ScalarReplAggregates.cpp revision 3ecfc861b4365f341c5c969b40e1afccde676e6f
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 if (FromVal->getType() == ToType && Offset == 0) 694 return FromVal; 695 696 // If the result alloca is a vector type, this is either an element 697 // access or a bitcast to another vector type of the same size. 698 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 699 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 700 if (ToTypeSize == AllocaSize) 701 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 702 703 if (ToType->isVectorTy()) { 704 assert(isPowerOf2_64(AllocaSize / ToTypeSize) && 705 "Partial vector access of an alloca must have a power-of-2 size " 706 "ratio."); 707 assert(Offset == 0 && "Can't extract a value of a smaller vector type " 708 "from a nonzero offset."); 709 710 const Type *ToElementTy = cast<VectorType>(ToType)->getElementType(); 711 const Type *CastElementTy = getScaledElementType(ToElementTy, 712 ToTypeSize * 8); 713 unsigned NumCastVectorElements = AllocaSize / ToTypeSize; 714 715 LLVMContext &Context = FromVal->getContext(); 716 const Type *CastTy = VectorType::get(CastElementTy, 717 NumCastVectorElements); 718 Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp"); 719 Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get( 720 Type::getInt32Ty(Context), 0), "tmp"); 721 return Builder.CreateBitCast(Extract, ToType, "tmp"); 722 } 723 724 // Otherwise it must be an element access. 725 unsigned Elt = 0; 726 if (Offset) { 727 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 728 Elt = Offset/EltSize; 729 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 730 } 731 // Return the element extracted out of it. 732 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 733 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 734 if (V->getType() != ToType) 735 V = Builder.CreateBitCast(V, ToType, "tmp"); 736 return V; 737 } 738 739 // If ToType is a first class aggregate, extract out each of the pieces and 740 // use insertvalue's to form the FCA. 741 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 742 const StructLayout &Layout = *TD.getStructLayout(ST); 743 Value *Res = UndefValue::get(ST); 744 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 745 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 746 Offset+Layout.getElementOffsetInBits(i), 747 Builder); 748 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 749 } 750 return Res; 751 } 752 753 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 754 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 755 Value *Res = UndefValue::get(AT); 756 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 757 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 758 Offset+i*EltSize, Builder); 759 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 760 } 761 return Res; 762 } 763 764 // Otherwise, this must be a union that was converted to an integer value. 765 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 766 767 // If this is a big-endian system and the load is narrower than the 768 // full alloca type, we need to do a shift to get the right bits. 769 int ShAmt = 0; 770 if (TD.isBigEndian()) { 771 // On big-endian machines, the lowest bit is stored at the bit offset 772 // from the pointer given by getTypeStoreSizeInBits. This matters for 773 // integers with a bitwidth that is not a multiple of 8. 774 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 775 TD.getTypeStoreSizeInBits(ToType) - Offset; 776 } else { 777 ShAmt = Offset; 778 } 779 780 // Note: we support negative bitwidths (with shl) which are not defined. 781 // We do this to support (f.e.) loads off the end of a structure where 782 // only some bits are used. 783 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 784 FromVal = Builder.CreateLShr(FromVal, 785 ConstantInt::get(FromVal->getType(), 786 ShAmt), "tmp"); 787 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 788 FromVal = Builder.CreateShl(FromVal, 789 ConstantInt::get(FromVal->getType(), 790 -ShAmt), "tmp"); 791 792 // Finally, unconditionally truncate the integer to the right width. 793 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 794 if (LIBitWidth < NTy->getBitWidth()) 795 FromVal = 796 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 797 LIBitWidth), "tmp"); 798 else if (LIBitWidth > NTy->getBitWidth()) 799 FromVal = 800 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 801 LIBitWidth), "tmp"); 802 803 // If the result is an integer, this is a trunc or bitcast. 804 if (ToType->isIntegerTy()) { 805 // Should be done. 806 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 807 // Just do a bitcast, we know the sizes match up. 808 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 809 } else { 810 // Otherwise must be a pointer. 811 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 812 } 813 assert(FromVal->getType() == ToType && "Didn't convert right?"); 814 return FromVal; 815} 816 817/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 818/// or vector value "Old" at the offset specified by Offset. 819/// 820/// This happens when we are converting an "integer union" to a 821/// single integer scalar, or when we are converting a "vector union" to a 822/// vector with insert/extractelement instructions. 823/// 824/// Offset is an offset from the original alloca, in bits that need to be 825/// shifted to the right. 826Value *ConvertToScalarInfo:: 827ConvertScalar_InsertValue(Value *SV, Value *Old, 828 uint64_t Offset, IRBuilder<> &Builder) { 829 // Convert the stored type to the actual type, shift it left to insert 830 // then 'or' into place. 831 const Type *AllocaType = Old->getType(); 832 LLVMContext &Context = Old->getContext(); 833 834 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 835 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 836 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 837 838 // Changing the whole vector with memset or with an access of a different 839 // vector type? 840 if (ValSize == VecSize) 841 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 842 843 if (SV->getType()->isVectorTy() && isPowerOf2_64(VecSize / ValSize)) { 844 assert(Offset == 0 && "Can't insert a value of a smaller vector type at " 845 "a nonzero offset."); 846 847 const Type *ToElementTy = 848 cast<VectorType>(SV->getType())->getElementType(); 849 const Type *CastElementTy = getScaledElementType(ToElementTy, ValSize); 850 unsigned NumCastVectorElements = VecSize / ValSize; 851 852 LLVMContext &Context = SV->getContext(); 853 const Type *OldCastTy = VectorType::get(CastElementTy, 854 NumCastVectorElements); 855 Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp"); 856 857 Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp"); 858 Value *Insert = 859 Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get( 860 Type::getInt32Ty(Context), 0), "tmp"); 861 return Builder.CreateBitCast(Insert, AllocaType, "tmp"); 862 } 863 864 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 865 866 // Must be an element insertion. 867 unsigned Elt = Offset/EltSize; 868 869 if (SV->getType() != VTy->getElementType()) 870 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 871 872 SV = Builder.CreateInsertElement(Old, SV, 873 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 874 "tmp"); 875 return SV; 876 } 877 878 // If SV is a first-class aggregate value, insert each value recursively. 879 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 880 const StructLayout &Layout = *TD.getStructLayout(ST); 881 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 882 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 883 Old = ConvertScalar_InsertValue(Elt, Old, 884 Offset+Layout.getElementOffsetInBits(i), 885 Builder); 886 } 887 return Old; 888 } 889 890 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 891 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 892 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 893 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 894 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 895 } 896 return Old; 897 } 898 899 // If SV is a float, convert it to the appropriate integer type. 900 // If it is a pointer, do the same. 901 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 902 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 903 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 904 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 905 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 906 SV = Builder.CreateBitCast(SV, 907 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 908 else if (SV->getType()->isPointerTy()) 909 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 910 911 // Zero extend or truncate the value if needed. 