ScalarReplAggregates.cpp revision 96cc1d0dfbcf9c7ffffc65f0aa008ff532d444f4
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/DIBuilder.h" 34#include "llvm/Analysis/Dominators.h" 35#include "llvm/Analysis/Loads.h" 36#include "llvm/Analysis/ValueTracking.h" 37#include "llvm/Target/TargetData.h" 38#include "llvm/Transforms/Utils/PromoteMemToReg.h" 39#include "llvm/Transforms/Utils/Local.h" 40#include "llvm/Transforms/Utils/SSAUpdater.h" 41#include "llvm/Support/CallSite.h" 42#include "llvm/Support/Debug.h" 43#include "llvm/Support/ErrorHandling.h" 44#include "llvm/Support/GetElementPtrTypeIterator.h" 45#include "llvm/Support/IRBuilder.h" 46#include "llvm/Support/MathExtras.h" 47#include "llvm/Support/raw_ostream.h" 48#include "llvm/ADT/SetVector.h" 49#include "llvm/ADT/SmallVector.h" 50#include "llvm/ADT/Statistic.h" 51using namespace llvm; 52 53STATISTIC(NumReplaced, "Number of allocas broken up"); 54STATISTIC(NumPromoted, "Number of allocas promoted"); 55STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); 56STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 57STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 58 59namespace { 60 struct SROA : public FunctionPass { 61 SROA(int T, bool hasDT, char &ID) 62 : FunctionPass(ID), HasDomTree(hasDT) { 63 if (T == -1) 64 SRThreshold = 128; 65 else 66 SRThreshold = T; 67 } 68 69 bool runOnFunction(Function &F); 70 71 bool performScalarRepl(Function &F); 72 bool performPromotion(Function &F); 73 74 private: 75 bool HasDomTree; 76 TargetData *TD; 77 78 /// DeadInsts - Keep track of instructions we have made dead, so that 79 /// we can remove them after we are done working. 80 SmallVector<Value*, 32> DeadInsts; 81 82 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 83 /// information about the uses. All these fields are initialized to false 84 /// and set to true when something is learned. 85 struct AllocaInfo { 86 /// The alloca to promote. 87 AllocaInst *AI; 88 89 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite 90 /// looping and avoid redundant work. 91 SmallPtrSet<PHINode*, 8> CheckedPHIs; 92 93 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 94 bool isUnsafe : 1; 95 96 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 97 bool isMemCpySrc : 1; 98 99 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 100 bool isMemCpyDst : 1; 101 102 /// hasSubelementAccess - This is true if a subelement of the alloca is 103 /// ever accessed, or false if the alloca is only accessed with mem 104 /// intrinsics or load/store that only access the entire alloca at once. 105 bool hasSubelementAccess : 1; 106 107 /// hasALoadOrStore - This is true if there are any loads or stores to it. 108 /// The alloca may just be accessed with memcpy, for example, which would 109 /// not set this. 110 bool hasALoadOrStore : 1; 111 112 explicit AllocaInfo(AllocaInst *ai) 113 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 114 hasSubelementAccess(false), hasALoadOrStore(false) {} 115 }; 116 117 unsigned SRThreshold; 118 119 void MarkUnsafe(AllocaInfo &I, Instruction *User) { 120 I.isUnsafe = true; 121 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); 122 } 123 124 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 125 126 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info); 127 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset, 128 AllocaInfo &Info); 129 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info); 130 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 131 const Type *MemOpType, bool isStore, AllocaInfo &Info, 132 Instruction *TheAccess, bool AllowWholeAccess); 133 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 134 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 135 const Type *&IdxTy); 136 137 void DoScalarReplacement(AllocaInst *AI, 138 std::vector<AllocaInst*> &WorkList); 139 void DeleteDeadInstructions(); 140 141 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 142 SmallVector<AllocaInst*, 32> &NewElts); 143 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 144 SmallVector<AllocaInst*, 32> &NewElts); 145 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 146 SmallVector<AllocaInst*, 32> &NewElts); 147 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 148 AllocaInst *AI, 149 SmallVector<AllocaInst*, 32> &NewElts); 150 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 151 SmallVector<AllocaInst*, 32> &NewElts); 152 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 153 SmallVector<AllocaInst*, 32> &NewElts); 154 155 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 156 }; 157 158 // SROA_DT - SROA that uses DominatorTree. 