ScalarReplAggregates.cpp revision 1537ce75ed25bbca58096383bb1fb9dd427bf1aa
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); 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 // If we already decided to turn this into a blob of integer memory, there is 320 // nothing to be done. 321 if (VectorTy && VectorTy->isVoidTy()) 322 return; 323 324 // If this could be contributing to a vector, analyze it. 325 326 // If the In type is a vector that is the same size as the alloca, see if it 327 // matches the existing VecTy. 328 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 329 if (MergeInVectorType(VInTy, Offset)) 330 return; 331 } else if (In->isFloatTy() || In->isDoubleTy() || 332 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 333 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 334 // If we're accessing something that could be an element of a vector, see 335 // if the implied vector agrees with what we already have and if Offset is 336 // compatible with it. 337 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 338 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 339 (VectorTy == 0 || 340 cast<VectorType>(VectorTy)->getElementType() 341 ->getPrimitiveSizeInBits()/8 == EltSize)) { 342 if (VectorTy == 0) 343 VectorTy = VectorType::get(In, AllocaSize/EltSize); 344 return; 345 } 346 } 347 348 // Otherwise, we have a case that we can't handle with an optimized vector 349 // form. We can still turn this into a large integer. 350 VectorTy = Type::getVoidTy(In->getContext()); 351} 352 353/// MergeInVectorType - Handles the vector case of MergeInType, returning true 354/// if the type was successfully merged and false otherwise. 355bool ConvertToScalarInfo::MergeInVectorType(const VectorType *VInTy, 356 uint64_t Offset) { 357 // Remember if we saw a vector type. 358 HadAVector = true; 359 360 // TODO: Support nonzero offsets? 361 if (Offset != 0) 362 return false; 363 364 // Only allow vectors that are a power-of-2 away from the size of the alloca. 365 if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8))) 366 return false; 367 368 // If this the first vector we see, remember the type so that we know the 369 // element size. 370 if (!VectorTy) { 371 VectorTy = VInTy; 372 return true; 373 } 374 375 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth(); 376 unsigned InBitWidth = VInTy->getBitWidth(); 377 378 // Vectors of the same size can be converted using a simple bitcast. 379 if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8)) 380 return true; 381 382 const Type *ElementTy = cast<VectorType>(VectorTy)->getElementType(); 383 const Type *InElementTy = cast<VectorType>(VectorTy)->getElementType(); 384 385 // Do not allow mixed integer and floating-point accesses from vectors of 386 // different sizes. 387 if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy()) 388 return false; 389 390 if (ElementTy->isFloatingPointTy()) { 391 // Only allow floating-point vectors of different sizes if they have the 392 // same element type. 393 // TODO: This could be loosened a bit, but would anything benefit? 394 if (ElementTy != InElementTy) 395 return false; 396 397 // There are no arbitrary-precision floating-point types, which limits the 398 // number of legal vector types with larger element types that we can form 399 // to bitcast and extract a subvector. 400 // TODO: We could support some more cases with mixed fp128 and double here. 401 if (!(BitWidth == 64 || BitWidth == 128) || 402 !(InBitWidth == 64 || InBitWidth == 128)) 403 return false; 404 } else { 405 assert(ElementTy->isIntegerTy() && "Vector elements must be either integer " 406 "or floating-point."); 407 unsigned BitWidth = ElementTy->getPrimitiveSizeInBits(); 408 unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits(); 409 410 // Do not allow integer types smaller than a byte or types whose widths are 411 // not a multiple of a byte. 412 if (BitWidth < 8 || InBitWidth < 8 || 413 BitWidth % 8 != 0 || InBitWidth % 8 != 0) 414 return false; 415 } 416 417 // Pick the largest of the two vector types. 418 if (InBitWidth > BitWidth) 419 VectorTy = VInTy; 420 421 return true; 422} 423 424/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 425/// its accesses to a single vector type, return true and set VecTy to 426/// the new type. If we could convert the alloca into a single promotable 427/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 428/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 429/// is the current offset from the base of the alloca being analyzed. 430/// 431/// If we see at least one access to the value that is as a vector type, set the 432/// SawVec flag. 433bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 434 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 435 Instruction *User = cast<Instruction>(*UI); 436 437 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 438 // Don't break volatile loads. 439 if (LI->isVolatile()) 440 return false; 441 // Don't touch MMX operations. 442 if (LI->getType()->isX86_MMXTy()) 443 return false; 444 HadNonMemTransferAccess = true; 445 MergeInType(LI->getType(), Offset); 446 continue; 447 } 448 449 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 450 // Storing the pointer, not into the value? 451 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 452 // Don't touch MMX operations. 453 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 454 return false; 455 HadNonMemTransferAccess = true; 456 MergeInType(SI->getOperand(0)->getType(), Offset); 457 continue; 458 } 459 460 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 461 IsNotTrivial = true; // Can't be mem2reg'd. 462 if (!CanConvertToScalar(BCI, Offset)) 463 return false; 464 continue; 465 } 466 467 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 468 // If this is a GEP with a variable indices, we can't handle it. 469 if (!GEP->hasAllConstantIndices()) 470 return false; 471 472 // Compute the offset that this GEP adds to the pointer. 473 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 474 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 475 &Indices[0], Indices.size()); 476 // See if all uses can be converted. 477 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 478 return false; 479 IsNotTrivial = true; // Can't be mem2reg'd. 480 HadNonMemTransferAccess = true; 481 continue; 482 } 483 484 // If this is a constant sized memset of a constant value (e.g. 0) we can 485 // handle it. 486 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 487 // Store of constant value and constant size. 488 if (!isa<ConstantInt>(MSI->getValue()) || 489 !isa<ConstantInt>(MSI->getLength())) 490 return false; 491 IsNotTrivial = true; // Can't be mem2reg'd. 492 HadNonMemTransferAccess = true; 493 continue; 494 } 495 496 // If this is a memcpy or memmove into or out of the whole allocation, we 497 // can handle it like a load or store of the scalar type. 498 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 499 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 500 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 501 return false; 502 503 IsNotTrivial = true; // Can't be mem2reg'd. 504 continue; 505 } 506 507 // Otherwise, we cannot handle this! 508 return false; 509 } 510 511 return true; 512} 513 514/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 515/// directly. This happens when we are converting an "integer union" to a 516/// single integer scalar, or when we are converting a "vector union" to a 517/// vector with insert/extractelement instructions. 518/// 519/// Offset is an offset from the original alloca, in bits that need to be 520/// shifted to the right. By the end of this, there should be no uses of Ptr. 521void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 522 uint64_t Offset) { 523 while (!Ptr->use_empty()) { 524 Instruction *User = cast<Instruction>(Ptr->use_back()); 525 526 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 527 ConvertUsesToScalar(CI, NewAI, Offset); 528 CI->eraseFromParent(); 529 continue; 530 } 531 532 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 533 // Compute the offset that this GEP adds to the pointer. 534 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 535 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 536 &Indices[0], Indices.size()); 537 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 538 GEP->eraseFromParent(); 539 continue; 540 } 541 542 IRBuilder<> Builder(User); 543 544 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 545 // The load is a bit extract from NewAI shifted right by Offset bits. 546 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 547 Value *NewLoadVal 548 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 549 LI->replaceAllUsesWith(NewLoadVal); 550 LI->eraseFromParent(); 551 continue; 552 } 553 554 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 555 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 556 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 557 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 558 Builder); 559 Builder.CreateStore(New, NewAI); 560 SI->eraseFromParent(); 561 562 // If the load we just inserted is now dead, then the inserted store 563 // overwrote the entire thing. 564 if (Old->use_empty()) 565 Old->eraseFromParent(); 566 continue; 567 } 568 569 // If this is a constant sized memset of a constant value (e.g. 0) we can 570 // transform it into a store of the expanded constant value. 571 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 572 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 573 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 574 if (NumBytes != 0) { 575 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 576 577 // Compute the value replicated the right number of times. 578 APInt APVal(NumBytes*8, Val); 579 580 // Splat the value if non-zero. 581 if (Val) 582 for (unsigned i = 1; i != NumBytes; ++i) 583 APVal |= APVal << 8; 584 585 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 586 Value *New = ConvertScalar_InsertValue( 587 ConstantInt::get(User->getContext(), APVal), 588 Old, Offset, Builder); 589 Builder.CreateStore(New, NewAI); 590 591 // If the load we just inserted is now dead, then the memset overwrote 592 // the entire thing. 