912 if (SV->getType() != AllocaType) { 913 if (SV->getType()->getPrimitiveSizeInBits() < 914 AllocaType->getPrimitiveSizeInBits()) 915 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 916 else { 917 // Truncation may be needed if storing more than the alloca can hold 918 // (undefined behavior). 919 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 920 SrcWidth = DestWidth; 921 SrcStoreWidth = DestStoreWidth; 922 } 923 } 924 925 // If this is a big-endian system and the store is narrower than the 926 // full alloca type, we need to do a shift to get the right bits. 927 int ShAmt = 0; 928 if (TD.isBigEndian()) { 929 // On big-endian machines, the lowest bit is stored at the bit offset 930 // from the pointer given by getTypeStoreSizeInBits. This matters for 931 // integers with a bitwidth that is not a multiple of 8. 932 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 933 } else { 934 ShAmt = Offset; 935 } 936 937 // Note: we support negative bitwidths (with shr) which are not defined. 938 // We do this to support (f.e.) stores off the end of a structure where 939 // only some bits in the structure are set. 940 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 941 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 942 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 943 ShAmt), "tmp"); 944 Mask <<= ShAmt; 945 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 946 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 947 -ShAmt), "tmp"); 948 Mask = Mask.lshr(-ShAmt); 949 } 950 951 // Mask out the bits we are about to insert from the old value, and or 952 // in the new bits. 953 if (SrcWidth != DestWidth) { 954 assert(DestWidth > SrcWidth); 955 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 956 SV = Builder.CreateOr(Old, SV, "ins"); 957 } 958 return SV; 959} 960 961 962//===----------------------------------------------------------------------===// 963// SRoA Driver 964//===----------------------------------------------------------------------===// 965 966 967bool SROA::runOnFunction(Function &F) { 968 TD = getAnalysisIfAvailable<TargetData>(); 969 970 bool Changed = performPromotion(F); 971 972 // FIXME: ScalarRepl currently depends on TargetData more than it 973 // theoretically needs to. It should be refactored in order to support 974 // target-independent IR. Until this is done, just skip the actual 975 // scalar-replacement portion of this pass. 976 if (!TD) return Changed; 977 978 while (1) { 979 bool LocalChange = performScalarRepl(F); 980 if (!LocalChange) break; // No need to repromote if no scalarrepl 981 Changed = true; 982 LocalChange = performPromotion(F); 983 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 984 } 985 986 return Changed; 987} 988 989namespace { 990class AllocaPromoter : public LoadAndStorePromoter { 991 AllocaInst *AI; 992public: 993 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S) 994 : LoadAndStorePromoter(Insts, S), AI(0) {} 995 996 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 997 // Remember which alloca we're promoting (for isInstInList). 998 this->AI = AI; 999 LoadAndStorePromoter::run(Insts); 1000 AI->eraseFromParent(); 1001 } 1002 1003 virtual bool isInstInList(Instruction *I, 1004 const SmallVectorImpl<Instruction*> &Insts) const { 1005 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1006 return LI->getOperand(0) == AI; 1007 return cast<StoreInst>(I)->getPointerOperand() == AI; 1008 } 1009}; 1010} // end anon namespace 1011 1012/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1013/// subsequently loaded can be rewritten to load both input pointers and then 1014/// select between the result, allowing the load of the alloca to be promoted. 1015/// From this: 1016/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1017/// %V = load i32* %P2 1018/// to: 1019/// %V1 = load i32* %Alloca -> will be mem2reg'd 1020/// %V2 = load i32* %Other 1021/// %V = select i1 %cond, i32 %V1, i32 %V2 1022/// 1023/// We can do this to a select if its only uses are loads and if the operand to 1024/// the select can be loaded unconditionally. 1025static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1026 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1027 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1028 1029 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1030 UI != UE; ++UI) { 1031 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1032 if (LI == 0 || LI->isVolatile()) return false; 1033 1034 // Both operands to the select need to be dereferencable, either absolutely 1035 // (e.g. allocas) or at this point because we can see other accesses to it. 1036 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1037 LI->getAlignment(), TD)) 1038 return false; 1039 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1040 LI->getAlignment(), TD)) 1041 return false; 1042 } 1043 1044 return true; 1045} 1046 1047/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1048/// subsequently loaded can be rewritten to load both input pointers in the pred 1049/// blocks and then PHI the results, allowing the load of the alloca to be 1050/// promoted. 1051/// From this: 1052/// %P2 = phi [i32* %Alloca, i32* %Other] 1053/// %V = load i32* %P2 1054/// to: 1055/// %V1 = load i32* %Alloca -> will be mem2reg'd 1056/// ... 1057/// %V2 = load i32* %Other 1058/// ... 1059/// %V = phi [i32 %V1, i32 %V2] 1060/// 1061/// We can do this to a select if its only uses are loads and if the operand to 1062/// the select can be loaded unconditionally. 1063static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1064 // For now, we can only do this promotion if the load is in the same block as 1065 // the PHI, and if there are no stores between the phi and load. 1066 // TODO: Allow recursive phi users. 1067 // TODO: Allow stores. 1068 BasicBlock *BB = PN->getParent(); 1069 unsigned MaxAlign = 0; 1070 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1071 UI != UE; ++UI) { 1072 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1073 if (LI == 0 || LI->isVolatile()) return false; 1074 1075 // For now we only allow loads in the same block as the PHI. This is a 1076 // common case that happens when instcombine merges two loads through a PHI. 1077 if (LI->getParent() != BB) return false; 1078 1079 // Ensure that there are no instructions between the PHI and the load that 1080 // could store. 1081 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1082 if (BBI->mayWriteToMemory()) 1083 return false; 1084 1085 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1086 } 1087 1088 // Okay, we know that we have one or more loads in the same block as the PHI. 1089 // We can transform this if it is safe to push the loads into the predecessor 1090 // blocks. The only thing to watch out for is that we can't put a possibly 1091 // trapping load in the predecessor if it is a critical edge. 1092 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1093 BasicBlock *Pred = PN->getIncomingBlock(i); 1094 1095 // If the predecessor has a single successor, then the edge isn't critical. 1096 if (Pred->getTerminator()->getNumSuccessors() == 1) 1097 continue; 1098 1099 Value *InVal = PN->getIncomingValue(i); 1100 1101 // If the InVal is an invoke in the pred, we can't put a load on the edge. 1102 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 1103 if (II->getParent() == Pred) 1104 return false; 1105 1106 // If this pointer is always safe to load, or if we can prove that there is 1107 // already a load in the block, then we can move the load to the pred block. 1108 if (InVal->isDereferenceablePointer() || 1109 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1110 continue; 1111 1112 return false; 1113 } 1114 1115 return true; 1116} 1117 1118 1119/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1120/// direct (non-volatile) loads and stores to it. If the alloca is close but 1121/// not quite there, this will transform the code to allow promotion. As such, 1122/// it is a non-pure predicate. 