159 struct SROA_DT : public SROA { 160 static char ID; 161 public: 162 SROA_DT(int T = -1) : SROA(T, true, ID) { 163 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 164 } 165 166 // getAnalysisUsage - This pass does not require any passes, but we know it 167 // will not alter the CFG, so say so. 168 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 169 AU.addRequired<DominatorTree>(); 170 AU.setPreservesCFG(); 171 } 172 }; 173 174 // SROA_SSAUp - SROA that uses SSAUpdater. 175 struct SROA_SSAUp : public SROA { 176 static char ID; 177 public: 178 SROA_SSAUp(int T = -1) : SROA(T, false, ID) { 179 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 180 } 181 182 // getAnalysisUsage - This pass does not require any passes, but we know it 183 // will not alter the CFG, so say so. 184 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 185 AU.setPreservesCFG(); 186 } 187 }; 188 189} 190 191char SROA_DT::ID = 0; 192char SROA_SSAUp::ID = 0; 193 194INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 195 "Scalar Replacement of Aggregates (DT)", false, false) 196INITIALIZE_PASS_DEPENDENCY(DominatorTree) 197INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 198 "Scalar Replacement of Aggregates (DT)", false, false) 199 200INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 201 "Scalar Replacement of Aggregates (SSAUp)", false, false) 202INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 203 "Scalar Replacement of Aggregates (SSAUp)", false, false) 204 205// Public interface to the ScalarReplAggregates pass 206FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 207 bool UseDomTree) { 208 if (UseDomTree) 209 return new SROA_DT(Threshold); 210 return new SROA_SSAUp(Threshold); 211} 212 213 214//===----------------------------------------------------------------------===// 215// Convert To Scalar Optimization. 216//===----------------------------------------------------------------------===// 217 218namespace { 219/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 220/// optimization, which scans the uses of an alloca and determines if it can 221/// rewrite it in terms of a single new alloca that can be mem2reg'd. 222class ConvertToScalarInfo { 223 /// AllocaSize - The size of the alloca being considered in bytes. 224 unsigned AllocaSize; 225 const TargetData &TD; 226 227 /// IsNotTrivial - This is set to true if there is some access to the object 228 /// which means that mem2reg can't promote it. 229 bool IsNotTrivial; 230 231 /// VectorTy - This tracks the type that we should promote the vector to if 232 /// it is possible to turn it into a vector. This starts out null, and if it 233 /// isn't possible to turn into a vector type, it gets set to VoidTy. 234 const Type *VectorTy; 235 236 /// HadAVector - True if there is at least one vector access to the alloca. 237 /// We don't want to turn random arrays into vectors and use vector element 238 /// insert/extract, but if there are element accesses to something that is 239 /// also declared as a vector, we do want to promote to a vector. 240 bool HadAVector; 241 242 /// HadNonMemTransferAccess - True if there is at least one access to the 243 /// alloca that is not a MemTransferInst. We don't want to turn structs into 244 /// large integers unless there is some potential for optimization. 245 bool HadNonMemTransferAccess; 246 247public: 248 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 249 : AllocaSize(Size), TD(td), IsNotTrivial(false), VectorTy(0), 250 HadAVector(false), HadNonMemTransferAccess(false) { } 251 252 AllocaInst *TryConvert(AllocaInst *AI); 253 254private: 255 bool CanConvertToScalar(Value *V, uint64_t Offset); 256 void MergeInType(const Type *In, uint64_t Offset, bool IsLoadOrStore); 257 bool MergeInVectorType(const VectorType *VInTy, uint64_t Offset); 258 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 259 260 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 261 uint64_t Offset, IRBuilder<> &Builder); 262 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 263 uint64_t Offset, IRBuilder<> &Builder); 264}; 265} // end anonymous namespace. 266 267 268/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 269/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 270/// alloca if possible or null if not. 271AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 272 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 273 // out. 274 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 275 return 0; 276 277 // If we were able to find a vector type that can handle this with 278 // insert/extract elements, and if there was at least one use that had 279 // a vector type, promote this to a vector. We don't want to promote 280 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 281 // we just get a lot of insert/extracts. If at least one vector is 282 // involved, then we probably really do have a union of vector/array. 