593 if (Old->use_empty()) 594 Old->eraseFromParent(); 595 } 596 MSI->eraseFromParent(); 597 continue; 598 } 599 600 // If this is a memcpy or memmove into or out of the whole allocation, we 601 // can handle it like a load or store of the scalar type. 602 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 603 assert(Offset == 0 && "must be store to start of alloca"); 604 605 // If the source and destination are both to the same alloca, then this is 606 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 607 // as appropriate. 608 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 609 610 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 611 // Dest must be OrigAI, change this to be a load from the original 612 // pointer (bitcasted), then a store to our new alloca. 613 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 614 Value *SrcPtr = MTI->getSource(); 615 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 616 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 617 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 618 AIPTy = PointerType::get(AIPTy->getElementType(), 619 SPTy->getAddressSpace()); 620 } 621 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 622 623 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 624 SrcVal->setAlignment(MTI->getAlignment()); 625 Builder.CreateStore(SrcVal, NewAI); 626 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 627 // Src must be OrigAI, change this to be a load from NewAI then a store 628 // through the original dest pointer (bitcasted). 629 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 630 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 631 632 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 633 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 634 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 635 AIPTy = PointerType::get(AIPTy->getElementType(), 636 DPTy->getAddressSpace()); 637 } 638 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 639 640 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 641 NewStore->setAlignment(MTI->getAlignment()); 642 } else { 643 // Noop transfer. Src == Dst 644 } 645 646 MTI->eraseFromParent(); 647 continue; 648 } 649 650 llvm_unreachable("Unsupported operation!"); 651 } 652} 653 654/// getScaledElementType - Gets a scaled element type for a partial vector 655/// access of an alloca. The input type must be an integer or float, and 656/// the resulting type must be an integer, float or double. 657static const Type *getScaledElementType(const Type *OldTy, 658 unsigned NewBitWidth) { 659 assert((OldTy->isIntegerTy() || OldTy->isFloatTy()) && "Partial vector " 660 "accesses must be scaled from integer or float elements."); 661 662 LLVMContext &Context = OldTy->getContext(); 663 664 if (OldTy->isIntegerTy()) 665 return Type::getIntNTy(Context, NewBitWidth); 666 if (NewBitWidth == 32) 667 return Type::getFloatTy(Context); 668 if (NewBitWidth == 64) 669 return Type::getDoubleTy(Context); 670 671 llvm_unreachable("Invalid type for a partial vector access of an alloca!"); 672} 673 674/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 675/// or vector value FromVal, extracting the bits from the offset specified by 676/// Offset. This returns the value, which is of type ToType. 677/// 678/// This happens when we are converting an "integer union" to a single 679/// integer scalar, or when we are converting a "vector union" to a vector with 680/// insert/extractelement instructions. 681/// 682/// Offset is an offset from the original alloca, in bits that need to be 683/// shifted to the right. 684Value *ConvertToScalarInfo:: 685ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 686 uint64_t Offset, IRBuilder<> &Builder) { 687 // If the load is of the whole new alloca, no conversion is needed. 688 if (FromVal->getType() == ToType && Offset == 0) 689 return FromVal; 690 691 // If the result alloca is a vector type, this is either an element 692 // access or a bitcast to another vector type of the same size. 693 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 694 if (ToType->isVectorTy()) { 695 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 696 if (ToTypeSize == AllocaSize) 697 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 698 699 assert(isPowerOf2_64(AllocaSize / ToTypeSize) && 700 "Partial vector access of an alloca must have a power-of-2 size " 701 "ratio."); 702 assert(Offset == 0 && "Can't extract a value of a smaller vector type " 703 "from a nonzero offset."); 704 705 const Type *ToElementTy = cast<VectorType>(ToType)->getElementType(); 706 const Type *CastElementTy = getScaledElementType(ToElementTy, 707 ToTypeSize * 8); 708 unsigned NumCastVectorElements = AllocaSize / ToTypeSize; 709 710 LLVMContext &Context = FromVal->getContext(); 711 const Type *CastTy = VectorType::get(CastElementTy, 712 NumCastVectorElements); 713 Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp"); 714 Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get( 715 Type::getInt32Ty(Context), 0), "tmp"); 716 return Builder.CreateBitCast(Extract, ToType, "tmp"); 717 } 718 719 // Otherwise it must be an element access. 720 unsigned Elt = 0; 721 if (Offset) { 722 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 723 Elt = Offset/EltSize; 724 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 725 } 726 // Return the element extracted out of it. 727 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 728 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 729 if (V->getType() != ToType) 730 V = Builder.CreateBitCast(V, ToType, "tmp"); 731 return V; 732 } 733 734 // If ToType is a first class aggregate, extract out each of the pieces and 735 // use insertvalue's to form the FCA. 736 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 737 const StructLayout &Layout = *TD.getStructLayout(ST); 738 Value *Res = UndefValue::get(ST); 739 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 740 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 741 Offset+Layout.getElementOffsetInBits(i), 742 Builder); 743 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 744 } 745 return Res; 746 } 747 748 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 749 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 750 Value *Res = UndefValue::get(AT); 751 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 752 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 753 Offset+i*EltSize, Builder); 754 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 755 } 756 return Res; 757 } 758 759 // Otherwise, this must be a union that was converted to an integer value. 760 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 761 762 // If this is a big-endian system and the load is narrower than the 763 // full alloca type, we need to do a shift to get the right bits. 764 int ShAmt = 0; 765 if (TD.isBigEndian()) { 766 // On big-endian machines, the lowest bit is stored at the bit offset 767 // from the pointer given by getTypeStoreSizeInBits. This matters for 768 // integers with a bitwidth that is not a multiple of 8. 769 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 770 TD.getTypeStoreSizeInBits(ToType) - Offset; 771 } else { 772 ShAmt = Offset; 773 } 774 775 // Note: we support negative bitwidths (with shl) which are not defined. 776 // We do this to support (f.e.) loads off the end of a structure where 777 // only some bits are used. 778 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 779 FromVal = Builder.CreateLShr(FromVal, 780 ConstantInt::get(FromVal->getType(), 781 ShAmt), "tmp"); 782 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 783 FromVal = Builder.CreateShl(FromVal, 784 ConstantInt::get(FromVal->getType(), 785 -ShAmt), "tmp"); 786 787 // Finally, unconditionally truncate the integer to the right width. 788 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 789 if (LIBitWidth < NTy->getBitWidth()) 790 FromVal = 791 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 792 LIBitWidth), "tmp"); 793 else if (LIBitWidth > NTy->getBitWidth()) 794 FromVal = 795 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 796 LIBitWidth), "tmp"); 797 798 // If the result is an integer, this is a trunc or bitcast. 799 if (ToType->isIntegerTy()) { 800 // Should be done. 801 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 802 // Just do a bitcast, we know the sizes match up. 803 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 804 } else { 805 // Otherwise must be a pointer. 806 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 807 } 808 assert(FromVal->getType() == ToType && "Didn't convert right?"); 809 return FromVal; 810} 811 812/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 813/// or vector value "Old" at the offset specified by Offset. 814/// 815/// This happens when we are converting an "integer union" to a 816/// single integer scalar, or when we are converting a "vector union" to a 817/// vector with insert/extractelement instructions. 818/// 819/// Offset is an offset from the original alloca, in bits that need to be 820/// shifted to the right. 821Value *ConvertToScalarInfo:: 822ConvertScalar_InsertValue(Value *SV, Value *Old, 823 uint64_t Offset, IRBuilder<> &Builder) { 824 // Convert the stored type to the actual type, shift it left to insert 825 // then 'or' into place. 826 const Type *AllocaType = Old->getType(); 827 LLVMContext &Context = Old->getContext(); 828 829 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 830 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 831 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 832 833 // Changing the whole vector with memset or with an access of a different 834 // vector type? 835 if (ValSize == VecSize) 836 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 837 838 if (SV->getType()->isVectorTy() && isPowerOf2_64(VecSize / ValSize)) { 839 assert(Offset == 0 && "Can't insert a value of a smaller vector type at " 840 "a nonzero offset."); 841 842 const Type *ToElementTy = 843 cast<VectorType>(SV->getType())->getElementType(); 844 const Type *CastElementTy = getScaledElementType(ToElementTy, ValSize); 845 unsigned NumCastVectorElements = VecSize / ValSize; 846 847 LLVMContext &Context = SV->getContext(); 848 const Type *OldCastTy = VectorType::get(CastElementTy, 849 NumCastVectorElements); 850 Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp"); 851 852 Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp"); 853 Value *Insert = 854 Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get( 855 Type::getInt32Ty(Context), 0), "tmp"); 856 return Builder.CreateBitCast(Insert, AllocaType, "tmp"); 857 } 858 859 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 860 861 // Must be an element insertion. 