we see at least one access to the value that is as a vector type, set the 437/// SawVec flag. 438bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 439 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 440 Instruction *User = cast<Instruction>(*UI); 441 442 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 443 // Don't break volatile loads. 444 if (LI->isVolatile()) 445 return false; 446 // Don't touch MMX operations. 447 if (LI->getType()->isX86_MMXTy()) 448 return false; 449 HadNonMemTransferAccess = true; 450 MergeInType(LI->getType(), Offset, true); 451 continue; 452 } 453 454 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 455 // Storing the pointer, not into the value? 456 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 457 // Don't touch MMX operations. 458 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 459 return false; 460 HadNonMemTransferAccess = true; 461 MergeInType(SI->getOperand(0)->getType(), Offset, true); 462 continue; 463 } 464 465 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 466 IsNotTrivial = true; // Can't be mem2reg'd. 467 if (!CanConvertToScalar(BCI, Offset)) 468 return false; 469 continue; 470 } 471 472 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 473 // If this is a GEP with a variable indices, we can't handle it. 474 if (!GEP->hasAllConstantIndices()) 475 return false; 476 477 // Compute the offset that this GEP adds to the pointer. 478 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 479 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 480 &Indices[0], Indices.size()); 481 // See if all uses can be converted. 482 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 483 return false; 484 IsNotTrivial = true; // Can't be mem2reg'd. 485 HadNonMemTransferAccess = true; 486 continue; 487 } 488 489 // If this is a constant sized memset of a constant value (e.g. 0) we can 490 // handle it. 491 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 492 // Store of constant value and constant size. 493 if (!isa<ConstantInt>(MSI->getValue()) || 494 !isa<ConstantInt>(MSI->getLength())) 495 return false; 496 IsNotTrivial = true; // Can't be mem2reg'd. 497 HadNonMemTransferAccess = true; 498 continue; 499 } 500 501 // If this is a memcpy or memmove into or out of the whole allocation, we 502 // can handle it like a load or store of the scalar type. 503 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 504 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 505 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 506 return false; 507 508 IsNotTrivial = true; // Can't be mem2reg'd. 509 continue; 510 } 511 512 // Otherwise, we cannot handle this! 513 return false; 514 } 515 516 return true; 517} 518 519/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 520/// directly. This happens when we are converting an "integer union" to a 521/// single integer scalar, or when we are converting a "vector union" to a 522/// vector with insert/extractelement instructions. 523/// 524/// Offset is an offset from the original alloca, in bits that need to be 525/// shifted to the right. By the end of this, there should be no uses of Ptr. 526void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 527 uint64_t Offset) { 528 while (!Ptr->use_empty()) { 529 Instruction *User = cast<Instruction>(Ptr->use_back()); 530 531 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 532 ConvertUsesToScalar(CI, NewAI, Offset); 533 CI->eraseFromParent(); 534 continue; 535 } 536 537 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 538 // Compute the offset that this GEP adds to the pointer. 539 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 540 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 541 &Indices[0], Indices.size()); 542 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 543 GEP->eraseFromParent(); 544 continue; 545 } 546 547 IRBuilder<> Builder(User); 548 549 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 550 // The load is a bit extract from NewAI shifted right by Offset bits. 551 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 552 Value *NewLoadVal 553 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 554 LI->replaceAllUsesWith(NewLoadVal); 555 LI->eraseFromParent(); 556 continue; 557 } 558 559 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 560 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 561 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 562 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 563 Builder); 564 Builder.CreateStore(New, NewAI); 565 SI->eraseFromParent(); 566 567 // If the load we just inserted is now dead, then the inserted store 568 // overwrote the entire thing. 569 if (Old->use_empty()) 570 Old->eraseFromParent(); 571 continue; 572 } 573 574 // If this is a constant sized memset of a constant value (e.g. 0) we can 575 // transform it into a store of the expanded constant value. 576 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 577 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 578 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 579 if (NumBytes != 0) { 580 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 581 582 // Compute the value replicated the right number of times. 583 APInt APVal(NumBytes*8, Val); 584 585 // Splat the value if non-zero. 586 if (Val) 587 for (unsigned i = 1; i != NumBytes; ++i) 588 APVal |= APVal << 8; 589 590 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 591 Value *New = ConvertScalar_InsertValue( 592 ConstantInt::get(User->getContext(), APVal), 593 Old, Offset, Builder); 594 Builder.CreateStore(New, NewAI); 595 596 // If the load we just inserted is now dead, then the memset overwrote 597 // the entire thing. 598 if (Old->use_empty()) 599 Old->eraseFromParent(); 600 } 601 MSI->eraseFromParent(); 602 continue; 603 } 604 605 // If this is a memcpy or memmove into or out of the whole allocation, we 606 // can handle it like a load or store of the scalar type. 607 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 608 assert(Offset == 0 && "must be store to start of alloca"); 609 610 // If the source and destination are both to the same alloca, then this is 611 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 612 // as appropriate. 613 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 614 615 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 616 // Dest must be OrigAI, change this to be a load from the original 617 // pointer (bitcasted), then a store to our new alloca. 618 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 619 Value *SrcPtr = MTI->getSource(); 620 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 621 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 622 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 623 AIPTy = PointerType::get(AIPTy->getElementType(), 624 SPTy->getAddressSpace()); 625 } 626 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 627 628 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 629 SrcVal->setAlignment(MTI->getAlignment()); 630 Builder.CreateStore(SrcVal, NewAI); 631 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 632 // Src must be OrigAI, change this to be a load from NewAI then a store 633 // through the original dest pointer (bitcasted). 634 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 635 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 636 637 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 638 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 639 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 640 AIPTy = PointerType::get(AIPTy->getElementType(), 641 DPTy->getAddressSpace()); 642 } 643 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 644 645 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 646 NewStore->setAlignment(MTI->getAlignment()); 647 } else { 648 // Noop transfer. Src == Dst 649 } 650 651 MTI->eraseFromParent(); 652 continue; 653 } 654 655 llvm_unreachable("Unsupported operation!"); 656 } 657} 658 659/// getScaledElementType - Gets a scaled element type for a partial vector 660/// access of an alloca. The input type must be an integer or float, and 661/// the resulting type must be an integer, float or double. 662static const Type *getScaledElementType(const Type *OldTy, 663 unsigned NewBitWidth) { 664 assert((OldTy->isIntegerTy() || OldTy->isFloatTy()) && "Partial vector " 665 "accesses must be scaled from integer or float elements."); 666 667 LLVMContext &Context = OldTy->getContext(); 668 669 if (OldTy->isIntegerTy()) 670 return Type::getIntNTy(Context, NewBitWidth); 671 if (NewBitWidth == 32) 672 return Type::getFloatTy(Context); 673 if (NewBitWidth == 64) 674 return Type::getDoubleTy(Context); 675 676 llvm_unreachable("Invalid type for a partial vector access of an alloca!"); 677} 678 679/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 680/// or vector value FromVal, extracting the bits from the offset specified by 681/// Offset. This returns the value, which is of type ToType. 682/// 683/// This happens when we are converting an "integer union" to a single 684/// integer scalar, or when we are converting a "vector union" to a vector with 685/// insert/extractelement instructions. 686/// 687/// Offset is an offset from the original alloca, in bits that need to be 688/// shifted to the right. 689Value *ConvertToScalarInfo:: 690ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 691 uint64_t Offset, IRBuilder<> &Builder) { 692 // If the load is of the whole new alloca, no conversion is needed. 