283 const Type *NewTy; 284 if (VectorTy && VectorTy->isVectorTy() && HadAVector) { 285 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 286 << *VectorTy << '\n'); 287 NewTy = VectorTy; // Use the vector type. 288 } else { 289 unsigned BitWidth = AllocaSize * 8; 290 if (!HadAVector && !HadNonMemTransferAccess && 291 !TD.fitsInLegalInteger(BitWidth)) 292 return 0; 293 294 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 295 // Create and insert the integer alloca. 296 NewTy = IntegerType::get(AI->getContext(), BitWidth); 297 } 298 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 299 ConvertUsesToScalar(AI, NewAI, 0); 300 return NewAI; 301} 302 303/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy) 304/// so far at the offset specified by Offset (which is specified in bytes). 305/// 306/// There are three cases we handle here: 307/// 1) A union of vector types of the same size and potentially its elements. 308/// Here we turn element accesses into insert/extract element operations. 309/// This promotes a <4 x float> with a store of float to the third element 310/// into a <4 x float> that uses insert element. 311/// 2) A union of vector types with power-of-2 size differences, e.g. a float, 312/// <2 x float> and <4 x float>. Here we turn element accesses into insert 313/// and extract element operations, and <2 x float> accesses into a cast to 314/// <2 x double>, an extract, and a cast back to <2 x float>. 315/// 3) A fully general blob of memory, which we turn into some (potentially 316/// large) integer type with extract and insert operations where the loads 317/// and stores would mutate the memory. We mark this by setting VectorTy 318/// to VoidTy. 319void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset, 320 bool IsLoadOrStore) { 321 // If we already decided to turn this into a blob of integer memory, there is 322 // nothing to be done. 323 if (VectorTy && VectorTy->isVoidTy()) 324 return; 325 326 // If this could be contributing to a vector, analyze it. 327 328 // If the In type is a vector that is the same size as the alloca, see if it 329 // matches the existing VecTy. 330 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 331 if (MergeInVectorType(VInTy, Offset)) 332 return; 333 } else if (In->isFloatTy() || In->isDoubleTy() || 334 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 335 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 336 // Full width accesses can be ignored, because they can always be turned 337 // into bitcasts. 338 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 339 if (IsLoadOrStore && EltSize == AllocaSize) 340 return; 341 342 // If we're accessing something that could be an element of a vector, see 343 // if the implied vector agrees with what we already have and if Offset is 344 // compatible with it. 345 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 346 Offset * 8 < 347 (VectorTy ? VectorTy->getPrimitiveSizeInBits() 348 : (AllocaSize / EltSize) * In->getPrimitiveSizeInBits())) { 349 if (!VectorTy) { 350 VectorTy = VectorType::get(In, AllocaSize/EltSize); 351 return; 352 } 353 354 unsigned CurrentEltSize = cast<VectorType>(VectorTy)->getElementType() 355 ->getPrimitiveSizeInBits()/8; 356 if (EltSize == CurrentEltSize) 357 return; 358 359 if (In->isIntegerTy() && isPowerOf2_32(AllocaSize / EltSize)) 360 return; 361 } 362 } 363 364 // Otherwise, we have a case that we can't handle with an optimized vector 365 // form. We can still turn this into a large integer. 366 VectorTy = Type::getVoidTy(In->getContext()); 367} 368 369/// MergeInVectorType - Handles the vector case of MergeInType, returning true 370/// if the type was successfully merged and false otherwise. 371bool ConvertToScalarInfo::MergeInVectorType(const VectorType *VInTy, 372 uint64_t Offset) { 373 // Remember if we saw a vector type. 374 HadAVector = true; 375 376 // TODO: Support nonzero offsets? 377 if (Offset != 0) 378 return false; 379 380 // Only allow vectors that are a power-of-2 away from the size of the alloca. 381 if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8))) 382 return false; 383 384 // If this the first vector we see, remember the type so that we know the 385 // element size. 386 if (!VectorTy) { 387 VectorTy = VInTy; 388 return true; 389 } 390 391 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth(); 392 unsigned InBitWidth = VInTy->getBitWidth(); 393 394 // Vectors of the same size can be converted using a simple bitcast. 