862 unsigned Elt = Offset/EltSize; 863 864 if (SV->getType() != VTy->getElementType()) 865 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 866 867 SV = Builder.CreateInsertElement(Old, SV, 868 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 869 "tmp"); 870 return SV; 871 } 872 873 // If SV is a first-class aggregate value, insert each value recursively. 874 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 875 const StructLayout &Layout = *TD.getStructLayout(ST); 876 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 877 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 878 Old = ConvertScalar_InsertValue(Elt, Old, 879 Offset+Layout.getElementOffsetInBits(i), 880 Builder); 881 } 882 return Old; 883 } 884 885 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 886 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 887 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 888 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 889 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 890 } 891 return Old; 892 } 893 894 // If SV is a float, convert it to the appropriate integer type. 895 // If it is a pointer, do the same. 896 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 897 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 898 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 899 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 900 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 901 SV = Builder.CreateBitCast(SV, 902 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 903 else if (SV->getType()->isPointerTy()) 904 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 905 906 // Zero extend or truncate the value if needed. 907 if (SV->getType() != AllocaType) { 908 if (SV->getType()->getPrimitiveSizeInBits() < 909 AllocaType->getPrimitiveSizeInBits()) 910 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 911 else { 912 // Truncation may be needed if storing more than the alloca can hold 913 // (undefined behavior). 914 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 915 SrcWidth = DestWidth; 916 SrcStoreWidth = DestStoreWidth; 917 } 918 } 919 920 // If this is a big-endian system and the store is narrower than the 921 // full alloca type, we need to do a shift to get the right bits. 922 int ShAmt = 0; 923 if (TD.isBigEndian()) { 924 // On big-endian machines, the lowest bit is stored at the bit offset 925 // from the pointer given by getTypeStoreSizeInBits. This matters for 926 // integers with a bitwidth that is not a multiple of 8. 927 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 928 } else { 929 ShAmt = Offset; 930 } 931 932 // Note: we support negative bitwidths (with shr) which are not defined. 933 // We do this to support (f.e.) stores off the end of a structure where 934 // only some bits in the structure are set. 935 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 936 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 937 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 938 ShAmt), "tmp"); 939 Mask <<= ShAmt; 940 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 941 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 942 -ShAmt), "tmp"); 943 Mask = Mask.lshr(-ShAmt); 944 } 945 946 // Mask out the bits we are about to insert from the old value, and or 947 // in the new bits. 948 if (SrcWidth != DestWidth) { 949 assert(DestWidth > SrcWidth); 950 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 951 SV = Builder.CreateOr(Old, SV, "ins"); 952 } 953 return SV; 954} 955 956 957//===----------------------------------------------------------------------===// 958// SRoA Driver 959//===----------------------------------------------------------------------===// 960 961 962bool SROA::runOnFunction(Function &F) { 963 TD = getAnalysisIfAvailable<TargetData>(); 964 965 bool Changed = performPromotion(F); 966 967 // FIXME: ScalarRepl currently depends on TargetData more than it 968 // theoretically needs to. It should be refactored in order to support 969 // target-independent IR. Until this is done, just skip the actual 970 // scalar-replacement portion of this pass. 971 if (!TD) return Changed; 972 973 while (1) { 974 bool LocalChange = performScalarRepl(F); 975 if (!LocalChange) break; // No need to repromote if no scalarrepl 976 Changed = true; 977 LocalChange = performPromotion(F); 978 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 979 } 980 981 return Changed; 982} 983 984namespace { 985class AllocaPromoter : public LoadAndStorePromoter { 986 AllocaInst *AI; 987public: 988 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S) 989 : LoadAndStorePromoter(Insts, S), AI(0) {} 990 991 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 992 // Remember which alloca we're promoting (for isInstInList). 993 this->AI = AI; 994 LoadAndStorePromoter::run(Insts); 995 AI->eraseFromParent(); 996 } 997 998 virtual bool isInstInList(Instruction *I, 999 const SmallVectorImpl<Instruction*> &Insts) const { 1000 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1001 return LI->getOperand(0) == AI; 1002 return cast<StoreInst>(I)->getPointerOperand() == AI; 1003 } 1004}; 1005} // end anon namespace 1006 1007/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1008/// subsequently loaded can be rewritten to load both input pointers and then 1009/// select between the result, allowing the load of the alloca to be promoted. 1010/// From this: 1011/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1012/// %V = load i32* %P2 1013/// to: 1014/// %V1 = load i32* %Alloca -> will be mem2reg'd 1015/// %V2 = load i32* %Other 1016/// %V = select i1 %cond, i32 %V1, i32 %V2 1017/// 1018/// We can do this to a select if its only uses are loads and if the operand to 1019/// the select can be loaded unconditionally. 1020static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1021 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1022 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1023 1024 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1025 UI != UE; ++UI) { 1026 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1027 if (LI == 0 || LI->isVolatile()) return false; 1028 1029 // Both operands to the select need to be dereferencable, either absolutely 1030 // (e.g. allocas) or at this point because we can see other accesses to it. 1031 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1032 LI->getAlignment(), TD)) 1033 return false; 1034 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1035 LI->getAlignment(), TD)) 1036 return false; 1037 } 1038 1039 return true; 1040} 1041 1042/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1043/// subsequently loaded can be rewritten to load both input pointers in the pred 1044/// blocks and then PHI the results, allowing the load of the alloca to be 1045/// promoted. 1046/// From this: 1047/// %P2 = phi [i32* %Alloca, i32* %Other] 1048/// %V = load i32* %P2 1049/// to: 1050/// %V1 = load i32* %Alloca -> will be mem2reg'd 1051/// ... 1052/// %V2 = load i32* %Other 1053/// ... 1054/// %V = phi [i32 %V1, i32 %V2] 1055/// 1056/// We can do this to a select if its only uses are loads and if the operand to 1057/// the select can be loaded unconditionally. 1058static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1059 // For now, we can only do this promotion if the load is in the same block as 1060 // the PHI, and if there are no stores between the phi and load. 1061 // TODO: Allow recursive phi users. 1062 // TODO: Allow stores. 1063 BasicBlock *BB = PN->getParent(); 1064 unsigned MaxAlign = 0; 1065 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1066 UI != UE; ++UI) { 1067 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1068 if (LI == 0 || LI->isVolatile()) return false; 1069 1070 // For now we only allow loads in the same block as the PHI. This is a 1071 // common case that happens when instcombine merges two loads through a PHI. 1072 if (LI->getParent() != BB) return false; 1073 1074 // Ensure that there are no instructions between the PHI and the load that 1075 // could store. 1076 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1077 if (BBI->mayWriteToMemory()) 1078 return false; 1079 1080 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1081 } 1082 1083 // Okay, we know that we have one or more loads in the same block as the PHI. 1084 // We can transform this if it is safe to push the loads into the predecessor 1085 // blocks. The only thing to watch out for is that we can't put a possibly 1086 // trapping load in the predecessor if it is a critical edge. 1087 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1088 BasicBlock *Pred = PN->getIncomingBlock(i); 1089 1090 // If the predecessor has a single successor, then the edge isn't critical. 1091 if (Pred->getTerminator()->getNumSuccessors() == 1) 1092 continue; 1093 1094 Value *InVal = PN->getIncomingValue(i); 1095 1096 // If the InVal is an invoke in the pred, we can't put a load on the edge. 1097 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 1098 if (II->getParent() == Pred) 1099 return false; 1100 1101 // If this pointer is always safe to load, or if we can prove that there is 1102 // already a load in the block, then we can move the load to the pred block. 1103 if (InVal->isDereferenceablePointer() || 1104 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1105 continue; 1106 1107 return false; 1108 } 1109 1110 return true; 1111} 1112 1113 1114/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1115/// direct (non-volatile) loads and stores to it. If the alloca is close but 1116/// not quite there, this will transform the code to allow promotion. As such, 1117/// it is a non-pure predicate. 