1123static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1124 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1125 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1126 1127 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1128 UI != UE; ++UI) { 1129 User *U = *UI; 1130 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1131 if (LI->isVolatile()) 1132 return false; 1133 continue; 1134 } 1135 1136 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1137 if (SI->getOperand(0) == AI || SI->isVolatile()) 1138 return false; // Don't allow a store OF the AI, only INTO the AI. 1139 continue; 1140 } 1141 1142 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1143 // If the condition being selected on is a constant, fold the select, yes 1144 // this does (rarely) happen early on. 1145 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1146 Value *Result = SI->getOperand(1+CI->isZero()); 1147 SI->replaceAllUsesWith(Result); 1148 SI->eraseFromParent(); 1149 1150 // This is very rare and we just scrambled the use list of AI, start 1151 // over completely. 1152 return tryToMakeAllocaBePromotable(AI, TD); 1153 } 1154 1155 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1156 // loads, then we can transform this by rewriting the select. 1157 if (!isSafeSelectToSpeculate(SI, TD)) 1158 return false; 1159 1160 InstsToRewrite.insert(SI); 1161 continue; 1162 } 1163 1164 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1165 if (PN->use_empty()) { // Dead PHIs can be stripped. 1166 InstsToRewrite.insert(PN); 1167 continue; 1168 } 1169 1170 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1171 // in the pred blocks, then we can transform this by rewriting the PHI. 1172 if (!isSafePHIToSpeculate(PN, TD)) 1173 return false; 1174 1175 InstsToRewrite.insert(PN); 1176 continue; 1177 } 1178 1179 return false; 1180 } 1181 1182 // If there are no instructions to rewrite, then all uses are load/stores and 1183 // we're done! 1184 if (InstsToRewrite.empty()) 1185 return true; 1186 1187 // If we have instructions that need to be rewritten for this to be promotable 1188 // take care of it now. 1189 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1190 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1191 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1192 // loads with a new select. 1193 while (!SI->use_empty()) { 1194 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1195 1196 IRBuilder<> Builder(LI); 1197 LoadInst *TrueLoad = 1198 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1199 LoadInst *FalseLoad = 1200 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".t"); 1201 1202 // Transfer alignment and TBAA info if present. 1203 TrueLoad->setAlignment(LI->getAlignment()); 1204 FalseLoad->setAlignment(LI->getAlignment()); 1205 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1206 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1207 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1208 } 1209 1210 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1211 V->takeName(LI); 1212 LI->replaceAllUsesWith(V); 1213 LI->eraseFromParent(); 1214 } 1215 1216 // Now that all the loads are gone, the select is gone too. 1217 SI->eraseFromParent(); 1218 continue; 1219 } 1220 1221 // Otherwise, we have a PHI node which allows us to push the loads into the 1222 // predecessors. 1223 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1224 if (PN->use_empty()) { 1225 PN->eraseFromParent(); 1226 continue; 1227 } 1228 1229 const Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1230 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1231 PN->getName()+".ld", PN); 1232 1233 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1234 // matter which one we get and if any differ, it doesn't matter. 1235 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1236 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1237 unsigned Align = SomeLoad->getAlignment(); 1238 1239 // Rewrite all loads of the PN to use the new PHI. 1240 while (!PN->use_empty()) { 1241 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1242 LI->replaceAllUsesWith(NewPN); 1243 LI->eraseFromParent(); 1244 } 1245 1246 // Inject loads into all of the pred blocks. Keep track of which blocks we 1247 // insert them into in case we have multiple edges from the same block. 1248 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1249 1250 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1251 BasicBlock *Pred = PN->getIncomingBlock(i); 1252 LoadInst *&Load = InsertedLoads[Pred]; 1253 if (Load == 0) { 1254 Load = new LoadInst(PN->getIncomingValue(i), 1255 PN->getName() + "." + Pred->getName(), 1256 Pred->getTerminator()); 1257 Load->setAlignment(Align); 1258 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1259 } 1260 1261 NewPN->addIncoming(Load, Pred); 1262 } 1263 1264 PN->eraseFromParent(); 1265 } 1266 1267 ++NumAdjusted; 1268 return true; 1269} 1270 1271 1272bool SROA::performPromotion(Function &F) { 1273 std::vector<AllocaInst*> Allocas; 1274 DominatorTree *DT = 0; 1275 if (HasDomTree) 1276 DT = &getAnalysis<DominatorTree>(); 1277 1278 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1279 1280 bool Changed = false; 1281 SmallVector<Instruction*, 64> Insts; 1282 while (1) { 1283 Allocas.clear(); 1284 1285 // Find allocas that are safe to promote, by looking at all instructions in 1286 // the entry node 1287 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1288 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1289 if (tryToMakeAllocaBePromotable(AI, TD)) 1290 Allocas.push_back(AI); 1291 1292 if (Allocas.empty()) break; 1293 1294 if (HasDomTree) 1295 PromoteMemToReg(Allocas, *DT); 1296 else { 1297 SSAUpdater SSA; 1298 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1299 AllocaInst *AI = Allocas[i]; 1300 1301 // Build list of instructions to promote. 1302 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1303 UI != E; ++UI) 1304 Insts.push_back(cast<Instruction>(*UI)); 1305 1306 AllocaPromoter(Insts, SSA).run(AI, Insts); 1307 Insts.clear(); 1308 } 1309 } 1310 NumPromoted += Allocas.size(); 1311 Changed = true; 1312 } 1313 1314 return Changed; 1315} 1316 1317 1318/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1319/// SROA. It must be a struct or array type with a small number of elements. 1320static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1321 const Type *T = AI->getAllocatedType(); 1322 // Do not promote any struct into more than 32 separate vars. 1323 if (const StructType *ST = dyn_cast<StructType>(T)) 1324 return ST->getNumElements() <= 32; 1325 // Arrays are much less likely to be safe for SROA; only consider 1326 // them if they are very small. 1327 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 1328 return AT->getNumElements() <= 8; 1329 return false; 1330} 1331 1332 1333// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1334// which runs on all of the malloc/alloca instructions in the function, removing 1335// them if they are only used by getelementptr instructions. 1336// 1337bool SROA::performScalarRepl(Function &F) { 1338 std::vector<AllocaInst*> WorkList; 1339 1340 // Scan the entry basic block, adding allocas to the worklist. 1341 BasicBlock &BB = F.getEntryBlock(); 1342 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1343 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1344 WorkList.push_back(A); 1345 1346 // Process the worklist 1347 bool Changed = false; 1348 while (!WorkList.empty()) { 1349 AllocaInst *AI = WorkList.back(); 1350 WorkList.pop_back(); 1351 1352 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1353 // with unused elements. 1354 if (AI->use_empty()) { 1355 AI->eraseFromParent(); 1356 Changed = true; 1357 continue; 1358 } 1359 1360 // If this alloca is impossible for us to promote, reject it early. 1361 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1362 continue; 1363 1364 // Check to see if this allocation is only modified by a memcpy/memmove from 1365 // a constant global. If this is the case, we can change all users to use 1366 // the constant global instead. This is commonly produced by the CFE by 1367 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1368 // is only subsequently read. 1369 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 1370 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1371 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 1372 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 1373 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1374 TheCopy->eraseFromParent(); // Don't mutate the global. 