693 const Type *FromType = FromVal->getType(); 694 if (FromType == ToType && Offset == 0) 695 return FromVal; 696 697 // If the result alloca is a vector type, this is either an element 698 // access or a bitcast to another vector type of the same size. 699 if (const VectorType *VTy = dyn_cast<VectorType>(FromType)) { 700 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 701 if (ToTypeSize == AllocaSize) { 702 if (FromType->getPrimitiveSizeInBits() == 703 ToType->getPrimitiveSizeInBits()) 704 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 705 else { 706 // Vectors with the same element type can have the same allocation 707 // size but different primitive sizes (e.g., <3 x i32> and <4 x i32>) 708 // In this case, use a shuffle vector instead of a bit cast. 709 const VectorType *ToVTy = dyn_cast<VectorType>(ToType); 710 assert(ToVTy && (ToVTy->getElementType() == VTy->getElementType()) && 711 "Vectors must have the same element type"); 712 LLVMContext &Context = FromVal->getContext(); 713 Value *UnV = UndefValue::get(FromType); 714 unsigned numEltsFrom = VTy->getNumElements(); 715 unsigned numEltsTo = ToVTy->getNumElements(); 716 717 SmallVector<Constant*, 3> Args; 718 unsigned minNumElts = std::min(numEltsFrom, numEltsTo); 719 unsigned i; 720 for (i=0; i != minNumElts; ++i) 721 Args.push_back(ConstantInt::get(Type::getInt32Ty(Context), i)); 722 723 if (i < numEltsTo) { 724 Constant* UnC = UndefValue::get(Type::getInt32Ty(Context)); 725 for (; i != numEltsTo; ++i) 726 Args.push_back(UnC); 727 } 728 Constant *Mask = ConstantVector::get(Args); 729 return Builder.CreateShuffleVector(FromVal, UnV, Mask, "tmpV"); 730 } 731 } 732 733 if (ToType->isVectorTy()) { 734 assert(isPowerOf2_64(AllocaSize / ToTypeSize) && 735 "Partial vector access of an alloca must have a power-of-2 size " 736 "ratio."); 737 assert(Offset == 0 && "Can't extract a value of a smaller vector type " 738 "from a nonzero offset."); 739 740 const Type *ToElementTy = cast<VectorType>(ToType)->getElementType(); 741 const Type *CastElementTy = getScaledElementType(ToElementTy, 742 ToTypeSize * 8); 743 unsigned NumCastVectorElements = AllocaSize / ToTypeSize; 744 745 LLVMContext &Context = FromVal->getContext(); 746 const Type *CastTy = VectorType::get(CastElementTy, 747 NumCastVectorElements); 748 Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp"); 749 Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get( 750 Type::getInt32Ty(Context), 0), "tmp"); 751 return Builder.CreateBitCast(Extract, ToType, "tmp"); 752 } 753 754 // Otherwise it must be an element access. 755 unsigned Elt = 0; 756 if (Offset) { 757 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 758 Elt = Offset/EltSize; 759 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 760 } 761 // Return the element extracted out of it. 762 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 763 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 764 if (V->getType() != ToType) 765 V = Builder.CreateBitCast(V, ToType, "tmp"); 766 return V; 767 } 768 769 // If ToType is a first class aggregate, extract out each of the pieces and 770 // use insertvalue's to form the FCA. 771 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 772 const StructLayout &Layout = *TD.getStructLayout(ST); 773 Value *Res = UndefValue::get(ST); 774 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 775 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 776 Offset+Layout.getElementOffsetInBits(i), 777 Builder); 778 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 779 } 780 return Res; 781 } 782 783 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 784 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 785 Value *Res = UndefValue::get(AT); 786 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 787 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 788 Offset+i*EltSize, Builder); 789 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 790 } 791 return Res; 792 } 793 794 // Otherwise, this must be a union that was converted to an integer value. 795 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 796 797 // If this is a big-endian system and the load is narrower than the 798 // full alloca type, we need to do a shift to get the right bits. 799 int ShAmt = 0; 800 if (TD.isBigEndian()) { 801 // On big-endian machines, the lowest bit is stored at the bit offset 802 // from the pointer given by getTypeStoreSizeInBits. This matters for 803 // integers with a bitwidth that is not a multiple of 8. 804 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 805 TD.getTypeStoreSizeInBits(ToType) - Offset; 806 } else { 807 ShAmt = Offset; 808 } 809 810 // Note: we support negative bitwidths (with shl) which are not defined. 811 // We do this to support (f.e.) loads off the end of a structure where 812 // only some bits are used. 813 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 814 FromVal = Builder.CreateLShr(FromVal, 815 ConstantInt::get(FromVal->getType(), 816 ShAmt), "tmp"); 817 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 818 FromVal = Builder.CreateShl(FromVal, 819 ConstantInt::get(FromVal->getType(), 820 -ShAmt), "tmp"); 821 822 // Finally, unconditionally truncate the integer to the right width. 823 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 824 if (LIBitWidth < NTy->getBitWidth()) 825 FromVal = 826 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 827 LIBitWidth), "tmp"); 828 else if (LIBitWidth > NTy->getBitWidth()) 829 FromVal = 830 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 831 LIBitWidth), "tmp"); 832 833 // If the result is an integer, this is a trunc or bitcast. 834 if (ToType->isIntegerTy()) { 835 // Should be done. 836 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 837 // Just do a bitcast, we know the sizes match up. 838 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 839 } else { 840 // Otherwise must be a pointer. 841 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 842 } 843 assert(FromVal->getType() == ToType && "Didn't convert right?"); 844 return FromVal; 845} 846 847/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 848/// or vector value "Old" at the offset specified by Offset. 849/// 850/// This happens when we are converting an "integer union" to a 851/// single integer scalar, or when we are converting a "vector union" to a 852/// vector with insert/extractelement instructions. 853/// 854/// Offset is an offset from the original alloca, in bits that need to be 855/// shifted to the right. 856Value *ConvertToScalarInfo:: 857ConvertScalar_InsertValue(Value *SV, Value *Old, 858 uint64_t Offset, IRBuilder<> &Builder) { 859 // Convert the stored type to the actual type, shift it left to insert 860 // then 'or' into place. 861 const Type *AllocaType = Old->getType(); 862 LLVMContext &Context = Old->getContext(); 863 864 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 865 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 866 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 867 868 // Changing the whole vector with memset or with an access of a different 869 // vector type? 870 if (ValSize == VecSize) { 871 if (VTy->getPrimitiveSizeInBits() == 872 SV->getType()->getPrimitiveSizeInBits()) 873 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 874 else { 875 // Vectors with the same element type can have the same allocation 876 // size but different primitive sizes (e.g., <3 x i32> and <4 x i32>) 877 // In this case, use a shuffle vector instead of a bit cast. 878 const VectorType *SVVTy = dyn_cast<VectorType>(SV->getType()); 879 assert(SVVTy && (SVVTy->getElementType() == VTy->getElementType()) && 880 "Vectors must have the same element type"); 881 Value *UnV = UndefValue::get(SVVTy); 882 unsigned numEltsFrom = SVVTy->getNumElements(); 883 unsigned numEltsTo = VTy->getNumElements(); 884 885 SmallVector<Constant*, 3> Args; 886 unsigned minNumElts = std::min(numEltsFrom, numEltsTo); 887 unsigned i; 888 for (i=0; i != minNumElts; ++i) 889 Args.push_back(ConstantInt::get(Type::getInt32Ty(Context), i)); 890 891 if (i < numEltsTo) { 892 Constant* UnC = UndefValue::get(Type::getInt32Ty(Context)); 893 for (; i != numEltsTo; ++i) 894 Args.push_back(UnC); 895 } 896 Constant *Mask = ConstantVector::get(Args); 897 return Builder.CreateShuffleVector(SV, UnV, Mask, "tmpV"); 898 } 899 } 900 901 if (SV->getType()->isVectorTy() && isPowerOf2_64(VecSize / ValSize)) { 902 assert(Offset == 0 && "Can't insert a value of a smaller vector type at " 903 "a nonzero offset."); 904 905 const Type *ToElementTy = 906 cast<VectorType>(SV->getType())->getElementType(); 907 const Type *CastElementTy = getScaledElementType(ToElementTy, ValSize); 908 unsigned NumCastVectorElements = VecSize / ValSize; 909 910 LLVMContext &Context = SV->getContext(); 911 const Type *OldCastTy = VectorType::get(CastElementTy, 912 NumCastVectorElements); 913 Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp"); 914 915 Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp"); 916 Value *Insert = 917 Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get( 918 Type::getInt32Ty(Context), 0), "tmp"); 919 return Builder.CreateBitCast(Insert, AllocaType, "tmp"); 920 } 921 922 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 923 924 // Must be an element insertion. 925 unsigned Elt = Offset/EltSize; 926 927 if (SV->getType() != VTy->getElementType()) 928 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 929 930 SV = Builder.CreateInsertElement(Old, SV, 931 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 932 "tmp"); 933 return SV; 934 } 935 936 // If SV is a first-class aggregate value, insert each value recursively. 937 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 938 const StructLayout &Layout = *TD.getStructLayout(ST); 939 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 940 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 941 Old = ConvertScalar_InsertValue(Elt, Old, 942 Offset+Layout.getElementOffsetInBits(i), 943 Builder); 944 } 945 return Old; 946 } 947 948 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 949 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 950 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 951 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 952 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 953 } 954 return Old; 955 } 956 957 // If SV is a float, convert it to the appropriate integer type. 958 // If it is a pointer, do the same. 959 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 960 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 961 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 962 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 963 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 964 SV = Builder.CreateBitCast(SV, 965 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 966 else if (SV->getType()->isPointerTy()) 967 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 968 969 // Zero extend or truncate the value if needed. 