1118static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1119 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1120 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1121 1122 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1123 UI != UE; ++UI) { 1124 User *U = *UI; 1125 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1126 if (LI->isVolatile()) 1127 return false; 1128 continue; 1129 } 1130 1131 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1132 if (SI->getOperand(0) == AI || SI->isVolatile()) 1133 return false; // Don't allow a store OF the AI, only INTO the AI. 1134 continue; 1135 } 1136 1137 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1138 // If the condition being selected on is a constant, fold the select, yes 1139 // this does (rarely) happen early on. 1140 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1141 Value *Result = SI->getOperand(1+CI->isZero()); 1142 SI->replaceAllUsesWith(Result); 1143 SI->eraseFromParent(); 1144 1145 // This is very rare and we just scrambled the use list of AI, start 1146 // over completely. 1147 return tryToMakeAllocaBePromotable(AI, TD); 1148 } 1149 1150 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1151 // loads, then we can transform this by rewriting the select. 1152 if (!isSafeSelectToSpeculate(SI, TD)) 1153 return false; 1154 1155 InstsToRewrite.insert(SI); 1156 continue; 1157 } 1158 1159 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1160 if (PN->use_empty()) { // Dead PHIs can be stripped. 1161 InstsToRewrite.insert(PN); 1162 continue; 1163 } 1164 1165 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1166 // in the pred blocks, then we can transform this by rewriting the PHI. 1167 if (!isSafePHIToSpeculate(PN, TD)) 1168 return false; 1169 1170 InstsToRewrite.insert(PN); 1171 continue; 1172 } 1173 1174 return false; 1175 } 1176 1177 // If there are no instructions to rewrite, then all uses are load/stores and 1178 // we're done! 1179 if (InstsToRewrite.empty()) 1180 return true; 1181 1182 // If we have instructions that need to be rewritten for this to be promotable 1183 // take care of it now. 1184 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1185 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1186 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1187 // loads with a new select. 1188 while (!SI->use_empty()) { 1189 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1190 1191 IRBuilder<> Builder(LI); 1192 LoadInst *TrueLoad = 1193 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1194 LoadInst *FalseLoad = 1195 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".t"); 1196 1197 // Transfer alignment and TBAA info if present. 1198 TrueLoad->setAlignment(LI->getAlignment()); 1199 FalseLoad->setAlignment(LI->getAlignment()); 1200 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1201 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1202 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1203 } 1204 1205 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1206 V->takeName(LI); 1207 LI->replaceAllUsesWith(V); 1208 LI->eraseFromParent(); 1209 } 1210 1211 // Now that all the loads are gone, the select is gone too. 1212 SI->eraseFromParent(); 1213 continue; 1214 } 1215 1216 // Otherwise, we have a PHI node which allows us to push the loads into the 1217 // predecessors. 1218 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1219 if (PN->use_empty()) { 1220 PN->eraseFromParent(); 1221 continue; 1222 } 1223 1224 const Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1225 PHINode *NewPN = PHINode::Create(LoadTy, PN->getName()+".ld", PN); 1226 1227 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1228 // matter which one we get and if any differ, it doesn't matter. 1229 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1230 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1231 unsigned Align = SomeLoad->getAlignment(); 1232 1233 // Rewrite all loads of the PN to use the new PHI. 1234 while (!PN->use_empty()) { 1235 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1236 LI->replaceAllUsesWith(NewPN); 1237 LI->eraseFromParent(); 1238 } 1239 1240 // Inject loads into all of the pred blocks. Keep track of which blocks we 1241 // insert them into in case we have multiple edges from the same block. 1242 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1243 1244 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1245 BasicBlock *Pred = PN->getIncomingBlock(i); 1246 LoadInst *&Load = InsertedLoads[Pred]; 1247 if (Load == 0) { 1248 Load = new LoadInst(PN->getIncomingValue(i), 1249 PN->getName() + "." + Pred->getName(), 1250 Pred->getTerminator()); 1251 Load->setAlignment(Align); 1252 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1253 } 1254 1255 NewPN->addIncoming(Load, Pred); 1256 } 1257 1258 PN->eraseFromParent(); 1259 } 1260 1261 ++NumAdjusted; 1262 return true; 1263} 1264 1265 1266bool SROA::performPromotion(Function &F) { 1267 std::vector<AllocaInst*> Allocas; 1268 DominatorTree *DT = 0; 1269 if (HasDomTree) 1270 DT = &getAnalysis<DominatorTree>(); 1271 1272 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1273 1274 bool Changed = false; 1275 SmallVector<Instruction*, 64> Insts; 1276 while (1) { 1277 Allocas.clear(); 1278 1279 // Find allocas that are safe to promote, by looking at all instructions in 1280 // the entry node 1281 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1282 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1283 if (tryToMakeAllocaBePromotable(AI, TD)) 1284 Allocas.push_back(AI); 1285 1286 if (Allocas.empty()) break; 1287 1288 if (HasDomTree) 1289 PromoteMemToReg(Allocas, *DT); 1290 else { 1291 SSAUpdater SSA; 1292 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1293 AllocaInst *AI = Allocas[i]; 1294 1295 // Build list of instructions to promote. 1296 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1297 UI != E; ++UI) 1298 Insts.push_back(cast<Instruction>(*UI)); 1299 1300 AllocaPromoter(Insts, SSA).run(AI, Insts); 1301 Insts.clear(); 1302 } 1303 } 1304 NumPromoted += Allocas.size(); 1305 Changed = true; 1306 } 1307 1308 return Changed; 1309} 1310 1311 1312/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1313/// SROA. It must be a struct or array type with a small number of elements. 1314static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1315 const Type *T = AI->getAllocatedType(); 1316 // Do not promote any struct into more than 32 separate vars. 1317 if (const StructType *ST = dyn_cast<StructType>(T)) 1318 return ST->getNumElements() <= 32; 1319 // Arrays are much less likely to be safe for SROA; only consider 1320 // them if they are very small. 1321 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 1322 return AT->getNumElements() <= 8; 1323 return false; 1324} 1325 1326 1327// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1328// which runs on all of the malloc/alloca instructions in the function, removing 1329// them if they are only used by getelementptr instructions. 1330// 1331bool SROA::performScalarRepl(Function &F) { 1332 std::vector<AllocaInst*> WorkList; 1333 1334 // Scan the entry basic block, adding allocas to the worklist. 1335 BasicBlock &BB = F.getEntryBlock(); 1336 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1337 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1338 WorkList.push_back(A); 1339 1340 // Process the worklist 1341 bool Changed = false; 1342 while (!WorkList.empty()) { 1343 AllocaInst *AI = WorkList.back(); 1344 WorkList.pop_back(); 1345 1346 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1347 // with unused elements. 1348 if (AI->use_empty()) { 1349 AI->eraseFromParent(); 1350 Changed = true; 1351 continue; 1352 } 1353 1354 // If this alloca is impossible for us to promote, reject it early. 1355 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1356 continue; 1357 1358 // Check to see if this allocation is only modified by a memcpy/memmove from 1359 // a constant global. If this is the case, we can change all users to use 1360 // the constant global instead. This is commonly produced by the CFE by 1361 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1362 // is only subsequently read. 1363 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 1364 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1365 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 1366 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 1367 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1368 TheCopy->eraseFromParent(); // Don't mutate the global. 1369 AI->eraseFromParent(); 1370 ++NumGlobals; 1371 Changed = true; 1372 continue; 1373 } 1374 1375 // Check to see if we can perform the core SROA transformation. We cannot 1376 // transform the allocation instruction if it is an array allocation 1377 // (allocations OF arrays are ok though), and an allocation of a scalar 1378 // value cannot be decomposed at all. 1379 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1380 1381 // Do not promote [0 x %struct]. 1382 if (AllocaSize == 0) continue; 1383 1384 // Do not promote any struct whose size is too big. 1385 if (AllocaSize > SRThreshold) continue; 1386 1387 // If the alloca looks like a good candidate for scalar replacement, and if 1388 // all its users can be transformed, then split up the aggregate into its 1389 // separate elements. 1390 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1391 DoScalarReplacement(AI, WorkList); 1392 Changed = true; 1393 continue; 1394 } 1395 1396 // If we can turn this aggregate value (potentially with casts) into a 1397 // simple scalar value that can be mem2reg'd into a register value. 1398 // IsNotTrivial tracks whether this is something that mem2reg could have 1399 // promoted itself. If so, we don't want to transform it needlessly. Note 1400 // that we can't just check based on the type: the alloca may be of an i32 1401 // but that has pointer arithmetic to set byte 3 of it or something. 1402 if (AllocaInst *NewAI = 1403 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1404 NewAI->takeName(AI); 1405 AI->eraseFromParent(); 1406 ++NumConverted; 1407 Changed = true; 1408 continue; 1409 } 1410 1411 // Otherwise, couldn't process this alloca. 