1375 AI->eraseFromParent(); 1376 ++NumGlobals; 1377 Changed = true; 1378 continue; 1379 } 1380 1381 // Check to see if we can perform the core SROA transformation. We cannot 1382 // transform the allocation instruction if it is an array allocation 1383 // (allocations OF arrays are ok though), and an allocation of a scalar 1384 // value cannot be decomposed at all. 1385 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1386 1387 // Do not promote [0 x %struct]. 1388 if (AllocaSize == 0) continue; 1389 1390 // Do not promote any struct whose size is too big. 1391 if (AllocaSize > SRThreshold) continue; 1392 1393 // If the alloca looks like a good candidate for scalar replacement, and if 1394 // all its users can be transformed, then split up the aggregate into its 1395 // separate elements. 1396 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1397 DoScalarReplacement(AI, WorkList); 1398 Changed = true; 1399 continue; 1400 } 1401 1402 // If we can turn this aggregate value (potentially with casts) into a 1403 // simple scalar value that can be mem2reg'd into a register value. 1404 // IsNotTrivial tracks whether this is something that mem2reg could have 1405 // promoted itself. If so, we don't want to transform it needlessly. Note 1406 // that we can't just check based on the type: the alloca may be of an i32 1407 // but that has pointer arithmetic to set byte 3 of it or something. 1408 if (AllocaInst *NewAI = 1409 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1410 NewAI->takeName(AI); 1411 AI->eraseFromParent(); 1412 ++NumConverted; 1413 Changed = true; 1414 continue; 1415 } 1416 1417 // Otherwise, couldn't process this alloca. 1418 } 1419 1420 return Changed; 1421} 1422 1423/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1424/// predicate, do SROA now. 1425void SROA::DoScalarReplacement(AllocaInst *AI, 1426 std::vector<AllocaInst*> &WorkList) { 1427 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1428 SmallVector<AllocaInst*, 32> ElementAllocas; 1429 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1430 ElementAllocas.reserve(ST->getNumContainedTypes()); 1431 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1432 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1433 AI->getAlignment(), 1434 AI->getName() + "." + Twine(i), AI); 1435 ElementAllocas.push_back(NA); 1436 WorkList.push_back(NA); // Add to worklist for recursive processing 1437 } 1438 } else { 1439 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1440 ElementAllocas.reserve(AT->getNumElements()); 1441 const Type *ElTy = AT->getElementType(); 1442 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1443 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1444 AI->getName() + "." + Twine(i), AI); 1445 ElementAllocas.push_back(NA); 1446 WorkList.push_back(NA); // Add to worklist for recursive processing 1447 } 1448 } 1449 1450 // Now that we have created the new alloca instructions, rewrite all the 1451 // uses of the old alloca. 1452 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1453 1454 // Now erase any instructions that were made dead while rewriting the alloca. 1455 DeleteDeadInstructions(); 1456 AI->eraseFromParent(); 1457 1458 ++NumReplaced; 1459} 1460 1461/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1462/// recursively including all their operands that become trivially dead. 1463void SROA::DeleteDeadInstructions() { 1464 while (!DeadInsts.empty()) { 1465 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1466 1467 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1468 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1469 // Zero out the operand and see if it becomes trivially dead. 1470 // (But, don't add allocas to the dead instruction list -- they are 1471 // already on the worklist and will be deleted separately.) 1472 *OI = 0; 1473 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1474 DeadInsts.push_back(U); 1475 } 1476 1477 I->eraseFromParent(); 1478 } 1479} 1480 1481/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1482/// performing scalar replacement of alloca AI. The results are flagged in 1483/// the Info parameter. Offset indicates the position within AI that is 1484/// referenced by this instruction. 1485void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1486 AllocaInfo &Info) { 1487 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1488 Instruction *User = cast<Instruction>(*UI); 1489 1490 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1491 isSafeForScalarRepl(BC, Offset, Info); 1492 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1493 uint64_t GEPOffset = Offset; 1494 isSafeGEP(GEPI, GEPOffset, Info); 1495 if (!Info.isUnsafe) 1496 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1497 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1498 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1499 if (Length == 0) 1500 return MarkUnsafe(Info, User); 1501 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1502 UI.getOperandNo() == 0, Info, MI, 1503 true /*AllowWholeAccess*/); 1504 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1505 if (LI->isVolatile()) 1506 return MarkUnsafe(Info, User); 1507 const Type *LIType = LI->getType(); 1508 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1509 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1510 Info.hasALoadOrStore = true; 1511 1512 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1513 // Store is ok if storing INTO the pointer, not storing the pointer 1514 if (SI->isVolatile() || SI->getOperand(0) == I) 1515 return MarkUnsafe(Info, User); 1516 1517 const Type *SIType = SI->getOperand(0)->getType(); 1518 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1519 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1520 Info.hasALoadOrStore = true; 1521 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1522 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1523 } else { 1524 return MarkUnsafe(Info, User); 1525 } 1526 if (Info.isUnsafe) return; 1527 } 1528} 1529 1530 1531/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1532/// derived from the alloca, we can often still split the alloca into elements. 1533/// This is useful if we have a large alloca where one element is phi'd 1534/// together somewhere: we can SRoA and promote all the other elements even if 1535/// we end up not being able to promote this one. 1536/// 1537/// All we require is that the uses of the PHI do not index into other parts of 1538/// the alloca. The most important use case for this is single load and stores 1539/// that are PHI'd together, which can happen due to code sinking. 1540void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1541 AllocaInfo &Info) { 1542 // If we've already checked this PHI, don't do it again. 1543 if (PHINode *PN = dyn_cast<PHINode>(I)) 1544 if (!Info.CheckedPHIs.insert(PN)) 1545 return; 1546 1547 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1548 Instruction *User = cast<Instruction>(*UI); 1549 1550 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1551 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1552 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1553 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1554 // but would have to prove that we're staying inside of an element being 1555 // promoted. 1556 if (!GEPI->hasAllZeroIndices()) 1557 return MarkUnsafe(Info, User); 1558 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1559 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1560 if (LI->isVolatile()) 1561 return MarkUnsafe(Info, User); 1562 const Type *LIType = LI->getType(); 1563 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1564 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1565 Info.hasALoadOrStore = true; 1566 1567 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1568 // Store is ok if storing INTO the pointer, not storing the pointer 1569 if (SI->isVolatile() || SI->getOperand(0) == I) 1570 return MarkUnsafe(Info, User); 1571 1572 const Type *SIType = SI->getOperand(0)->getType(); 1573 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1574 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1575 Info.hasALoadOrStore = true; 1576 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1577 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1578 } else { 1579 return MarkUnsafe(Info, User); 1580 } 1581 if (Info.isUnsafe) return; 1582 } 1583} 1584 1585/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1586/// replacement. It is safe when all the indices are constant, in-bounds 1587/// references, and when the resulting offset corresponds to an element within 1588/// the alloca type. The results are flagged in the Info parameter. Upon 1589/// return, Offset is adjusted as specified by the GEP indices. 