970 if (SV->getType() != AllocaType) { 971 if (SV->getType()->getPrimitiveSizeInBits() < 972 AllocaType->getPrimitiveSizeInBits()) 973 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 974 else { 975 // Truncation may be needed if storing more than the alloca can hold 976 // (undefined behavior). 977 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 978 SrcWidth = DestWidth; 979 SrcStoreWidth = DestStoreWidth; 980 } 981 } 982 983 // If this is a big-endian system and the store is narrower than the 984 // full alloca type, we need to do a shift to get the right bits. 985 int ShAmt = 0; 986 if (TD.isBigEndian()) { 987 // On big-endian machines, the lowest bit is stored at the bit offset 988 // from the pointer given by getTypeStoreSizeInBits. This matters for 989 // integers with a bitwidth that is not a multiple of 8. 990 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 991 } else { 992 ShAmt = Offset; 993 } 994 995 // Note: we support negative bitwidths (with shr) which are not defined. 996 // We do this to support (f.e.) stores off the end of a structure where 997 // only some bits in the structure are set. 998 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 999 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1000 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1001 ShAmt), "tmp"); 1002 Mask <<= ShAmt; 1003 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1004 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1005 -ShAmt), "tmp"); 1006 Mask = Mask.lshr(-ShAmt); 1007 } 1008 1009 // Mask out the bits we are about to insert from the old value, and or 1010 // in the new bits. 1011 if (SrcWidth != DestWidth) { 1012 assert(DestWidth > SrcWidth); 1013 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1014 SV = Builder.CreateOr(Old, SV, "ins"); 1015 } 1016 return SV; 1017} 1018 1019 1020//===----------------------------------------------------------------------===// 1021// SRoA Driver 1022//===----------------------------------------------------------------------===// 1023 1024 1025bool SROA::runOnFunction(Function &F) { 1026 TD = getAnalysisIfAvailable<TargetData>(); 1027 1028 bool Changed = performPromotion(F); 1029 1030 // FIXME: ScalarRepl currently depends on TargetData more than it 1031 // theoretically needs to. It should be refactored in order to support 1032 // target-independent IR. Until this is done, just skip the actual 1033 // scalar-replacement portion of this pass. 1034 if (!TD) return Changed; 1035 1036 while (1) { 1037 bool LocalChange = performScalarRepl(F); 1038 if (!LocalChange) break; // No need to repromote if no scalarrepl 1039 Changed = true; 1040 LocalChange = performPromotion(F); 1041 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 1042 } 1043 1044 return Changed; 1045} 1046 1047namespace { 1048class AllocaPromoter : public LoadAndStorePromoter { 1049 AllocaInst *AI; 1050public: 1051 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S) 1052 : LoadAndStorePromoter(Insts, S), AI(0) {} 1053 1054 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 1055 // Remember which alloca we're promoting (for isInstInList). 1056 this->AI = AI; 1057 LoadAndStorePromoter::run(Insts); 1058 AI->eraseFromParent(); 1059 } 1060 1061 virtual bool isInstInList(Instruction *I, 1062 const SmallVectorImpl<Instruction*> &Insts) const { 1063 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1064 return LI->getOperand(0) == AI; 1065 return cast<StoreInst>(I)->getPointerOperand() == AI; 1066 } 1067}; 1068} // end anon namespace 1069 1070/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1071/// subsequently loaded can be rewritten to load both input pointers and then 1072/// select between the result, allowing the load of the alloca to be promoted. 1073/// From this: 1074/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1075/// %V = load i32* %P2 1076/// to: 1077/// %V1 = load i32* %Alloca -> will be mem2reg'd 1078/// %V2 = load i32* %Other 1079/// %V = select i1 %cond, i32 %V1, i32 %V2 1080/// 1081/// We can do this to a select if its only uses are loads and if the operand to 1082/// the select can be loaded unconditionally. 1083static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1084 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1085 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1086 1087 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1088 UI != UE; ++UI) { 1089 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1090 if (LI == 0 || LI->isVolatile()) return false; 1091 1092 // Both operands to the select need to be dereferencable, either absolutely 1093 // (e.g. allocas) or at this point because we can see other accesses to it. 1094 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1095 LI->getAlignment(), TD)) 1096 return false; 1097 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1098 LI->getAlignment(), TD)) 1099 return false; 1100 } 1101 1102 return true; 1103} 1104 1105/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1106/// subsequently loaded can be rewritten to load both input pointers in the pred 1107/// blocks and then PHI the results, allowing the load of the alloca to be 1108/// promoted. 1109/// From this: 1110/// %P2 = phi [i32* %Alloca, i32* %Other] 1111/// %V = load i32* %P2 1112/// to: 1113/// %V1 = load i32* %Alloca -> will be mem2reg'd 1114/// ... 1115/// %V2 = load i32* %Other 1116/// ... 1117/// %V = phi [i32 %V1, i32 %V2] 1118/// 1119/// We can do this to a select if its only uses are loads and if the operand to 1120/// the select can be loaded unconditionally. 1121static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1122 // For now, we can only do this promotion if the load is in the same block as 1123 // the PHI, and if there are no stores between the phi and load. 1124 // TODO: Allow recursive phi users. 1125 // TODO: Allow stores. 1126 BasicBlock *BB = PN->getParent(); 1127 unsigned MaxAlign = 0; 1128 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1129 UI != UE; ++UI) { 1130 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1131 if (LI == 0 || LI->isVolatile()) return false; 1132 1133 // For now we only allow loads in the same block as the PHI. This is a 1134 // common case that happens when instcombine merges two loads through a PHI. 1135 if (LI->getParent() != BB) return false; 1136 1137 // Ensure that there are no instructions between the PHI and the load that 1138 // could store. 1139 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1140 if (BBI->mayWriteToMemory()) 1141 return false; 1142 1143 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1144 } 1145 1146 // Okay, we know that we have one or more loads in the same block as the PHI. 1147 // We can transform this if it is safe to push the loads into the predecessor 1148 // blocks. The only thing to watch out for is that we can't put a possibly 1149 // trapping load in the predecessor if it is a critical edge. 1150 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1151 BasicBlock *Pred = PN->getIncomingBlock(i); 1152 1153 // If the predecessor has a single successor, then the edge isn't critical. 1154 if (Pred->getTerminator()->getNumSuccessors() == 1) 1155 continue; 1156 1157 Value *InVal = PN->getIncomingValue(i); 1158 1159 // If the InVal is an invoke in the pred, we can't put a load on the edge. 1160 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 1161 if (II->getParent() == Pred) 1162 return false; 1163 1164 // If this pointer is always safe to load, or if we can prove that there is 1165 // already a load in the block, then we can move the load to the pred block. 1166 if (InVal->isDereferenceablePointer() || 1167 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1168 continue; 1169 1170 return false; 1171 } 1172 1173 return true; 1174} 1175 1176 1177/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1178/// direct (non-volatile) loads and stores to it. If the alloca is close but 1179/// not quite there, this will transform the code to allow promotion. As such, 1180/// it is a non-pure predicate. 1181static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1182 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1183 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1184 1185 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1186 UI != UE; ++UI) { 1187 User *U = *UI; 1188 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1189 if (LI->isVolatile()) 1190 return false; 1191 continue; 1192 } 1193 1194 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1195 if (SI->getOperand(0) == AI || SI->isVolatile()) 1196 return false; // Don't allow a store OF the AI, only INTO the AI. 1197 continue; 1198 } 1199 1200 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1201 // If the condition being selected on is a constant, fold the select, yes 1202 // this does (rarely) happen early on. 1203 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1204 Value *Result = SI->getOperand(1+CI->isZero()); 1205 SI->replaceAllUsesWith(Result); 1206 SI->eraseFromParent(); 1207 1208 // This is very rare and we just scrambled the use list of AI, start 1209 // over completely. 1210 return tryToMakeAllocaBePromotable(AI, TD); 1211 } 1212 1213 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1214 // loads, then we can transform this by rewriting the select. 