1412 } 1413 1414 return Changed; 1415} 1416 1417/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1418/// predicate, do SROA now. 1419void SROA::DoScalarReplacement(AllocaInst *AI, 1420 std::vector<AllocaInst*> &WorkList) { 1421 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1422 SmallVector<AllocaInst*, 32> ElementAllocas; 1423 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1424 ElementAllocas.reserve(ST->getNumContainedTypes()); 1425 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1426 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1427 AI->getAlignment(), 1428 AI->getName() + "." + Twine(i), AI); 1429 ElementAllocas.push_back(NA); 1430 WorkList.push_back(NA); // Add to worklist for recursive processing 1431 } 1432 } else { 1433 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1434 ElementAllocas.reserve(AT->getNumElements()); 1435 const Type *ElTy = AT->getElementType(); 1436 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1437 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1438 AI->getName() + "." + Twine(i), AI); 1439 ElementAllocas.push_back(NA); 1440 WorkList.push_back(NA); // Add to worklist for recursive processing 1441 } 1442 } 1443 1444 // Now that we have created the new alloca instructions, rewrite all the 1445 // uses of the old alloca. 1446 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1447 1448 // Now erase any instructions that were made dead while rewriting the alloca. 1449 DeleteDeadInstructions(); 1450 AI->eraseFromParent(); 1451 1452 ++NumReplaced; 1453} 1454 1455/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1456/// recursively including all their operands that become trivially dead. 1457void SROA::DeleteDeadInstructions() { 1458 while (!DeadInsts.empty()) { 1459 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1460 1461 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1462 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1463 // Zero out the operand and see if it becomes trivially dead. 1464 // (But, don't add allocas to the dead instruction list -- they are 1465 // already on the worklist and will be deleted separately.) 1466 *OI = 0; 1467 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1468 DeadInsts.push_back(U); 1469 } 1470 1471 I->eraseFromParent(); 1472 } 1473} 1474 1475/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1476/// performing scalar replacement of alloca AI. The results are flagged in 1477/// the Info parameter. Offset indicates the position within AI that is 1478/// referenced by this instruction. 1479void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1480 AllocaInfo &Info) { 1481 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1482 Instruction *User = cast<Instruction>(*UI); 1483 1484 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1485 isSafeForScalarRepl(BC, Offset, Info); 1486 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1487 uint64_t GEPOffset = Offset; 1488 isSafeGEP(GEPI, GEPOffset, Info); 1489 if (!Info.isUnsafe) 1490 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1491 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1492 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1493 if (Length == 0) 1494 return MarkUnsafe(Info, User); 1495 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1496 UI.getOperandNo() == 0, Info, MI, 1497 true /*AllowWholeAccess*/); 1498 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1499 if (LI->isVolatile()) 1500 return MarkUnsafe(Info, User); 1501 const Type *LIType = LI->getType(); 1502 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1503 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1504 Info.hasALoadOrStore = true; 1505 1506 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1507 // Store is ok if storing INTO the pointer, not storing the pointer 1508 if (SI->isVolatile() || SI->getOperand(0) == I) 1509 return MarkUnsafe(Info, User); 1510 1511 const Type *SIType = SI->getOperand(0)->getType(); 1512 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1513 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1514 Info.hasALoadOrStore = true; 1515 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1516 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1517 } else { 1518 return MarkUnsafe(Info, User); 1519 } 1520 if (Info.isUnsafe) return; 1521 } 1522} 1523 1524 1525/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1526/// derived from the alloca, we can often still split the alloca into elements. 1527/// This is useful if we have a large alloca where one element is phi'd 1528/// together somewhere: we can SRoA and promote all the other elements even if 1529/// we end up not being able to promote this one. 1530/// 1531/// All we require is that the uses of the PHI do not index into other parts of 1532/// the alloca. The most important use case for this is single load and stores 1533/// that are PHI'd together, which can happen due to code sinking. 1534void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1535 AllocaInfo &Info) { 1536 // If we've already checked this PHI, don't do it again. 1537 if (PHINode *PN = dyn_cast<PHINode>(I)) 1538 if (!Info.CheckedPHIs.insert(PN)) 1539 return; 1540 1541 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1542 Instruction *User = cast<Instruction>(*UI); 1543 1544 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1545 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1546 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1547 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1548 // but would have to prove that we're staying inside of an element being 1549 // promoted. 1550 if (!GEPI->hasAllZeroIndices()) 1551 return MarkUnsafe(Info, User); 1552 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1553 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1554 if (LI->isVolatile()) 1555 return MarkUnsafe(Info, User); 1556 const Type *LIType = LI->getType(); 1557 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1558 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1559 Info.hasALoadOrStore = true; 1560 1561 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1562 // Store is ok if storing INTO the pointer, not storing the pointer 1563 if (SI->isVolatile() || SI->getOperand(0) == I) 1564 return MarkUnsafe(Info, User); 1565 1566 const Type *SIType = SI->getOperand(0)->getType(); 1567 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1568 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1569 Info.hasALoadOrStore = true; 1570 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1571 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1572 } else { 1573 return MarkUnsafe(Info, User); 1574 } 1575 if (Info.isUnsafe) return; 1576 } 1577} 1578 1579/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1580/// replacement. It is safe when all the indices are constant, in-bounds 1581/// references, and when the resulting offset corresponds to an element within 1582/// the alloca type. The results are flagged in the Info parameter. Upon 1583/// return, Offset is adjusted as specified by the GEP indices. 1584void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1585 uint64_t &Offset, AllocaInfo &Info) { 1586 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1587 if (GEPIt == E) 1588 return; 1589 1590 // Walk through the GEP type indices, checking the types that this indexes 1591 // into. 1592 for (; GEPIt != E; ++GEPIt) { 1593 // Ignore struct elements, no extra checking needed for these. 1594 if ((*GEPIt)->isStructTy()) 1595 continue; 1596 1597 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1598 if (!IdxVal) 1599 return MarkUnsafe(Info, GEPI); 1600 } 1601 1602 // Compute the offset due to this GEP and check if the alloca has a 1603 // component element at that offset. 1604 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1605 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1606 &Indices[0], Indices.size()); 1607 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1608 MarkUnsafe(Info, GEPI); 1609} 1610 1611/// isHomogeneousAggregate - Check if type T is a struct or array containing 1612/// elements of the same type (which is always true for arrays). If so, 1613/// return true with NumElts and EltTy set to the number of elements and the 1614/// element type, respectively. 1615static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1616 const Type *&EltTy) { 1617 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1618 NumElts = AT->getNumElements(); 1619 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1620 return true; 1621 } 1622 if (const StructType *ST = dyn_cast<StructType>(T)) { 1623 NumElts = ST->getNumContainedTypes(); 1624 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1625 for (unsigned n = 1; n < NumElts; ++n) { 1626 if (ST->getContainedType(n) != EltTy) 1627 return false; 1628 } 1629 return true; 1630 } 1631 return false; 1632} 1633 1634/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1635/// "homogeneous" aggregates with the same element type and number of elements. 1636static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1637 if (T1 == T2) 1638 return true; 1639 1640 unsigned NumElts1, NumElts2; 1641 const Type *EltTy1, *EltTy2; 1642 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1643 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1644 NumElts1 == NumElts2 && 1645 EltTy1 == EltTy2) 1646 return true; 1647 1648 return false; 1649} 1650 1651/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1652/// alloca or has an offset and size that corresponds to a component element 1653/// within it. The offset checked here may have been formed from a GEP with a 1654/// pointer bitcasted to a different type. 1655/// 1656/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1657/// unit. If false, it only allows accesses known to be in a single element. 1658void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1659 const Type *MemOpType, bool isStore, 1660 AllocaInfo &Info, Instruction *TheAccess, 1661 bool AllowWholeAccess) { 1662 // Check if this is a load/store of the entire alloca. 