1590void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1591 uint64_t &Offset, AllocaInfo &Info) { 1592 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1593 if (GEPIt == E) 1594 return; 1595 1596 // Walk through the GEP type indices, checking the types that this indexes 1597 // into. 1598 for (; GEPIt != E; ++GEPIt) { 1599 // Ignore struct elements, no extra checking needed for these. 1600 if ((*GEPIt)->isStructTy()) 1601 continue; 1602 1603 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1604 if (!IdxVal) 1605 return MarkUnsafe(Info, GEPI); 1606 } 1607 1608 // Compute the offset due to this GEP and check if the alloca has a 1609 // component element at that offset. 1610 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1611 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1612 &Indices[0], Indices.size()); 1613 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1614 MarkUnsafe(Info, GEPI); 1615} 1616 1617/// isHomogeneousAggregate - Check if type T is a struct or array containing 1618/// elements of the same type (which is always true for arrays). If so, 1619/// return true with NumElts and EltTy set to the number of elements and the 1620/// element type, respectively. 1621static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1622 const Type *&EltTy) { 1623 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1624 NumElts = AT->getNumElements(); 1625 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1626 return true; 1627 } 1628 if (const StructType *ST = dyn_cast<StructType>(T)) { 1629 NumElts = ST->getNumContainedTypes(); 1630 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1631 for (unsigned n = 1; n < NumElts; ++n) { 1632 if (ST->getContainedType(n) != EltTy) 1633 return false; 1634 } 1635 return true; 1636 } 1637 return false; 1638} 1639 1640/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1641/// "homogeneous" aggregates with the same element type and number of elements. 1642static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1643 if (T1 == T2) 1644 return true; 1645 1646 unsigned NumElts1, NumElts2; 1647 const Type *EltTy1, *EltTy2; 1648 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1649 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1650 NumElts1 == NumElts2 && 1651 EltTy1 == EltTy2) 1652 return true; 1653 1654 return false; 1655} 1656 1657/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1658/// alloca or has an offset and size that corresponds to a component element 1659/// within it. The offset checked here may have been formed from a GEP with a 1660/// pointer bitcasted to a different type. 1661/// 1662/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1663/// unit. If false, it only allows accesses known to be in a single element. 1664void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1665 const Type *MemOpType, bool isStore, 1666 AllocaInfo &Info, Instruction *TheAccess, 1667 bool AllowWholeAccess) { 1668 // Check if this is a load/store of the entire alloca. 1669 if (Offset == 0 && AllowWholeAccess && 1670 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1671 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1672 // loads/stores (which are essentially the same as the MemIntrinsics with 1673 // regard to copying padding between elements). But, if an alloca is 1674 // flagged as both a source and destination of such operations, we'll need 1675 // to check later for padding between elements. 1676 if (!MemOpType || MemOpType->isIntegerTy()) { 1677 if (isStore) 1678 Info.isMemCpyDst = true; 1679 else 1680 Info.isMemCpySrc = true; 1681 return; 1682 } 1683 // This is also safe for references using a type that is compatible with 1684 // the type of the alloca, so that loads/stores can be rewritten using 1685 // insertvalue/extractvalue. 1686 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1687 Info.hasSubelementAccess = true; 1688 return; 1689 } 1690 } 1691 // Check if the offset/size correspond to a component within the alloca type. 1692 const Type *T = Info.AI->getAllocatedType(); 1693 if (TypeHasComponent(T, Offset, MemSize)) { 1694 Info.hasSubelementAccess = true; 1695 return; 1696 } 1697 1698 return MarkUnsafe(Info, TheAccess); 1699} 1700 1701/// TypeHasComponent - Return true if T has a component type with the 1702/// specified offset and size. If Size is zero, do not check the size. 1703bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1704 const Type *EltTy; 1705 uint64_t EltSize; 1706 if (const StructType *ST = dyn_cast<StructType>(T)) { 1707 const StructLayout *Layout = TD->getStructLayout(ST); 1708 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1709 EltTy = ST->getContainedType(EltIdx); 1710 EltSize = TD->getTypeAllocSize(EltTy); 1711 Offset -= Layout->getElementOffset(EltIdx); 1712 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1713 EltTy = AT->getElementType(); 1714 EltSize = TD->getTypeAllocSize(EltTy); 1715 if (Offset >= AT->getNumElements() * EltSize) 1716 return false; 1717 Offset %= EltSize; 1718 } else { 1719 return false; 1720 } 1721 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1722 return true; 1723 // Check if the component spans multiple elements. 1724 if (Offset + Size > EltSize) 1725 return false; 1726 return TypeHasComponent(EltTy, Offset, Size); 1727} 1728 1729/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1730/// the instruction I, which references it, to use the separate elements. 1731/// Offset indicates the position within AI that is referenced by this 1732/// instruction. 1733void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1734 SmallVector<AllocaInst*, 32> &NewElts) { 1735 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1736 Use &TheUse = UI.getUse(); 1737 Instruction *User = cast<Instruction>(*UI++); 1738 1739 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1740 RewriteBitCast(BC, AI, Offset, NewElts); 1741 continue; 1742 } 1743 1744 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1745 RewriteGEP(GEPI, AI, Offset, NewElts); 1746 continue; 1747 } 1748 1749 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1750 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1751 uint64_t MemSize = Length->getZExtValue(); 1752 if (Offset == 0 && 1753 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1754 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1755 // Otherwise the intrinsic can only touch a single element and the 1756 // address operand will be updated, so nothing else needs to be done. 1757 continue; 1758 } 1759 1760 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1761 const Type *LIType = LI->getType(); 1762 1763 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1764 // Replace: 1765 // %res = load { i32, i32 }* %alloc 1766 // with: 1767 // %load.0 = load i32* %alloc.0 1768 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1769 // %load.1 = load i32* %alloc.1 1770 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1771 // (Also works for arrays instead of structs) 1772 Value *Insert = UndefValue::get(LIType); 1773 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1774 Value *Load = new LoadInst(NewElts[i], "load", LI); 1775 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1776 } 1777 LI->replaceAllUsesWith(Insert); 1778 DeadInsts.push_back(LI); 1779 } else if (LIType->isIntegerTy() && 1780 TD->getTypeAllocSize(LIType) == 1781 TD->getTypeAllocSize(AI->getAllocatedType())) { 1782 // If this is a load of the entire alloca to an integer, rewrite it. 1783 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1784 } 1785 continue; 1786 } 1787 1788 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1789 Value *Val = SI->getOperand(0); 1790 const Type *SIType = Val->getType(); 1791 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1792 // Replace: 1793 // store { i32, i32 } %val, { i32, i32 }* %alloc 1794 // with: 1795 // %val.0 = extractvalue { i32, i32 } %val, 0 1796 // store i32 %val.0, i32* %alloc.0 1797 // %val.1 = extractvalue { i32, i32 } %val, 1 1798 // store i32 %val.1, i32* %alloc.1 1799 // (Also works for arrays instead of structs) 1800 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1801 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1802 new StoreInst(Extract, NewElts[i], SI); 1803 } 1804 DeadInsts.push_back(SI); 1805 } else if (SIType->isIntegerTy() && 1806 TD->getTypeAllocSize(SIType) == 1807 TD->getTypeAllocSize(AI->getAllocatedType())) { 1808 // If this is a store of the entire alloca from an integer, rewrite it. 1809 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1810 } 1811 continue; 1812 } 1813 1814 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1815 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1816 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1817 // the new pointer. 