1215 if (!isSafeSelectToSpeculate(SI, TD)) 1216 return false; 1217 1218 InstsToRewrite.insert(SI); 1219 continue; 1220 } 1221 1222 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1223 if (PN->use_empty()) { // Dead PHIs can be stripped. 1224 InstsToRewrite.insert(PN); 1225 continue; 1226 } 1227 1228 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1229 // in the pred blocks, then we can transform this by rewriting the PHI. 1230 if (!isSafePHIToSpeculate(PN, TD)) 1231 return false; 1232 1233 InstsToRewrite.insert(PN); 1234 continue; 1235 } 1236 1237 return false; 1238 } 1239 1240 // If there are no instructions to rewrite, then all uses are load/stores and 1241 // we're done! 1242 if (InstsToRewrite.empty()) 1243 return true; 1244 1245 // If we have instructions that need to be rewritten for this to be promotable 1246 // take care of it now. 1247 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1248 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1249 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1250 // loads with a new select. 1251 while (!SI->use_empty()) { 1252 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1253 1254 IRBuilder<> Builder(LI); 1255 LoadInst *TrueLoad = 1256 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1257 LoadInst *FalseLoad = 1258 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".t"); 1259 1260 // Transfer alignment and TBAA info if present. 1261 TrueLoad->setAlignment(LI->getAlignment()); 1262 FalseLoad->setAlignment(LI->getAlignment()); 1263 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1264 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1265 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1266 } 1267 1268 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1269 V->takeName(LI); 1270 LI->replaceAllUsesWith(V); 1271 LI->eraseFromParent(); 1272 } 1273 1274 // Now that all the loads are gone, the select is gone too. 1275 SI->eraseFromParent(); 1276 continue; 1277 } 1278 1279 // Otherwise, we have a PHI node which allows us to push the loads into the 1280 // predecessors. 1281 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1282 if (PN->use_empty()) { 1283 PN->eraseFromParent(); 1284 continue; 1285 } 1286 1287 const Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1288 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1289 PN->getName()+".ld", PN); 1290 1291 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1292 // matter which one we get and if any differ, it doesn't matter. 1293 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1294 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1295 unsigned Align = SomeLoad->getAlignment(); 1296 1297 // Rewrite all loads of the PN to use the new PHI. 1298 while (!PN->use_empty()) { 1299 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1300 LI->replaceAllUsesWith(NewPN); 1301 LI->eraseFromParent(); 1302 } 1303 1304 // Inject loads into all of the pred blocks. Keep track of which blocks we 1305 // insert them into in case we have multiple edges from the same block. 1306 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1307 1308 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1309 BasicBlock *Pred = PN->getIncomingBlock(i); 1310 LoadInst *&Load = InsertedLoads[Pred]; 1311 if (Load == 0) { 1312 Load = new LoadInst(PN->getIncomingValue(i), 1313 PN->getName() + "." + Pred->getName(), 1314 Pred->getTerminator()); 1315 Load->setAlignment(Align); 1316 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1317 } 1318 1319 NewPN->addIncoming(Load, Pred); 1320 } 1321 1322 PN->eraseFromParent(); 1323 } 1324 1325 ++NumAdjusted; 1326 return true; 1327} 1328 1329 1330bool SROA::performPromotion(Function &F) { 1331 std::vector<AllocaInst*> Allocas; 1332 DominatorTree *DT = 0; 1333 if (HasDomTree) 1334 DT = &getAnalysis<DominatorTree>(); 1335 1336 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1337 1338 bool Changed = false; 1339 SmallVector<Instruction*, 64> Insts; 1340 while (1) { 1341 Allocas.clear(); 1342 1343 // Find allocas that are safe to promote, by looking at all instructions in 1344 // the entry node 1345 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1346 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1347 if (tryToMakeAllocaBePromotable(AI, TD)) 1348 Allocas.push_back(AI); 1349 1350 if (Allocas.empty()) break; 1351 1352 if (HasDomTree) 1353 PromoteMemToReg(Allocas, *DT); 1354 else { 1355 SSAUpdater SSA; 1356 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1357 AllocaInst *AI = Allocas[i]; 1358 1359 // Build list of instructions to promote. 1360 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1361 UI != E; ++UI) 1362 Insts.push_back(cast<Instruction>(*UI)); 1363 1364 AllocaPromoter(Insts, SSA).run(AI, Insts); 1365 Insts.clear(); 1366 } 1367 } 1368 NumPromoted += Allocas.size(); 1369 Changed = true; 1370 } 1371 1372 return Changed; 1373} 1374 1375 1376/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1377/// SROA. It must be a struct or array type with a small number of elements. 1378static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1379 const Type *T = AI->getAllocatedType(); 1380 // Do not promote any struct into more than 32 separate vars. 1381 if (const StructType *ST = dyn_cast<StructType>(T)) 1382 return ST->getNumElements() <= 32; 1383 // Arrays are much less likely to be safe for SROA; only consider 1384 // them if they are very small. 1385 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 1386 return AT->getNumElements() <= 8; 1387 return false; 1388} 1389 1390 1391// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1392// which runs on all of the malloc/alloca instructions in the function, removing 1393// them if they are only used by getelementptr instructions. 1394// 1395bool SROA::performScalarRepl(Function &F) { 1396 std::vector<AllocaInst*> WorkList; 1397 1398 // Scan the entry basic block, adding allocas to the worklist. 1399 BasicBlock &BB = F.getEntryBlock(); 1400 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1401 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1402 WorkList.push_back(A); 1403 1404 // Process the worklist 1405 bool Changed = false; 1406 while (!WorkList.empty()) { 1407 AllocaInst *AI = WorkList.back(); 1408 WorkList.pop_back(); 1409 1410 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1411 // with unused elements. 1412 if (AI->use_empty()) { 1413 AI->eraseFromParent(); 1414 Changed = true; 1415 continue; 1416 } 1417 1418 // If this alloca is impossible for us to promote, reject it early. 1419 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1420 continue; 1421 1422 // Check to see if this allocation is only modified by a memcpy/memmove from 1423 // a constant global. If this is the case, we can change all users to use 1424 // the constant global instead. This is commonly produced by the CFE by 1425 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1426 // is only subsequently read. 1427 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 1428 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1429 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 1430 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 1431 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1432 TheCopy->eraseFromParent(); // Don't mutate the global. 1433 AI->eraseFromParent(); 1434 ++NumGlobals; 1435 Changed = true; 1436 continue; 1437 } 1438 1439 // Check to see if we can perform the core SROA transformation. We cannot 1440 // transform the allocation instruction if it is an array allocation 1441 // (allocations OF arrays are ok though), and an allocation of a scalar 1442 // value cannot be decomposed at all. 1443 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1444 1445 // Do not promote [0 x %struct]. 1446 if (AllocaSize == 0) continue; 1447 1448 // Do not promote any struct whose size is too big. 1449 if (AllocaSize > SRThreshold) continue; 1450 1451 // If the alloca looks like a good candidate for scalar replacement, and if 1452 // all its users can be transformed, then split up the aggregate into its 1453 // separate elements. 1454 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1455 DoScalarReplacement(AI, WorkList); 1456 Changed = true; 1457 continue; 1458 } 1459 1460 // If we can turn this aggregate value (potentially with casts) into a 1461 // simple scalar value that can be mem2reg'd into a register value. 1462 // IsNotTrivial tracks whether this is something that mem2reg could have 1463 // promoted itself. If so, we don't want to transform it needlessly. Note 1464 // that we can't just check based on the type: the alloca may be of an i32 1465 // but that has pointer arithmetic to set byte 3 of it or something. 1466 if (AllocaInst *NewAI = 1467 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1468 NewAI->takeName(AI); 1469 AI->eraseFromParent(); 1470 ++NumConverted; 1471 Changed = true; 1472 continue; 1473 } 1474 1475 // Otherwise, couldn't process this alloca. 1476 } 1477 1478 return Changed; 1479} 1480 1481/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1482/// predicate, do SROA now. 1483void SROA::DoScalarReplacement(AllocaInst *AI, 1484 std::vector<AllocaInst*> &WorkList) { 1485 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1486 SmallVector<AllocaInst*, 32> ElementAllocas; 1487 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1488 ElementAllocas.reserve(ST->getNumContainedTypes()); 1489 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1490 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1491 AI->getAlignment(), 1492 AI->getName() + "." + Twine(i), AI); 1493 ElementAllocas.push_back(NA); 1494 WorkList.push_back(NA); // Add to worklist for recursive processing 1495 } 1496 } else { 1497 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1498 ElementAllocas.reserve(AT->getNumElements()); 1499 const Type *ElTy = AT->getElementType(); 1500 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1501 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1502 AI->getName() + "." + Twine(i), AI); 1503 ElementAllocas.push_back(NA); 1504 WorkList.push_back(NA); // Add to worklist for recursive processing 1505 } 1506 } 1507 1508 // Now that we have created the new alloca instructions, rewrite all the 1509 // uses of the old alloca. 