1663 if (Offset == 0 && AllowWholeAccess && 1664 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1665 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1666 // loads/stores (which are essentially the same as the MemIntrinsics with 1667 // regard to copying padding between elements). But, if an alloca is 1668 // flagged as both a source and destination of such operations, we'll need 1669 // to check later for padding between elements. 1670 if (!MemOpType || MemOpType->isIntegerTy()) { 1671 if (isStore) 1672 Info.isMemCpyDst = true; 1673 else 1674 Info.isMemCpySrc = true; 1675 return; 1676 } 1677 // This is also safe for references using a type that is compatible with 1678 // the type of the alloca, so that loads/stores can be rewritten using 1679 // insertvalue/extractvalue. 1680 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1681 Info.hasSubelementAccess = true; 1682 return; 1683 } 1684 } 1685 // Check if the offset/size correspond to a component within the alloca type. 1686 const Type *T = Info.AI->getAllocatedType(); 1687 if (TypeHasComponent(T, Offset, MemSize)) { 1688 Info.hasSubelementAccess = true; 1689 return; 1690 } 1691 1692 return MarkUnsafe(Info, TheAccess); 1693} 1694 1695/// TypeHasComponent - Return true if T has a component type with the 1696/// specified offset and size. If Size is zero, do not check the size. 1697bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1698 const Type *EltTy; 1699 uint64_t EltSize; 1700 if (const StructType *ST = dyn_cast<StructType>(T)) { 1701 const StructLayout *Layout = TD->getStructLayout(ST); 1702 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1703 EltTy = ST->getContainedType(EltIdx); 1704 EltSize = TD->getTypeAllocSize(EltTy); 1705 Offset -= Layout->getElementOffset(EltIdx); 1706 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1707 EltTy = AT->getElementType(); 1708 EltSize = TD->getTypeAllocSize(EltTy); 1709 if (Offset >= AT->getNumElements() * EltSize) 1710 return false; 1711 Offset %= EltSize; 1712 } else { 1713 return false; 1714 } 1715 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1716 return true; 1717 // Check if the component spans multiple elements. 1718 if (Offset + Size > EltSize) 1719 return false; 1720 return TypeHasComponent(EltTy, Offset, Size); 1721} 1722 1723/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1724/// the instruction I, which references it, to use the separate elements. 1725/// Offset indicates the position within AI that is referenced by this 1726/// instruction. 1727void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1728 SmallVector<AllocaInst*, 32> &NewElts) { 1729 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1730 Use &TheUse = UI.getUse(); 1731 Instruction *User = cast<Instruction>(*UI++); 1732 1733 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1734 RewriteBitCast(BC, AI, Offset, NewElts); 1735 continue; 1736 } 1737 1738 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1739 RewriteGEP(GEPI, AI, Offset, NewElts); 1740 continue; 1741 } 1742 1743 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1744 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1745 uint64_t MemSize = Length->getZExtValue(); 1746 if (Offset == 0 && 1747 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1748 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1749 // Otherwise the intrinsic can only touch a single element and the 1750 // address operand will be updated, so nothing else needs to be done. 1751 continue; 1752 } 1753 1754 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1755 const Type *LIType = LI->getType(); 1756 1757 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1758 // Replace: 1759 // %res = load { i32, i32 }* %alloc 1760 // with: 1761 // %load.0 = load i32* %alloc.0 1762 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1763 // %load.1 = load i32* %alloc.1 1764 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1765 // (Also works for arrays instead of structs) 1766 Value *Insert = UndefValue::get(LIType); 1767 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1768 Value *Load = new LoadInst(NewElts[i], "load", LI); 1769 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1770 } 1771 LI->replaceAllUsesWith(Insert); 1772 DeadInsts.push_back(LI); 1773 } else if (LIType->isIntegerTy() && 1774 TD->getTypeAllocSize(LIType) == 1775 TD->getTypeAllocSize(AI->getAllocatedType())) { 1776 // If this is a load of the entire alloca to an integer, rewrite it. 1777 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1778 } 1779 continue; 1780 } 1781 1782 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1783 Value *Val = SI->getOperand(0); 1784 const Type *SIType = Val->getType(); 1785 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1786 // Replace: 1787 // store { i32, i32 } %val, { i32, i32 }* %alloc 1788 // with: 1789 // %val.0 = extractvalue { i32, i32 } %val, 0 1790 // store i32 %val.0, i32* %alloc.0 1791 // %val.1 = extractvalue { i32, i32 } %val, 1 1792 // store i32 %val.1, i32* %alloc.1 1793 // (Also works for arrays instead of structs) 1794 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1795 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1796 new StoreInst(Extract, NewElts[i], SI); 1797 } 1798 DeadInsts.push_back(SI); 1799 } else if (SIType->isIntegerTy() && 1800 TD->getTypeAllocSize(SIType) == 1801 TD->getTypeAllocSize(AI->getAllocatedType())) { 1802 // If this is a store of the entire alloca from an integer, rewrite it. 1803 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1804 } 1805 continue; 1806 } 1807 1808 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1809 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1810 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1811 // the new pointer. 1812 if (!isa<AllocaInst>(I)) continue; 1813 1814 assert(Offset == 0 && NewElts[0] && 1815 "Direct alloca use should have a zero offset"); 1816 1817 // If we have a use of the alloca, we know the derived uses will be 1818 // utilizing just the first element of the scalarized result. Insert a 1819 // bitcast of the first alloca before the user as required. 1820 AllocaInst *NewAI = NewElts[0]; 1821 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1822 NewAI->moveBefore(BCI); 1823 TheUse = BCI; 1824 continue; 1825 } 1826 } 1827} 1828 1829/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1830/// and recursively continue updating all of its uses. 1831void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1832 SmallVector<AllocaInst*, 32> &NewElts) { 1833 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1834 if (BC->getOperand(0) != AI) 1835 return; 1836 1837 // The bitcast references the original alloca. Replace its uses with 1838 // references to the first new element alloca. 1839 Instruction *Val = NewElts[0]; 1840 if (Val->getType() != BC->getDestTy()) { 1841 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1842 Val->takeName(BC); 1843 } 1844 BC->replaceAllUsesWith(Val); 1845 DeadInsts.push_back(BC); 1846} 1847 1848/// FindElementAndOffset - Return the index of the element containing Offset 1849/// within the specified type, which must be either a struct or an array. 1850/// Sets T to the type of the element and Offset to the offset within that 1851/// element. IdxTy is set to the type of the index result to be used in a 1852/// GEP instruction. 1853uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1854 const Type *&IdxTy) { 1855 uint64_t Idx = 0; 1856 if (const StructType *ST = dyn_cast<StructType>(T)) { 1857 const StructLayout *Layout = TD->getStructLayout(ST); 1858 Idx = Layout->getElementContainingOffset(Offset); 1859 T = ST->getContainedType(Idx); 1860 Offset -= Layout->getElementOffset(Idx); 1861 IdxTy = Type::getInt32Ty(T->getContext()); 1862 return Idx; 1863 } 1864 const ArrayType *AT = cast<ArrayType>(T); 1865 T = AT->getElementType(); 1866 uint64_t EltSize = TD->getTypeAllocSize(T); 1867 Idx = Offset / EltSize; 1868 Offset -= Idx * EltSize; 1869 IdxTy = Type::getInt64Ty(T->getContext()); 1870 return Idx; 1871} 1872 1873/// RewriteGEP - Check if this GEP instruction moves the pointer across 1874/// elements of the alloca that are being split apart, and if so, rewrite 1875/// the GEP to be relative to the new element. 1876void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1877 SmallVector<AllocaInst*, 32> &NewElts) { 1878 uint64_t OldOffset = Offset; 1879 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1880 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1881 &Indices[0], Indices.size()); 1882 1883 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1884 1885 const Type *T = AI->getAllocatedType(); 1886 const Type *IdxTy; 1887 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1888 if (GEPI->getOperand(0) == AI) 1889 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1890 1891 T = AI->getAllocatedType(); 1892 uint64_t EltOffset = Offset; 1893 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1894 1895 // If this GEP does not move the pointer across elements of the alloca 1896 // being split, then it does not needs to be rewritten. 1897 if (Idx == OldIdx) 1898 return; 1899 1900 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1901 SmallVector<Value*, 8> NewArgs; 1902 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1903 while (EltOffset != 0) { 1904 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1905 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1906 } 1907 Instruction *Val = NewElts[Idx]; 1908 if (NewArgs.size() > 1) { 1909 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1910 NewArgs.end(), "", GEPI); 1911 Val->takeName(GEPI); 1912 } 1913 if (Val->getType() != GEPI->getType()) 1914 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1915 GEPI->replaceAllUsesWith(Val); 1916 DeadInsts.push_back(GEPI); 1917} 1918 1919/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1920/// Rewrite it to copy or set the elements of the scalarized memory. 1921void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1922 AllocaInst *AI, 1923 SmallVector<AllocaInst*, 32> &NewElts) { 1924 // If this is a memcpy/memmove, construct the other pointer as the 1925 // appropriate type. The "Other" pointer is the pointer that goes to memory 1926 // that doesn't have anything to do with the alloca that we are promoting. For 1927 // memset, this Value* stays null. 1928 Value *OtherPtr = 0; 1929 unsigned MemAlignment = MI->getAlignment(); 1930 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1931 if (Inst == MTI->getRawDest()) 1932 OtherPtr = MTI->getRawSource(); 1933 else { 1934 assert(Inst == MTI->getRawSource()); 1935 OtherPtr = MTI->getRawDest(); 1936 } 1937 } 1938 1939 // If there is an other pointer, we want to convert it to the same pointer 1940 // type as AI has, so we can GEP through it safely. 