1818 if (!isa<AllocaInst>(I)) continue; 1819 1820 assert(Offset == 0 && NewElts[0] && 1821 "Direct alloca use should have a zero offset"); 1822 1823 // If we have a use of the alloca, we know the derived uses will be 1824 // utilizing just the first element of the scalarized result. Insert a 1825 // bitcast of the first alloca before the user as required. 1826 AllocaInst *NewAI = NewElts[0]; 1827 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1828 NewAI->moveBefore(BCI); 1829 TheUse = BCI; 1830 continue; 1831 } 1832 } 1833} 1834 1835/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1836/// and recursively continue updating all of its uses. 1837void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1838 SmallVector<AllocaInst*, 32> &NewElts) { 1839 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1840 if (BC->getOperand(0) != AI) 1841 return; 1842 1843 // The bitcast references the original alloca. Replace its uses with 1844 // references to the first new element alloca. 1845 Instruction *Val = NewElts[0]; 1846 if (Val->getType() != BC->getDestTy()) { 1847 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1848 Val->takeName(BC); 1849 } 1850 BC->replaceAllUsesWith(Val); 1851 DeadInsts.push_back(BC); 1852} 1853 1854/// FindElementAndOffset - Return the index of the element containing Offset 1855/// within the specified type, which must be either a struct or an array. 1856/// Sets T to the type of the element and Offset to the offset within that 1857/// element. IdxTy is set to the type of the index result to be used in a 1858/// GEP instruction. 1859uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1860 const Type *&IdxTy) { 1861 uint64_t Idx = 0; 1862 if (const StructType *ST = dyn_cast<StructType>(T)) { 1863 const StructLayout *Layout = TD->getStructLayout(ST); 1864 Idx = Layout->getElementContainingOffset(Offset); 1865 T = ST->getContainedType(Idx); 1866 Offset -= Layout->getElementOffset(Idx); 1867 IdxTy = Type::getInt32Ty(T->getContext()); 1868 return Idx; 1869 } 1870 const ArrayType *AT = cast<ArrayType>(T); 1871 T = AT->getElementType(); 1872 uint64_t EltSize = TD->getTypeAllocSize(T); 1873 Idx = Offset / EltSize; 1874 Offset -= Idx * EltSize; 1875 IdxTy = Type::getInt64Ty(T->getContext()); 1876 return Idx; 1877} 1878 1879/// RewriteGEP - Check if this GEP instruction moves the pointer across 1880/// elements of the alloca that are being split apart, and if so, rewrite 1881/// the GEP to be relative to the new element. 1882void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1883 SmallVector<AllocaInst*, 32> &NewElts) { 1884 uint64_t OldOffset = Offset; 1885 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1886 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1887 &Indices[0], Indices.size()); 1888 1889 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1890 1891 const Type *T = AI->getAllocatedType(); 1892 const Type *IdxTy; 1893 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1894 if (GEPI->getOperand(0) == AI) 1895 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1896 1897 T = AI->getAllocatedType(); 1898 uint64_t EltOffset = Offset; 1899 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1900 1901 // If this GEP does not move the pointer across elements of the alloca 1902 // being split, then it does not needs to be rewritten. 1903 if (Idx == OldIdx) 1904 return; 1905 1906 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1907 SmallVector<Value*, 8> NewArgs; 1908 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1909 while (EltOffset != 0) { 1910 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1911 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1912 } 1913 Instruction *Val = NewElts[Idx]; 1914 if (NewArgs.size() > 1) { 1915 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1916 NewArgs.end(), "", GEPI); 1917 Val->takeName(GEPI); 1918 } 1919 if (Val->getType() != GEPI->getType()) 1920 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1921 GEPI->replaceAllUsesWith(Val); 1922 DeadInsts.push_back(GEPI); 1923} 1924 1925/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1926/// Rewrite it to copy or set the elements of the scalarized memory. 1927void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1928 AllocaInst *AI, 1929 SmallVector<AllocaInst*, 32> &NewElts) { 1930 // If this is a memcpy/memmove, construct the other pointer as the 1931 // appropriate type. The "Other" pointer is the pointer that goes to memory 1932 // that doesn't have anything to do with the alloca that we are promoting. For 1933 // memset, this Value* stays null. 1934 Value *OtherPtr = 0; 1935 unsigned MemAlignment = MI->getAlignment(); 1936 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1937 if (Inst == MTI->getRawDest()) 1938 OtherPtr = MTI->getRawSource(); 1939 else { 1940 assert(Inst == MTI->getRawSource()); 1941 OtherPtr = MTI->getRawDest(); 1942 } 1943 } 1944 1945 // If there is an other pointer, we want to convert it to the same pointer 1946 // type as AI has, so we can GEP through it safely. 1947 if (OtherPtr) { 1948 unsigned AddrSpace = 1949 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 1950 1951 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1952 // optimization, but it's also required to detect the corner case where 1953 // both pointer operands are referencing the same memory, and where 1954 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1955 // function is only called for mem intrinsics that access the whole 1956 // aggregate, so non-zero GEPs are not an issue here.) 1957 OtherPtr = OtherPtr->stripPointerCasts(); 1958 1959 // Copying the alloca to itself is a no-op: just delete it. 1960 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1961 // This code will run twice for a no-op memcpy -- once for each operand. 1962 // Put only one reference to MI on the DeadInsts list. 1963 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1964 E = DeadInsts.end(); I != E; ++I) 1965 if (*I == MI) return; 1966 DeadInsts.push_back(MI); 1967 return; 1968 } 1969 1970 // If the pointer is not the right type, insert a bitcast to the right 1971 // type. 1972 const Type *NewTy = 1973 PointerType::get(AI->getType()->getElementType(), AddrSpace); 1974 1975 if (OtherPtr->getType() != NewTy) 1976 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 1977 } 1978 1979 // Process each element of the aggregate. 1980 bool SROADest = MI->getRawDest() == Inst; 1981 1982 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1983 1984 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1985 // If this is a memcpy/memmove, emit a GEP of the other element address. 1986 Value *OtherElt = 0; 1987 unsigned OtherEltAlign = MemAlignment; 1988 1989 if (OtherPtr) { 1990 Value *Idx[2] = { Zero, 1991 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1992 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1993 OtherPtr->getName()+"."+Twine(i), 1994 MI); 1995 uint64_t EltOffset; 1996 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1997 const Type *OtherTy = OtherPtrTy->getElementType(); 1998 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1999 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2000 } else { 2001 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2002 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2003 } 2004 2005 // The alignment of the other pointer is the guaranteed alignment of the 2006 // element, which is affected by both the known alignment of the whole 2007 // mem intrinsic and the alignment of the element. If the alignment of 2008 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2009 // known alignment is just 4 bytes. 2010 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2011 } 2012 2013 Value *EltPtr = NewElts[i]; 2014 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2015 2016 // If we got down to a scalar, insert a load or store as appropriate. 2017 if (EltTy->isSingleValueType()) { 2018 if (isa<MemTransferInst>(MI)) { 2019 if (SROADest) { 2020 // From Other to Alloca. 2021 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2022 new StoreInst(Elt, EltPtr, MI); 2023 } else { 2024 // From Alloca to Other. 