1510 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1511 1512 // Now erase any instructions that were made dead while rewriting the alloca. 1513 DeleteDeadInstructions(); 1514 AI->eraseFromParent(); 1515 1516 ++NumReplaced; 1517} 1518 1519/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1520/// recursively including all their operands that become trivially dead. 1521void SROA::DeleteDeadInstructions() { 1522 while (!DeadInsts.empty()) { 1523 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1524 1525 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1526 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1527 // Zero out the operand and see if it becomes trivially dead. 1528 // (But, don't add allocas to the dead instruction list -- they are 1529 // already on the worklist and will be deleted separately.) 1530 *OI = 0; 1531 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1532 DeadInsts.push_back(U); 1533 } 1534 1535 I->eraseFromParent(); 1536 } 1537} 1538 1539/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1540/// performing scalar replacement of alloca AI. The results are flagged in 1541/// the Info parameter. Offset indicates the position within AI that is 1542/// referenced by this instruction. 1543void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1544 AllocaInfo &Info) { 1545 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1546 Instruction *User = cast<Instruction>(*UI); 1547 1548 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1549 isSafeForScalarRepl(BC, Offset, Info); 1550 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1551 uint64_t GEPOffset = Offset; 1552 isSafeGEP(GEPI, GEPOffset, Info); 1553 if (!Info.isUnsafe) 1554 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1555 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1556 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1557 if (Length == 0) 1558 return MarkUnsafe(Info, User); 1559 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1560 UI.getOperandNo() == 0, Info, MI, 1561 true /*AllowWholeAccess*/); 1562 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1563 if (LI->isVolatile()) 1564 return MarkUnsafe(Info, User); 1565 const Type *LIType = LI->getType(); 1566 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1567 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1568 Info.hasALoadOrStore = true; 1569 1570 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1571 // Store is ok if storing INTO the pointer, not storing the pointer 1572 if (SI->isVolatile() || SI->getOperand(0) == I) 1573 return MarkUnsafe(Info, User); 1574 1575 const Type *SIType = SI->getOperand(0)->getType(); 1576 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1577 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1578 Info.hasALoadOrStore = true; 1579 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1580 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1581 } else { 1582 return MarkUnsafe(Info, User); 1583 } 1584 if (Info.isUnsafe) return; 1585 } 1586} 1587 1588 1589/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1590/// derived from the alloca, we can often still split the alloca into elements. 1591/// This is useful if we have a large alloca where one element is phi'd 1592/// together somewhere: we can SRoA and promote all the other elements even if 1593/// we end up not being able to promote this one. 1594/// 1595/// All we require is that the uses of the PHI do not index into other parts of 1596/// the alloca. The most important use case for this is single load and stores 1597/// that are PHI'd together, which can happen due to code sinking. 1598void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1599 AllocaInfo &Info) { 1600 // If we've already checked this PHI, don't do it again. 1601 if (PHINode *PN = dyn_cast<PHINode>(I)) 1602 if (!Info.CheckedPHIs.insert(PN)) 1603 return; 1604 1605 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1606 Instruction *User = cast<Instruction>(*UI); 1607 1608 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1609 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1610 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1611 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1612 // but would have to prove that we're staying inside of an element being 1613 // promoted. 1614 if (!GEPI->hasAllZeroIndices()) 1615 return MarkUnsafe(Info, User); 1616 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1617 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1618 if (LI->isVolatile()) 1619 return MarkUnsafe(Info, User); 1620 const Type *LIType = LI->getType(); 1621 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1622 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1623 Info.hasALoadOrStore = true; 1624 1625 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1626 // Store is ok if storing INTO the pointer, not storing the pointer 1627 if (SI->isVolatile() || SI->getOperand(0) == I) 1628 return MarkUnsafe(Info, User); 1629 1630 const Type *SIType = SI->getOperand(0)->getType(); 1631 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1632 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1633 Info.hasALoadOrStore = true; 1634 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1635 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1636 } else { 1637 return MarkUnsafe(Info, User); 1638 } 1639 if (Info.isUnsafe) return; 1640 } 1641} 1642 1643/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1644/// replacement. It is safe when all the indices are constant, in-bounds 1645/// references, and when the resulting offset corresponds to an element within 1646/// the alloca type. The results are flagged in the Info parameter. Upon 1647/// return, Offset is adjusted as specified by the GEP indices. 1648void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1649 uint64_t &Offset, AllocaInfo &Info) { 1650 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1651 if (GEPIt == E) 1652 return; 1653 1654 // Walk through the GEP type indices, checking the types that this indexes 1655 // into. 1656 for (; GEPIt != E; ++GEPIt) { 1657 // Ignore struct elements, no extra checking needed for these. 1658 if ((*GEPIt)->isStructTy()) 1659 continue; 1660 1661 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1662 if (!IdxVal) 1663 return MarkUnsafe(Info, GEPI); 1664 } 1665 1666 // Compute the offset due to this GEP and check if the alloca has a 1667 // component element at that offset. 1668 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1669 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1670 &Indices[0], Indices.size()); 1671 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1672 MarkUnsafe(Info, GEPI); 1673} 1674 1675/// isHomogeneousAggregate - Check if type T is a struct or array containing 1676/// elements of the same type (which is always true for arrays). If so, 1677/// return true with NumElts and EltTy set to the number of elements and the 1678/// element type, respectively. 1679static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1680 const Type *&EltTy) { 1681 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1682 NumElts = AT->getNumElements(); 1683 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1684 return true; 1685 } 1686 if (const StructType *ST = dyn_cast<StructType>(T)) { 1687 NumElts = ST->getNumContainedTypes(); 1688 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1689 for (unsigned n = 1; n < NumElts; ++n) { 1690 if (ST->getContainedType(n) != EltTy) 1691 return false; 1692 } 1693 return true; 1694 } 1695 return false; 1696} 1697 1698/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1699/// "homogeneous" aggregates with the same element type and number of elements. 1700static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1701 if (T1 == T2) 1702 return true; 1703 1704 unsigned NumElts1, NumElts2; 1705 const Type *EltTy1, *EltTy2; 1706 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1707 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1708 NumElts1 == NumElts2 && 1709 EltTy1 == EltTy2) 1710 return true; 1711 1712 return false; 1713} 1714 1715/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1716/// alloca or has an offset and size that corresponds to a component element 1717/// within it. The offset checked here may have been formed from a GEP with a 1718/// pointer bitcasted to a different type. 1719/// 1720/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1721/// unit. If false, it only allows accesses known to be in a single element. 1722void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1723 const Type *MemOpType, bool isStore, 1724 AllocaInfo &Info, Instruction *TheAccess, 1725 bool AllowWholeAccess) { 1726 // Check if this is a load/store of the entire alloca. 1727 if (Offset == 0 && AllowWholeAccess && 1728 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1729 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1730 // loads/stores (which are essentially the same as the MemIntrinsics with 1731 // regard to copying padding between elements). But, if an alloca is 1732 // flagged as both a source and destination of such operations, we'll need 1733 // to check later for padding between elements. 1734 if (!MemOpType || MemOpType->isIntegerTy()) { 1735 if (isStore) 1736 Info.isMemCpyDst = true; 1737 else 1738 Info.isMemCpySrc = true; 1739 return; 1740 } 1741 // This is also safe for references using a type that is compatible with 1742 // the type of the alloca, so that loads/stores can be rewritten using 1743 // insertvalue/extractvalue. 