1941 if (OtherPtr) { 1942 unsigned AddrSpace = 1943 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 1944 1945 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1946 // optimization, but it's also required to detect the corner case where 1947 // both pointer operands are referencing the same memory, and where 1948 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1949 // function is only called for mem intrinsics that access the whole 1950 // aggregate, so non-zero GEPs are not an issue here.) 1951 OtherPtr = OtherPtr->stripPointerCasts(); 1952 1953 // Copying the alloca to itself is a no-op: just delete it. 1954 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1955 // This code will run twice for a no-op memcpy -- once for each operand. 1956 // Put only one reference to MI on the DeadInsts list. 1957 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1958 E = DeadInsts.end(); I != E; ++I) 1959 if (*I == MI) return; 1960 DeadInsts.push_back(MI); 1961 return; 1962 } 1963 1964 // If the pointer is not the right type, insert a bitcast to the right 1965 // type. 1966 const Type *NewTy = 1967 PointerType::get(AI->getType()->getElementType(), AddrSpace); 1968 1969 if (OtherPtr->getType() != NewTy) 1970 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 1971 } 1972 1973 // Process each element of the aggregate. 1974 bool SROADest = MI->getRawDest() == Inst; 1975 1976 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1977 1978 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1979 // If this is a memcpy/memmove, emit a GEP of the other element address. 1980 Value *OtherElt = 0; 1981 unsigned OtherEltAlign = MemAlignment; 1982 1983 if (OtherPtr) { 1984 Value *Idx[2] = { Zero, 1985 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1986 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1987 OtherPtr->getName()+"."+Twine(i), 1988 MI); 1989 uint64_t EltOffset; 1990 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1991 const Type *OtherTy = OtherPtrTy->getElementType(); 1992 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1993 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1994 } else { 1995 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 1996 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1997 } 1998 1999 // The alignment of the other pointer is the guaranteed alignment of the 2000 // element, which is affected by both the known alignment of the whole 2001 // mem intrinsic and the alignment of the element. If the alignment of 2002 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2003 // known alignment is just 4 bytes. 2004 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2005 } 2006 2007 Value *EltPtr = NewElts[i]; 2008 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2009 2010 // If we got down to a scalar, insert a load or store as appropriate. 2011 if (EltTy->isSingleValueType()) { 2012 if (isa<MemTransferInst>(MI)) { 2013 if (SROADest) { 2014 // From Other to Alloca. 2015 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2016 new StoreInst(Elt, EltPtr, MI); 2017 } else { 2018 // From Alloca to Other. 2019 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2020 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2021 } 2022 continue; 2023 } 2024 assert(isa<MemSetInst>(MI)); 2025 2026 // If the stored element is zero (common case), just store a null 2027 // constant. 2028 Constant *StoreVal; 2029 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2030 if (CI->isZero()) { 2031 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2032 } else { 2033 // If EltTy is a vector type, get the element type. 2034 const Type *ValTy = EltTy->getScalarType(); 2035 2036 // Construct an integer with the right value. 2037 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2038 APInt OneVal(EltSize, CI->getZExtValue()); 2039 APInt TotalVal(OneVal); 2040 // Set each byte. 2041 for (unsigned i = 0; 8*i < EltSize; ++i) { 2042 TotalVal = TotalVal.shl(8); 2043 TotalVal |= OneVal; 2044 } 2045 2046 // Convert the integer value to the appropriate type. 2047 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2048 if (ValTy->isPointerTy()) 2049 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2050 else if (ValTy->isFloatingPointTy()) 2051 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2052 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2053 2054 // If the requested value was a vector constant, create it. 2055 if (EltTy != ValTy) { 2056 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 2057 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2058 StoreVal = ConstantVector::get(Elts); 2059 } 2060 } 2061 new StoreInst(StoreVal, EltPtr, MI); 2062 continue; 2063 } 2064 // Otherwise, if we're storing a byte variable, use a memset call for 2065 // this element. 2066 } 2067 2068 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2069 2070 IRBuilder<> Builder(MI); 2071 2072 // Finally, insert the meminst for this element. 2073 if (isa<MemSetInst>(MI)) { 2074 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2075 MI->isVolatile()); 2076 } else { 2077 assert(isa<MemTransferInst>(MI)); 2078 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2079 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2080 2081 if (isa<MemCpyInst>(MI)) 2082 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2083 else 2084 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2085 } 2086 } 2087 DeadInsts.push_back(MI); 2088} 2089 2090/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2091/// overwrites the entire allocation. Extract out the pieces of the stored 2092/// integer and store them individually. 2093void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2094 SmallVector<AllocaInst*, 32> &NewElts){ 2095 // Extract each element out of the integer according to its structure offset 2096 // and store the element value to the individual alloca. 2097 Value *SrcVal = SI->getOperand(0); 2098 const Type *AllocaEltTy = AI->getAllocatedType(); 2099 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2100 2101 IRBuilder<> Builder(SI); 2102 2103 // Handle tail padding by extending the operand 2104 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2105 SrcVal = Builder.CreateZExt(SrcVal, 2106 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2107 2108 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2109 << '\n'); 2110 2111 // There are two forms here: AI could be an array or struct. Both cases 2112 // have different ways to compute the element offset. 2113 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2114 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2115 2116 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2117 // Get the number of bits to shift SrcVal to get the value. 2118 const Type *FieldTy = EltSTy->getElementType(i); 2119 uint64_t Shift = Layout->getElementOffsetInBits(i); 2120 2121 if (TD->isBigEndian()) 2122 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2123 2124 Value *EltVal = SrcVal; 2125 if (Shift) { 2126 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2127 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2128 } 2129 2130 // Truncate down to an integer of the right size. 2131 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2132 2133 // Ignore zero sized fields like {}, they obviously contain no data. 2134 if (FieldSizeBits == 0) continue; 2135 2136 if (FieldSizeBits != AllocaSizeBits) 2137 EltVal = Builder.CreateTrunc(EltVal, 2138 IntegerType::get(SI->getContext(), FieldSizeBits)); 2139 Value *DestField = NewElts[i]; 2140 if (EltVal->getType() == FieldTy) { 2141 // Storing to an integer field of this size, just do it. 2142 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2143 // Bitcast to the right element type (for fp/vector values). 2144 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2145 } else { 2146 // Otherwise, bitcast the dest pointer (for aggregates). 2147 DestField = Builder.CreateBitCast(DestField, 2148 PointerType::getUnqual(EltVal->getType())); 2149 } 2150 new StoreInst(EltVal, DestField, SI); 2151 } 2152 2153 } else { 2154 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2155 const Type *ArrayEltTy = ATy->getElementType(); 2156 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2157 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2158 2159 uint64_t Shift; 2160 2161 if (TD->isBigEndian()) 2162 Shift = AllocaSizeBits-ElementOffset; 2163 else 2164 Shift = 0; 2165 2166 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2167 // Ignore zero sized fields like {}, they obviously contain no data. 2168 if (ElementSizeBits == 0) continue; 2169 2170 Value *EltVal = SrcVal; 2171 if (Shift) { 2172 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2173 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2174 } 2175 2176 // Truncate down to an integer of the right size. 2177 if (ElementSizeBits != AllocaSizeBits) 2178 EltVal = Builder.CreateTrunc(EltVal, 2179 IntegerType::get(SI->getContext(), 2180 ElementSizeBits)); 2181 Value *DestField = NewElts[i]; 2182 if (EltVal->getType() == ArrayEltTy) { 2183 // Storing to an integer field of this size, just do it. 2184 } else if (ArrayEltTy->isFloatingPointTy() || 2185 ArrayEltTy->isVectorTy()) { 2186 // Bitcast to the right element type (for fp/vector values). 2187 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2188 } else { 2189 // Otherwise, bitcast the dest pointer (for aggregates). 2190 DestField = Builder.CreateBitCast(DestField, 2191 PointerType::getUnqual(EltVal->getType())); 2192 } 2193 new StoreInst(EltVal, DestField, SI); 2194 2195 if (TD->isBigEndian()) 2196 Shift -= ElementOffset; 2197 else 2198 Shift += ElementOffset; 2199 } 2200 } 2201 2202 DeadInsts.push_back(SI); 2203} 2204 2205/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2206/// an integer. Load the individual pieces to form the aggregate value. 