2025 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2026 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2027 } 2028 continue; 2029 } 2030 assert(isa<MemSetInst>(MI)); 2031 2032 // If the stored element is zero (common case), just store a null 2033 // constant. 2034 Constant *StoreVal; 2035 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2036 if (CI->isZero()) { 2037 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2038 } else { 2039 // If EltTy is a vector type, get the element type. 2040 const Type *ValTy = EltTy->getScalarType(); 2041 2042 // Construct an integer with the right value. 2043 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2044 APInt OneVal(EltSize, CI->getZExtValue()); 2045 APInt TotalVal(OneVal); 2046 // Set each byte. 2047 for (unsigned i = 0; 8*i < EltSize; ++i) { 2048 TotalVal = TotalVal.shl(8); 2049 TotalVal |= OneVal; 2050 } 2051 2052 // Convert the integer value to the appropriate type. 2053 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2054 if (ValTy->isPointerTy()) 2055 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2056 else if (ValTy->isFloatingPointTy()) 2057 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2058 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2059 2060 // If the requested value was a vector constant, create it. 2061 if (EltTy != ValTy) { 2062 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 2063 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2064 StoreVal = ConstantVector::get(Elts); 2065 } 2066 } 2067 new StoreInst(StoreVal, EltPtr, MI); 2068 continue; 2069 } 2070 // Otherwise, if we're storing a byte variable, use a memset call for 2071 // this element. 2072 } 2073 2074 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2075 2076 IRBuilder<> Builder(MI); 2077 2078 // Finally, insert the meminst for this element. 2079 if (isa<MemSetInst>(MI)) { 2080 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2081 MI->isVolatile()); 2082 } else { 2083 assert(isa<MemTransferInst>(MI)); 2084 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2085 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2086 2087 if (isa<MemCpyInst>(MI)) 2088 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2089 else 2090 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2091 } 2092 } 2093 DeadInsts.push_back(MI); 2094} 2095 2096/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2097/// overwrites the entire allocation. Extract out the pieces of the stored 2098/// integer and store them individually. 2099void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2100 SmallVector<AllocaInst*, 32> &NewElts){ 2101 // Extract each element out of the integer according to its structure offset 2102 // and store the element value to the individual alloca. 2103 Value *SrcVal = SI->getOperand(0); 2104 const Type *AllocaEltTy = AI->getAllocatedType(); 2105 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2106 2107 IRBuilder<> Builder(SI); 2108 2109 // Handle tail padding by extending the operand 2110 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2111 SrcVal = Builder.CreateZExt(SrcVal, 2112 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2113 2114 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2115 << '\n'); 2116 2117 // There are two forms here: AI could be an array or struct. Both cases 2118 // have different ways to compute the element offset. 2119 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2120 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2121 2122 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2123 // Get the number of bits to shift SrcVal to get the value. 2124 const Type *FieldTy = EltSTy->getElementType(i); 2125 uint64_t Shift = Layout->getElementOffsetInBits(i); 2126 2127 if (TD->isBigEndian()) 2128 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2129 2130 Value *EltVal = SrcVal; 2131 if (Shift) { 2132 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2133 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2134 } 2135 2136 // Truncate down to an integer of the right size. 2137 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2138 2139 // Ignore zero sized fields like {}, they obviously contain no data. 2140 if (FieldSizeBits == 0) continue; 2141 2142 if (FieldSizeBits != AllocaSizeBits) 2143 EltVal = Builder.CreateTrunc(EltVal, 2144 IntegerType::get(SI->getContext(), FieldSizeBits)); 2145 Value *DestField = NewElts[i]; 2146 if (EltVal->getType() == FieldTy) { 2147 // Storing to an integer field of this size, just do it. 2148 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2149 // Bitcast to the right element type (for fp/vector values). 2150 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2151 } else { 2152 // Otherwise, bitcast the dest pointer (for aggregates). 2153 DestField = Builder.CreateBitCast(DestField, 2154 PointerType::getUnqual(EltVal->getType())); 2155 } 2156 new StoreInst(EltVal, DestField, SI); 2157 } 2158 2159 } else { 2160 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2161 const Type *ArrayEltTy = ATy->getElementType(); 2162 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2163 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2164 2165 uint64_t Shift; 2166 2167 if (TD->isBigEndian()) 2168 Shift = AllocaSizeBits-ElementOffset; 2169 else 2170 Shift = 0; 2171 2172 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2173 // Ignore zero sized fields like {}, they obviously contain no data. 2174 if (ElementSizeBits == 0) continue; 2175 2176 Value *EltVal = SrcVal; 2177 if (Shift) { 2178 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2179 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2180 } 2181 2182 // Truncate down to an integer of the right size. 2183 if (ElementSizeBits != AllocaSizeBits) 2184 EltVal = Builder.CreateTrunc(EltVal, 2185 IntegerType::get(SI->getContext(), 2186 ElementSizeBits)); 2187 Value *DestField = NewElts[i]; 2188 if (EltVal->getType() == ArrayEltTy) { 2189 // Storing to an integer field of this size, just do it. 2190 } else if (ArrayEltTy->isFloatingPointTy() || 2191 ArrayEltTy->isVectorTy()) { 2192 // Bitcast to the right element type (for fp/vector values). 2193 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2194 } else { 2195 // Otherwise, bitcast the dest pointer (for aggregates). 2196 DestField = Builder.CreateBitCast(DestField, 2197 PointerType::getUnqual(EltVal->getType())); 2198 } 2199 new StoreInst(EltVal, DestField, SI); 2200 2201 if (TD->isBigEndian()) 2202 Shift -= ElementOffset; 2203 else 2204 Shift += ElementOffset; 2205 } 2206 } 2207 2208 DeadInsts.push_back(SI); 2209} 2210 2211/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2212/// an integer. Load the individual pieces to form the aggregate value. 2213void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2214 SmallVector<AllocaInst*, 32> &NewElts) { 2215 // Extract each element out of the NewElts according to its structure offset 2216 // and form the result value. 2217 const Type *AllocaEltTy = AI->getAllocatedType(); 2218 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2219 2220 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2221 << '\n'); 2222 2223 // There are two forms here: AI could be an array or struct. Both cases 2224 // have different ways to compute the element offset. 2225 const StructLayout *Layout = 0; 2226 uint64_t ArrayEltBitOffset = 0; 2227 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2228 Layout = TD->getStructLayout(EltSTy); 2229 } else { 2230 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2231 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2232 } 2233 2234 Value *ResultVal = 2235 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2236 2237 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2238 // Load the value from the alloca. If the NewElt is an aggregate, cast 2239 // the pointer to an integer of the same size before doing the load. 2240 Value *SrcField = NewElts[i]; 2241 const Type *FieldTy = 2242 cast<PointerType>(SrcField->getType())->getElementType(); 2243 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2244 2245 // Ignore zero sized fields like {}, they obviously contain no data. 2246 if (FieldSizeBits == 0) continue; 2247 2248 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2249 FieldSizeBits); 2250 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2251 !FieldTy->isVectorTy()) 2252 SrcField = new BitCastInst(SrcField, 2253 PointerType::getUnqual(FieldIntTy), 2254 "", LI); 2255 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2256 2257 // If SrcField is a fp or vector of the right size but that isn't an 2258 // integer type, bitcast to an integer so we can shift it. 2259 if (SrcField->getType() != FieldIntTy) 2260 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2261 2262 // Zero extend the field to be the same size as the final alloca so that 2263 // we can shift and insert it. 2264 if (SrcField->getType() != ResultVal->getType()) 2265 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2266 2267 // Determine the number of bits to shift SrcField. 