1744 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1745 Info.hasSubelementAccess = true; 1746 return; 1747 } 1748 } 1749 // Check if the offset/size correspond to a component within the alloca type. 1750 const Type *T = Info.AI->getAllocatedType(); 1751 if (TypeHasComponent(T, Offset, MemSize)) { 1752 Info.hasSubelementAccess = true; 1753 return; 1754 } 1755 1756 return MarkUnsafe(Info, TheAccess); 1757} 1758 1759/// TypeHasComponent - Return true if T has a component type with the 1760/// specified offset and size. If Size is zero, do not check the size. 1761bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1762 const Type *EltTy; 1763 uint64_t EltSize; 1764 if (const StructType *ST = dyn_cast<StructType>(T)) { 1765 const StructLayout *Layout = TD->getStructLayout(ST); 1766 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1767 EltTy = ST->getContainedType(EltIdx); 1768 EltSize = TD->getTypeAllocSize(EltTy); 1769 Offset -= Layout->getElementOffset(EltIdx); 1770 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1771 EltTy = AT->getElementType(); 1772 EltSize = TD->getTypeAllocSize(EltTy); 1773 if (Offset >= AT->getNumElements() * EltSize) 1774 return false; 1775 Offset %= EltSize; 1776 } else { 1777 return false; 1778 } 1779 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1780 return true; 1781 // Check if the component spans multiple elements. 1782 if (Offset + Size > EltSize) 1783 return false; 1784 return TypeHasComponent(EltTy, Offset, Size); 1785} 1786 1787/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1788/// the instruction I, which references it, to use the separate elements. 1789/// Offset indicates the position within AI that is referenced by this 1790/// instruction. 1791void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1792 SmallVector<AllocaInst*, 32> &NewElts) { 1793 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1794 Use &TheUse = UI.getUse(); 1795 Instruction *User = cast<Instruction>(*UI++); 1796 1797 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1798 RewriteBitCast(BC, AI, Offset, NewElts); 1799 continue; 1800 } 1801 1802 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1803 RewriteGEP(GEPI, AI, Offset, NewElts); 1804 continue; 1805 } 1806 1807 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1808 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1809 uint64_t MemSize = Length->getZExtValue(); 1810 if (Offset == 0 && 1811 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1812 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1813 // Otherwise the intrinsic can only touch a single element and the 1814 // address operand will be updated, so nothing else needs to be done. 1815 continue; 1816 } 1817 1818 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1819 const Type *LIType = LI->getType(); 1820 1821 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1822 // Replace: 1823 // %res = load { i32, i32 }* %alloc 1824 // with: 1825 // %load.0 = load i32* %alloc.0 1826 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1827 // %load.1 = load i32* %alloc.1 1828 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1829 // (Also works for arrays instead of structs) 1830 Value *Insert = UndefValue::get(LIType); 1831 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1832 Value *Load = new LoadInst(NewElts[i], "load", LI); 1833 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1834 } 1835 LI->replaceAllUsesWith(Insert); 1836 DeadInsts.push_back(LI); 1837 } else if (LIType->isIntegerTy() && 1838 TD->getTypeAllocSize(LIType) == 1839 TD->getTypeAllocSize(AI->getAllocatedType())) { 1840 // If this is a load of the entire alloca to an integer, rewrite it. 1841 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1842 } 1843 continue; 1844 } 1845 1846 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1847 Value *Val = SI->getOperand(0); 1848 const Type *SIType = Val->getType(); 1849 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1850 // Replace: 1851 // store { i32, i32 } %val, { i32, i32 }* %alloc 1852 // with: 1853 // %val.0 = extractvalue { i32, i32 } %val, 0 1854 // store i32 %val.0, i32* %alloc.0 1855 // %val.1 = extractvalue { i32, i32 } %val, 1 1856 // store i32 %val.1, i32* %alloc.1 1857 // (Also works for arrays instead of structs) 1858 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1859 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1860 new StoreInst(Extract, NewElts[i], SI); 1861 } 1862 DeadInsts.push_back(SI); 1863 } else if (SIType->isIntegerTy() && 1864 TD->getTypeAllocSize(SIType) == 1865 TD->getTypeAllocSize(AI->getAllocatedType())) { 1866 // If this is a store of the entire alloca from an integer, rewrite it. 1867 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1868 } 1869 continue; 1870 } 1871 1872 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1873 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1874 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1875 // the new pointer. 1876 if (!isa<AllocaInst>(I)) continue; 1877 1878 assert(Offset == 0 && NewElts[0] && 1879 "Direct alloca use should have a zero offset"); 1880 1881 // If we have a use of the alloca, we know the derived uses will be 1882 // utilizing just the first element of the scalarized result. Insert a 1883 // bitcast of the first alloca before the user as required. 1884 AllocaInst *NewAI = NewElts[0]; 1885 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1886 NewAI->moveBefore(BCI); 1887 TheUse = BCI; 1888 continue; 1889 } 1890 } 1891} 1892 1893/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1894/// and recursively continue updating all of its uses. 1895void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1896 SmallVector<AllocaInst*, 32> &NewElts) { 1897 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1898 if (BC->getOperand(0) != AI) 1899 return; 1900 1901 // The bitcast references the original alloca. Replace its uses with 1902 // references to the first new element alloca. 1903 Instruction *Val = NewElts[0]; 1904 if (Val->getType() != BC->getDestTy()) { 1905 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1906 Val->takeName(BC); 1907 } 1908 BC->replaceAllUsesWith(Val); 1909 DeadInsts.push_back(BC); 1910} 1911 1912/// FindElementAndOffset - Return the index of the element containing Offset 1913/// within the specified type, which must be either a struct or an array. 1914/// Sets T to the type of the element and Offset to the offset within that 1915/// element. IdxTy is set to the type of the index result to be used in a 1916/// GEP instruction. 1917uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1918 const Type *&IdxTy) { 1919 uint64_t Idx = 0; 1920 if (const StructType *ST = dyn_cast<StructType>(T)) { 1921 const StructLayout *Layout = TD->getStructLayout(ST); 1922 Idx = Layout->getElementContainingOffset(Offset); 1923 T = ST->getContainedType(Idx); 1924 Offset -= Layout->getElementOffset(Idx); 1925 IdxTy = Type::getInt32Ty(T->getContext()); 1926 return Idx; 1927 } 1928 const ArrayType *AT = cast<ArrayType>(T); 1929 T = AT->getElementType(); 1930 uint64_t EltSize = TD->getTypeAllocSize(T); 1931 Idx = Offset / EltSize; 1932 Offset -= Idx * EltSize; 1933 IdxTy = Type::getInt64Ty(T->getContext()); 1934 return Idx; 1935} 1936 1937/// RewriteGEP - Check if this GEP instruction moves the pointer across 1938/// elements of the alloca that are being split apart, and if so, rewrite 1939/// the GEP to be relative to the new element. 1940void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1941 SmallVector<AllocaInst*, 32> &NewElts) { 1942 uint64_t OldOffset = Offset; 1943 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1944 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1945 &Indices[0], Indices.size()); 1946 1947 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1948 1949 const Type *T = AI->getAllocatedType(); 1950 const Type *IdxTy; 1951 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1952 if (GEPI->getOperand(0) == AI) 1953 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1954 1955 T = AI->getAllocatedType(); 1956 uint64_t EltOffset = Offset; 1957 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1958 1959 // If this GEP does not move the pointer across elements of the alloca 1960 // being split, then it does not needs to be rewritten. 1961 if (Idx == OldIdx) 1962 return; 1963 1964 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1965 SmallVector<Value*, 8> NewArgs; 1966 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1967 while (EltOffset != 0) { 1968 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1969 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1970 } 1971 Instruction *Val = NewElts[Idx]; 1972 if (NewArgs.size() > 1) { 1973 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1974 NewArgs.end(), "", GEPI); 1975 Val->takeName(GEPI); 1976 } 1977 if (Val->getType() != GEPI->getType()) 1978 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1979 GEPI->replaceAllUsesWith(Val); 1980 DeadInsts.push_back(GEPI); 1981} 1982 1983/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1984/// Rewrite it to copy or set the elements of the scalarized memory. 1985void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1986 AllocaInst *AI, 1987 SmallVector<AllocaInst*, 32> &NewElts) { 1988 // If this is a memcpy/memmove, construct the other pointer as the 1989 // appropriate type. The "Other" pointer is the pointer that goes to memory 1990 // that doesn't have anything to do with the alloca that we are promoting. For 1991 // memset, this Value* stays null. 1992 Value *OtherPtr = 0; 1993 unsigned MemAlignment = MI->getAlignment(); 1994 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1995 if (Inst == MTI->getRawDest()) 1996 OtherPtr = MTI->getRawSource(); 1997 else { 1998 assert(Inst == MTI->getRawSource()); 1999 OtherPtr = MTI->getRawDest(); 2000 } 2001 } 2002 2003 // If there is an other pointer, we want to convert it to the same pointer 2004 // type as AI has, so we can GEP through it safely. 