2207void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2208 SmallVector<AllocaInst*, 32> &NewElts) { 2209 // Extract each element out of the NewElts according to its structure offset 2210 // and form the result value. 2211 const Type *AllocaEltTy = AI->getAllocatedType(); 2212 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2213 2214 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2215 << '\n'); 2216 2217 // There are two forms here: AI could be an array or struct. Both cases 2218 // have different ways to compute the element offset. 2219 const StructLayout *Layout = 0; 2220 uint64_t ArrayEltBitOffset = 0; 2221 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2222 Layout = TD->getStructLayout(EltSTy); 2223 } else { 2224 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2225 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2226 } 2227 2228 Value *ResultVal = 2229 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2230 2231 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2232 // Load the value from the alloca. If the NewElt is an aggregate, cast 2233 // the pointer to an integer of the same size before doing the load. 2234 Value *SrcField = NewElts[i]; 2235 const Type *FieldTy = 2236 cast<PointerType>(SrcField->getType())->getElementType(); 2237 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2238 2239 // Ignore zero sized fields like {}, they obviously contain no data. 2240 if (FieldSizeBits == 0) continue; 2241 2242 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2243 FieldSizeBits); 2244 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2245 !FieldTy->isVectorTy()) 2246 SrcField = new BitCastInst(SrcField, 2247 PointerType::getUnqual(FieldIntTy), 2248 "", LI); 2249 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2250 2251 // If SrcField is a fp or vector of the right size but that isn't an 2252 // integer type, bitcast to an integer so we can shift it. 2253 if (SrcField->getType() != FieldIntTy) 2254 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2255 2256 // Zero extend the field to be the same size as the final alloca so that 2257 // we can shift and insert it. 2258 if (SrcField->getType() != ResultVal->getType()) 2259 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2260 2261 // Determine the number of bits to shift SrcField. 2262 uint64_t Shift; 2263 if (Layout) // Struct case. 2264 Shift = Layout->getElementOffsetInBits(i); 2265 else // Array case. 2266 Shift = i*ArrayEltBitOffset; 2267 2268 if (TD->isBigEndian()) 2269 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2270 2271 if (Shift) { 2272 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2273 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2274 } 2275 2276 // Don't create an 'or x, 0' on the first iteration. 2277 if (!isa<Constant>(ResultVal) || 2278 !cast<Constant>(ResultVal)->isNullValue()) 2279 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2280 else 2281 ResultVal = SrcField; 2282 } 2283 2284 // Handle tail padding by truncating the result 2285 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2286 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2287 2288 LI->replaceAllUsesWith(ResultVal); 2289 DeadInsts.push_back(LI); 2290} 2291 2292/// HasPadding - Return true if the specified type has any structure or 2293/// alignment padding in between the elements that would be split apart 2294/// by SROA; return false otherwise. 2295static bool HasPadding(const Type *Ty, const TargetData &TD) { 2296 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2297 Ty = ATy->getElementType(); 2298 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2299 } 2300 2301 // SROA currently handles only Arrays and Structs. 2302 const StructType *STy = cast<StructType>(Ty); 2303 const StructLayout *SL = TD.getStructLayout(STy); 2304 unsigned PrevFieldBitOffset = 0; 2305 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2306 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2307 2308 // Check to see if there is any padding between this element and the 2309 // previous one. 2310 if (i) { 2311 unsigned PrevFieldEnd = 2312 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2313 if (PrevFieldEnd < FieldBitOffset) 2314 return true; 2315 } 2316 PrevFieldBitOffset = FieldBitOffset; 2317 } 2318 // Check for tail padding. 2319 if (unsigned EltCount = STy->getNumElements()) { 2320 unsigned PrevFieldEnd = PrevFieldBitOffset + 2321 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2322 if (PrevFieldEnd < SL->getSizeInBits()) 2323 return true; 2324 } 2325 return false; 2326} 2327 2328/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2329/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2330/// or 1 if safe after canonicalization has been performed. 2331bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2332 // Loop over the use list of the alloca. We can only transform it if all of 2333 // the users are safe to transform. 2334 AllocaInfo Info(AI); 2335 2336 isSafeForScalarRepl(AI, 0, Info); 2337 if (Info.isUnsafe) { 2338 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2339 return false; 2340 } 2341 2342 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2343 // source and destination, we have to be careful. In particular, the memcpy 2344 // could be moving around elements that live in structure padding of the LLVM 2345 // types, but may actually be used. In these cases, we refuse to promote the 2346 // struct. 2347 if (Info.isMemCpySrc && Info.isMemCpyDst && 2348 HasPadding(AI->getAllocatedType(), *TD)) 2349 return false; 2350 2351 // If the alloca never has an access to just *part* of it, but is accessed 2352 // via loads and stores, then we should use ConvertToScalarInfo to promote 2353 // the alloca instead of promoting each piece at a time and inserting fission 2354 // and fusion code. 2355 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2356 // If the struct/array just has one element, use basic SRoA. 2357 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2358 if (ST->getNumElements() > 1) return false; 2359 } else { 2360 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2361 return false; 2362 } 2363 } 2364 2365 return true; 2366} 2367 2368 2369 2370/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2371/// some part of a constant global variable. This intentionally only accepts 2372/// constant expressions because we don't can't rewrite arbitrary instructions. 2373static bool PointsToConstantGlobal(Value *V) { 2374 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2375 return GV->isConstant(); 2376 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2377 if (CE->getOpcode() == Instruction::BitCast || 2378 CE->getOpcode() == Instruction::GetElementPtr) 2379 return PointsToConstantGlobal(CE->getOperand(0)); 2380 return false; 2381} 2382 2383/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2384/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2385/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2386/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2387/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2388/// the alloca, and if the source pointer is a pointer to a constant global, we 2389/// can optimize this. 2390static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2391 bool isOffset) { 2392 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2393 User *U = cast<Instruction>(*UI); 2394 2395 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2396 // Ignore non-volatile loads, they are always ok. 2397 if (LI->isVolatile()) return false; 2398 continue; 2399 } 2400 2401 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2402 // If uses of the bitcast are ok, we are ok. 2403 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 2404 return false; 2405 continue; 2406 } 2407 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2408 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2409 // doesn't, it does. 2410 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2411 isOffset || !GEP->hasAllZeroIndices())) 2412 return false; 2413 continue; 2414 } 2415 2416 if (CallSite CS = U) { 2417 // If this is a readonly/readnone call site, then we know it is just a 2418 // load and we can ignore it. 2419 if (CS.onlyReadsMemory()) 2420 continue; 2421 2422 // If this is the function being called then we treat it like a load and 2423 // ignore it. 2424 if (CS.isCallee(UI)) 2425 continue; 2426 2427 // If this is being passed as a byval argument, the caller is making a 2428 // copy, so it is only a read of the alloca. 2429 unsigned ArgNo = CS.getArgumentNo(UI); 2430 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2431 continue; 2432 } 2433 2434 // If this is isn't our memcpy/memmove, reject it as something we can't 2435 // handle. 2436 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2437 if (MI == 0) 2438 return false; 2439 2440 // If the transfer is using the alloca as a source of the transfer, then 2441 // ignore it since it is a load (unless the transfer is volatile). 2442 if (UI.getOperandNo() == 1) { 2443 if (MI->isVolatile()) return false; 2444 continue; 2445 } 2446 2447 // If we already have seen a copy, reject the second one. 2448 if (TheCopy) return false; 2449 2450 // If the pointer has been offset from the start of the alloca, we can't 2451 // safely handle this. 2452 if (isOffset) return false; 2453 2454 // If the memintrinsic isn't using the alloca as the dest, reject it. 2455 if (UI.getOperandNo() != 0) return false; 2456 2457 // If the source of the memcpy/move is not a constant global, reject it. 2458 if (!PointsToConstantGlobal(MI->getSource())) 2459 return false; 2460 2461 // Otherwise, the transform is safe. Remember the copy instruction. 2462 TheCopy = MI; 2463 } 2464 return true; 2465} 2466 2467/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2468/// modified by a copy from a constant global. If we can prove this, we can 2469/// replace any uses of the alloca with uses of the global directly. 2470MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 2471 MemTransferInst *TheCopy = 0; 2472 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 2473 return TheCopy; 2474 return 0; 2475} 2476