2268 uint64_t Shift; 2269 if (Layout) // Struct case. 2270 Shift = Layout->getElementOffsetInBits(i); 2271 else // Array case. 2272 Shift = i*ArrayEltBitOffset; 2273 2274 if (TD->isBigEndian()) 2275 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2276 2277 if (Shift) { 2278 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2279 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2280 } 2281 2282 // Don't create an 'or x, 0' on the first iteration. 2283 if (!isa<Constant>(ResultVal) || 2284 !cast<Constant>(ResultVal)->isNullValue()) 2285 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2286 else 2287 ResultVal = SrcField; 2288 } 2289 2290 // Handle tail padding by truncating the result 2291 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2292 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2293 2294 LI->replaceAllUsesWith(ResultVal); 2295 DeadInsts.push_back(LI); 2296} 2297 2298/// HasPadding - Return true if the specified type has any structure or 2299/// alignment padding in between the elements that would be split apart 2300/// by SROA; return false otherwise. 2301static bool HasPadding(const Type *Ty, const TargetData &TD) { 2302 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2303 Ty = ATy->getElementType(); 2304 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2305 } 2306 2307 // SROA currently handles only Arrays and Structs. 2308 const StructType *STy = cast<StructType>(Ty); 2309 const StructLayout *SL = TD.getStructLayout(STy); 2310 unsigned PrevFieldBitOffset = 0; 2311 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2312 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2313 2314 // Check to see if there is any padding between this element and the 2315 // previous one. 2316 if (i) { 2317 unsigned PrevFieldEnd = 2318 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2319 if (PrevFieldEnd < FieldBitOffset) 2320 return true; 2321 } 2322 PrevFieldBitOffset = FieldBitOffset; 2323 } 2324 // Check for tail padding. 2325 if (unsigned EltCount = STy->getNumElements()) { 2326 unsigned PrevFieldEnd = PrevFieldBitOffset + 2327 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2328 if (PrevFieldEnd < SL->getSizeInBits()) 2329 return true; 2330 } 2331 return false; 2332} 2333 2334/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2335/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2336/// or 1 if safe after canonicalization has been performed. 2337bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2338 // Loop over the use list of the alloca. We can only transform it if all of 2339 // the users are safe to transform. 2340 AllocaInfo Info(AI); 2341 2342 isSafeForScalarRepl(AI, 0, Info); 2343 if (Info.isUnsafe) { 2344 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2345 return false; 2346 } 2347 2348 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2349 // source and destination, we have to be careful. In particular, the memcpy 2350 // could be moving around elements that live in structure padding of the LLVM 2351 // types, but may actually be used. In these cases, we refuse to promote the 2352 // struct. 2353 if (Info.isMemCpySrc && Info.isMemCpyDst && 2354 HasPadding(AI->getAllocatedType(), *TD)) 2355 return false; 2356 2357 // If the alloca never has an access to just *part* of it, but is accessed 2358 // via loads and stores, then we should use ConvertToScalarInfo to promote 2359 // the alloca instead of promoting each piece at a time and inserting fission 2360 // and fusion code. 2361 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2362 // If the struct/array just has one element, use basic SRoA. 2363 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2364 if (ST->getNumElements() > 1) return false; 2365 } else { 2366 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2367 return false; 2368 } 2369 } 2370 2371 return true; 2372} 2373 2374 2375 2376/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2377/// some part of a constant global variable. This intentionally only accepts 2378/// constant expressions because we don't can't rewrite arbitrary instructions. 2379static bool PointsToConstantGlobal(Value *V) { 2380 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2381 return GV->isConstant(); 2382 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2383 if (CE->getOpcode() == Instruction::BitCast || 2384 CE->getOpcode() == Instruction::GetElementPtr) 2385 return PointsToConstantGlobal(CE->getOperand(0)); 2386 return false; 2387} 2388 2389/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2390/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2391/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2392/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2393/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2394/// the alloca, and if the source pointer is a pointer to a constant global, we 2395/// can optimize this. 2396static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2397 bool isOffset) { 2398 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2399 User *U = cast<Instruction>(*UI); 2400 2401 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2402 // Ignore non-volatile loads, they are always ok. 2403 if (LI->isVolatile()) return false; 2404 continue; 2405 } 2406 2407 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2408 // If uses of the bitcast are ok, we are ok. 2409 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 2410 return false; 2411 continue; 2412 } 2413 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2414 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2415 // doesn't, it does. 2416 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2417 isOffset || !GEP->hasAllZeroIndices())) 2418 return false; 2419 continue; 2420 } 2421 2422 if (CallSite CS = U) { 2423 // If this is a readonly/readnone call site, then we know it is just a 2424 // load and we can ignore it. 2425 if (CS.onlyReadsMemory()) 2426 continue; 2427 2428 // If this is the function being called then we treat it like a load and 2429 // ignore it. 2430 if (CS.isCallee(UI)) 2431 continue; 2432 2433 // If this is being passed as a byval argument, the caller is making a 2434 // copy, so it is only a read of the alloca. 2435 unsigned ArgNo = CS.getArgumentNo(UI); 2436 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2437 continue; 2438 } 2439 2440 // If this is isn't our memcpy/memmove, reject it as something we can't 2441 // handle. 2442 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2443 if (MI == 0) 2444 return false; 2445 2446 // If the transfer is using the alloca as a source of the transfer, then 2447 // ignore it since it is a load (unless the transfer is volatile). 2448 if (UI.getOperandNo() == 1) { 2449 if (MI->isVolatile()) return false; 2450 continue; 2451 } 2452 2453 // If we already have seen a copy, reject the second one. 2454 if (TheCopy) return false; 2455 2456 // If the pointer has been offset from the start of the alloca, we can't 2457 // safely handle this. 2458 if (isOffset) return false; 2459 2460 // If the memintrinsic isn't using the alloca as the dest, reject it. 2461 if (UI.getOperandNo() != 0) return false; 2462 2463 // If the source of the memcpy/move is not a constant global, reject it. 2464 if (!PointsToConstantGlobal(MI->getSource())) 2465 return false; 2466 2467 // Otherwise, the transform is safe. Remember the copy instruction. 2468 TheCopy = MI; 2469 } 2470 return true; 2471} 2472 2473/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2474/// modified by a copy from a constant global. If we can prove this, we can 2475/// replace any uses of the alloca with uses of the global directly. 2476MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 2477 MemTransferInst *TheCopy = 0; 2478 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 2479 return TheCopy; 2480 return 0; 2481} 2482