2005 if (OtherPtr) { 2006 unsigned AddrSpace = 2007 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2008 2009 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2010 // optimization, but it's also required to detect the corner case where 2011 // both pointer operands are referencing the same memory, and where 2012 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2013 // function is only called for mem intrinsics that access the whole 2014 // aggregate, so non-zero GEPs are not an issue here.) 2015 OtherPtr = OtherPtr->stripPointerCasts(); 2016 2017 // Copying the alloca to itself is a no-op: just delete it. 2018 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2019 // This code will run twice for a no-op memcpy -- once for each operand. 2020 // Put only one reference to MI on the DeadInsts list. 2021 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2022 E = DeadInsts.end(); I != E; ++I) 2023 if (*I == MI) return; 2024 DeadInsts.push_back(MI); 2025 return; 2026 } 2027 2028 // If the pointer is not the right type, insert a bitcast to the right 2029 // type. 2030 const Type *NewTy = 2031 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2032 2033 if (OtherPtr->getType() != NewTy) 2034 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2035 } 2036 2037 // Process each element of the aggregate. 2038 bool SROADest = MI->getRawDest() == Inst; 2039 2040 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2041 2042 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2043 // If this is a memcpy/memmove, emit a GEP of the other element address. 2044 Value *OtherElt = 0; 2045 unsigned OtherEltAlign = MemAlignment; 2046 2047 if (OtherPtr) { 2048 Value *Idx[2] = { Zero, 2049 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2050 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 2051 OtherPtr->getName()+"."+Twine(i), 2052 MI); 2053 uint64_t EltOffset; 2054 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2055 const Type *OtherTy = OtherPtrTy->getElementType(); 2056 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 2057 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2058 } else { 2059 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2060 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2061 } 2062 2063 // The alignment of the other pointer is the guaranteed alignment of the 2064 // element, which is affected by both the known alignment of the whole 2065 // mem intrinsic and the alignment of the element. If the alignment of 2066 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2067 // known alignment is just 4 bytes. 2068 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2069 } 2070 2071 Value *EltPtr = NewElts[i]; 2072 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2073 2074 // If we got down to a scalar, insert a load or store as appropriate. 2075 if (EltTy->isSingleValueType()) { 2076 if (isa<MemTransferInst>(MI)) { 2077 if (SROADest) { 2078 // From Other to Alloca. 2079 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2080 new StoreInst(Elt, EltPtr, MI); 2081 } else { 2082 // From Alloca to Other. 2083 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2084 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2085 } 2086 continue; 2087 } 2088 assert(isa<MemSetInst>(MI)); 2089 2090 // If the stored element is zero (common case), just store a null 2091 // constant. 2092 Constant *StoreVal; 2093 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2094 if (CI->isZero()) { 2095 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2096 } else { 2097 // If EltTy is a vector type, get the element type. 2098 const Type *ValTy = EltTy->getScalarType(); 2099 2100 // Construct an integer with the right value. 2101 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2102 APInt OneVal(EltSize, CI->getZExtValue()); 2103 APInt TotalVal(OneVal); 2104 // Set each byte. 2105 for (unsigned i = 0; 8*i < EltSize; ++i) { 2106 TotalVal = TotalVal.shl(8); 2107 TotalVal |= OneVal; 2108 } 2109 2110 // Convert the integer value to the appropriate type. 2111 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2112 if (ValTy->isPointerTy()) 2113 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2114 else if (ValTy->isFloatingPointTy()) 2115 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2116 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2117 2118 // If the requested value was a vector constant, create it. 2119 if (EltTy != ValTy) { 2120 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 2121 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2122 StoreVal = ConstantVector::get(Elts); 2123 } 2124 } 2125 new StoreInst(StoreVal, EltPtr, MI); 2126 continue; 2127 } 2128 // Otherwise, if we're storing a byte variable, use a memset call for 2129 // this element. 2130 } 2131 2132 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2133 2134 IRBuilder<> Builder(MI); 2135 2136 // Finally, insert the meminst for this element. 2137 if (isa<MemSetInst>(MI)) { 2138 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2139 MI->isVolatile()); 2140 } else { 2141 assert(isa<MemTransferInst>(MI)); 2142 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2143 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2144 2145 if (isa<MemCpyInst>(MI)) 2146 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2147 else 2148 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2149 } 2150 } 2151 DeadInsts.push_back(MI); 2152} 2153 2154/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2155/// overwrites the entire allocation. Extract out the pieces of the stored 2156/// integer and store them individually. 2157void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2158 SmallVector<AllocaInst*, 32> &NewElts){ 2159 // Extract each element out of the integer according to its structure offset 2160 // and store the element value to the individual alloca. 2161 Value *SrcVal = SI->getOperand(0); 2162 const Type *AllocaEltTy = AI->getAllocatedType(); 2163 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2164 2165 IRBuilder<> Builder(SI); 2166 2167 // Handle tail padding by extending the operand 2168 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2169 SrcVal = Builder.CreateZExt(SrcVal, 2170 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2171 2172 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2173 << '\n'); 2174 2175 // There are two forms here: AI could be an array or struct. Both cases 2176 // have different ways to compute the element offset. 2177 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2178 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2179 2180 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2181 // Get the number of bits to shift SrcVal to get the value. 2182 const Type *FieldTy = EltSTy->getElementType(i); 2183 uint64_t Shift = Layout->getElementOffsetInBits(i); 2184 2185 if (TD->isBigEndian()) 2186 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2187 2188 Value *EltVal = SrcVal; 2189 if (Shift) { 2190 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2191 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2192 } 2193 2194 // Truncate down to an integer of the right size. 2195 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2196 2197 // Ignore zero sized fields like {}, they obviously contain no data. 2198 if (FieldSizeBits == 0) continue; 2199 2200 if (FieldSizeBits != AllocaSizeBits) 2201 EltVal = Builder.CreateTrunc(EltVal, 2202 IntegerType::get(SI->getContext(), FieldSizeBits)); 2203 Value *DestField = NewElts[i]; 2204 if (EltVal->getType() == FieldTy) { 2205 // Storing to an integer field of this size, just do it. 2206 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2207 // Bitcast to the right element type (for fp/vector values). 2208 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2209 } else { 2210 // Otherwise, bitcast the dest pointer (for aggregates). 2211 DestField = Builder.CreateBitCast(DestField, 2212 PointerType::getUnqual(EltVal->getType())); 2213 } 2214 new StoreInst(EltVal, DestField, SI); 2215 } 2216 2217 } else { 2218 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2219 const Type *ArrayEltTy = ATy->getElementType(); 2220 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2221 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2222 2223 uint64_t Shift; 2224 2225 if (TD->isBigEndian()) 2226 Shift = AllocaSizeBits-ElementOffset; 2227 else 2228 Shift = 0; 2229 2230 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2231 // Ignore zero sized fields like {}, they obviously contain no data. 2232 if (ElementSizeBits == 0) continue; 2233 2234 Value *EltVal = SrcVal; 2235 if (Shift) { 2236 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2237 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2238 } 2239 2240 // Truncate down to an integer of the right size. 2241 if (ElementSizeBits != AllocaSizeBits) 2242 EltVal = Builder.CreateTrunc(EltVal, 2243 IntegerType::get(SI->getContext(), 2244 ElementSizeBits)); 2245 Value *DestField = NewElts[i]; 2246 if (EltVal->getType() == ArrayEltTy) { 2247 // Storing to an integer field of this size, just do it. 2248 } else if (ArrayEltTy->isFloatingPointTy() || 2249 ArrayEltTy->isVectorTy()) { 2250 // Bitcast to the right element type (for fp/vector values). 2251 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2252 } else { 2253 // Otherwise, bitcast the dest pointer (for aggregates). 2254 DestField = Builder.CreateBitCast(DestField, 2255 PointerType::getUnqual(EltVal->getType())); 2256 } 2257 new StoreInst(EltVal, DestField, SI); 2258 2259 if (TD->isBigEndian()) 2260 Shift -= ElementOffset; 2261 else 2262 Shift += ElementOffset; 2263 } 2264 } 2265 2266 DeadInsts.push_back(SI); 2267} 2268 2269/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2270/// an integer. Load the individual pieces to form the aggregate value. 2271void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2272 SmallVector<AllocaInst*, 32> &NewElts) { 2273 // Extract each element out of the NewElts according to its structure offset 2274 // and form the result value. 2275 const Type *AllocaEltTy = AI->getAllocatedType(); 2276 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2277 2278 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2279 << '\n'); 2280 2281 // There are two forms here: AI could be an array or struct. Both cases 2282 // have different ways to compute the element offset. 2283 const StructLayout *Layout = 0; 2284 uint64_t ArrayEltBitOffset = 0; 2285 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2286 Layout = TD->getStructLayout(EltSTy); 2287 } else { 2288 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2289 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2290 } 2291 2292 Value *ResultVal = 2293 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2294 2295 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2296 // Load the value from the alloca. If the NewElt is an aggregate, cast 2297 // the pointer to an integer of the same size before doing the load. 2298 Value *SrcField = NewElts[i]; 2299 const Type *FieldTy = 2300 cast<PointerType>(SrcField->getType())->getElementType(); 2301 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2302 2303 // Ignore zero sized fields like {}, they obviously contain no data. 2304 if (FieldSizeBits == 0) continue; 2305 2306 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2307 FieldSizeBits); 2308 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2309 !FieldTy->isVectorTy()) 2310 SrcField = new BitCastInst(SrcField, 2311 PointerType::getUnqual(FieldIntTy), 2312 "", LI); 2313 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2314 2315 // If SrcField is a fp or vector of the right size but that isn't an 2316 // integer type, bitcast to an integer so we can shift it. 2317 if (SrcField->getType() != FieldIntTy) 2318 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2319 2320 // Zero extend the field to be the same size as the final alloca so that 2321 // we can shift and insert it. 2322 if (SrcField->getType() != ResultVal->getType()) 2323 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2324 2325 // Determine the number of bits to shift SrcField. 2326 uint64_t Shift; 2327 if (Layout) // Struct case. 2328 Shift = Layout->getElementOffsetInBits(i); 2329 else // Array case. 2330 Shift = i*ArrayEltBitOffset; 2331 2332 if (TD->isBigEndian()) 2333 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2334 2335 if (Shift) { 2336 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2337 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2338 } 2339 2340 // Don't create an 'or x, 0' on the first iteration. 2341 if (!isa<Constant>(ResultVal) || 2342 !cast<Constant>(ResultVal)->isNullValue()) 2343 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2344 else 2345 ResultVal = SrcField; 2346 } 2347 2348 // Handle tail padding by truncating the result 2349 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2350 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2351 2352 LI->replaceAllUsesWith(ResultVal); 2353 DeadInsts.push_back(LI); 2354} 2355 2356/// HasPadding - Return true if the specified type has any structure or 2357/// alignment padding in between the elements that would be split apart 2358/// by SROA; return false otherwise. 2359static bool HasPadding(const Type *Ty, const TargetData &TD) { 2360 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2361 Ty = ATy->getElementType(); 2362 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2363 } 2364 2365 // SROA currently handles only Arrays and Structs. 2366 const StructType *STy = cast<StructType>(Ty); 2367 const StructLayout *SL = TD.getStructLayout(STy); 2368 unsigned PrevFieldBitOffset = 0; 2369 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2370 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2371 2372 // Check to see if there is any padding between this element and the 2373 // previous one. 2374 if (i) { 2375 unsigned PrevFieldEnd = 2376 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2377 if (PrevFieldEnd < FieldBitOffset) 2378 return true; 2379 } 2380 PrevFieldBitOffset = FieldBitOffset; 2381 } 2382 // Check for tail padding. 2383 if (unsigned EltCount = STy->getNumElements()) { 2384 unsigned PrevFieldEnd = PrevFieldBitOffset + 2385 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2386 if (PrevFieldEnd < SL->getSizeInBits()) 2387 return true; 2388 } 2389 return false; 2390} 2391 2392/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2393/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2394/// or 1 if safe after canonicalization has been performed. 2395bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2396 // Loop over the use list of the alloca. We can only transform it if all of 2397 // the users are safe to transform. 2398 AllocaInfo Info(AI); 2399 2400 isSafeForScalarRepl(AI, 0, Info); 2401 if (Info.isUnsafe) { 2402 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2403 return false; 2404 } 2405 2406 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2407 // source and destination, we have to be careful. In particular, the memcpy 2408 // could be moving around elements that live in structure padding of the LLVM 2409 // types, but may actually be used. In these cases, we refuse to promote the 2410 // struct. 2411 if (Info.isMemCpySrc && Info.isMemCpyDst && 2412 HasPadding(AI->getAllocatedType(), *TD)) 2413 return false; 2414 2415 // If the alloca never has an access to just *part* of it, but is accessed 2416 // via loads and stores, then we should use ConvertToScalarInfo to promote 2417 // the alloca instead of promoting each piece at a time and inserting fission 2418 // and fusion code. 2419 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2420 // If the struct/array just has one element, use basic SRoA. 2421 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2422 if (ST->getNumElements() > 1) return false; 2423 } else { 2424 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2425 return false; 2426 } 2427 } 2428 2429 return true; 2430} 2431 2432 2433 2434/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2435/// some part of a constant global variable. This intentionally only accepts 2436/// constant expressions because we don't can't rewrite arbitrary instructions. 2437static bool PointsToConstantGlobal(Value *V) { 2438 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2439 return GV->isConstant(); 2440 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2441 if (CE->getOpcode() == Instruction::BitCast || 2442 CE->getOpcode() == Instruction::GetElementPtr) 2443 return PointsToConstantGlobal(CE->getOperand(0)); 2444 return false; 2445} 2446 2447/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2448/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2449/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2450/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2451/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2452/// the alloca, and if the source pointer is a pointer to a constant global, we 2453/// can optimize this. 2454static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2455 bool isOffset) { 2456 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2457 User *U = cast<Instruction>(*UI); 2458 2459 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2460 // Ignore non-volatile loads, they are always ok. 2461 if (LI->isVolatile()) return false; 2462 continue; 2463 } 2464 2465 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2466 // If uses of the bitcast are ok, we are ok. 2467 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 2468 return false; 2469 continue; 2470 } 2471 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2472 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2473 // doesn't, it does. 2474 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2475 isOffset || !GEP->hasAllZeroIndices())) 2476 return false; 2477 continue; 2478 } 2479 2480 if (CallSite CS = U) { 2481 // If this is a readonly/readnone call site, then we know it is just a 2482 // load and we can ignore it. 2483 if (CS.onlyReadsMemory()) 2484 continue; 2485 2486 // If this is the function being called then we treat it like a load and 2487 // ignore it. 2488 if (CS.isCallee(UI)) 2489 continue; 2490 2491 // If this is being passed as a byval argument, the caller is making a 2492 // copy, so it is only a read of the alloca. 2493 unsigned ArgNo = CS.getArgumentNo(UI); 2494 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2495 continue; 2496 } 2497 2498 // If this is isn't our memcpy/memmove, reject it as something we can't 2499 // handle. 2500 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2501 if (MI == 0) 2502 return false; 2503 2504 // If the transfer is using the alloca as a source of the transfer, then 2505 // ignore it since it is a load (unless the transfer is volatile). 2506 if (UI.getOperandNo() == 1) { 2507 if (MI->isVolatile()) return false; 2508 continue; 2509 } 2510 2511 // If we already have seen a copy, reject the second one. 2512 if (TheCopy) return false; 2513 2514 // If the pointer has been offset from the start of the alloca, we can't 2515 // safely handle this. 2516 if (isOffset) return false; 2517 2518 // If the memintrinsic isn't using the alloca as the dest, reject it. 2519 if (UI.getOperandNo() != 0) return false; 2520 2521 // If the source of the memcpy/move is not a constant global, reject it. 2522 if (!PointsToConstantGlobal(MI->getSource())) 2523 return false; 2524 2525 // Otherwise, the transform is safe. Remember the copy instruction. 2526 TheCopy = MI; 2527 } 2528 return true; 2529} 2530 2531/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2532/// modified by a copy from a constant global. If we can prove this, we can 2533/// replace any uses of the alloca with uses of the global directly. 2534MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 2535 MemTransferInst *TheCopy = 0; 2536 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 2537 return TheCopy; 2538 return 0; 2539} 2540