395 if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8)) 396 return true; 397 398 const Type *ElementTy = cast<VectorType>(VectorTy)->getElementType(); 399 const Type *InElementTy = cast<VectorType>(VInTy)->getElementType(); 400 401 // Do not allow mixed integer and floating-point accesses from vectors of 402 // different sizes. 403 if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy()) 404 return false; 405 406 if (ElementTy->isFloatingPointTy()) { 407 // Only allow floating-point vectors of different sizes if they have the 408 // same element type. 409 // TODO: This could be loosened a bit, but would anything benefit? 410 if (ElementTy != InElementTy) 411 return false; 412 413 // There are no arbitrary-precision floating-point types, which limits the 414 // number of legal vector types with larger element types that we can form 415 // to bitcast and extract a subvector. 416 // TODO: We could support some more cases with mixed fp128 and double here. 417 if (!(BitWidth == 64 || BitWidth == 128) || 418 !(InBitWidth == 64 || InBitWidth == 128)) 419 return false; 420 } else { 421 assert(ElementTy->isIntegerTy() && "Vector elements must be either integer " 422 "or floating-point."); 423 unsigned BitWidth = ElementTy->getPrimitiveSizeInBits(); 424 unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits(); 425 426 // Do not allow integer types smaller than a byte or types whose widths are 427 // not a multiple of a byte. 428 if (BitWidth < 8 || InBitWidth < 8 || 429 BitWidth % 8 != 0 || InBitWidth % 8 != 0) 430 return false; 431 } 432 433 // Pick the largest of the two vector types. 434 if (InBitWidth > BitWidth) 435 VectorTy = VInTy; 436 437 return true; 438} 439 440/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 441/// its accesses to a single vector type, return true and set VecTy to 442/// the new type. If we could convert the alloca into a single promotable 443/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 444/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 445/// is the current offset from the base of the alloca being analyzed. 446/// 447/// If we see at least one access to the value that is as a vector type, set the 448/// SawVec flag. 449bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 450 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 451 Instruction *User = cast<Instruction>(*UI); 452 453 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 454 // Don't break volatile loads. 455 if (LI->isVolatile()) 456 return false; 457 // Don't touch MMX operations. 458 if (LI->getType()->isX86_MMXTy()) 459 return false; 460 HadNonMemTransferAccess = true; 461 MergeInType(LI->getType(), Offset, true); 462 continue; 463 } 464 465 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 466 // Storing the pointer, not into the value? 467 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 468 // Don't touch MMX operations. 469 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 470 return false; 471 HadNonMemTransferAccess = true; 472 MergeInType(SI->getOperand(0)->getType(), Offset, true); 473 continue; 474 } 475 476 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 477 IsNotTrivial = true; // Can't be mem2reg'd. 478 if (!CanConvertToScalar(BCI, Offset)) 479 return false; 480 continue; 481 } 482 483 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 484 // If this is a GEP with a variable indices, we can't handle it. 485 if (!GEP->hasAllConstantIndices()) 486 return false; 487 488 // Compute the offset that this GEP adds to the pointer. 489 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 490 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 491 &Indices[0], Indices.size()); 492 // See if all uses can be converted. 493 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 494 return false; 495 IsNotTrivial = true; // Can't be mem2reg'd. 496 HadNonMemTransferAccess = true; 497 continue; 498 } 499 500 // If this is a constant sized memset of a constant value (e.g. 0) we can 501 // handle it. 502 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 503 // Store of constant value and constant size. 504 if (!isa<ConstantInt>(MSI->getValue()) || 505 !isa<ConstantInt>(MSI->getLength())) 506 return false; 507 IsNotTrivial = true; // Can't be mem2reg'd. 508 HadNonMemTransferAccess = true; 509 continue; 510 } 511 512 // If this is a memcpy or memmove into or out of the whole allocation, we 513 // can handle it like a load or store of the scalar type. 514 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 515 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 516 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 517 return false; 518 519 IsNotTrivial = true; // Can't be mem2reg'd. 520 continue; 521 } 522 523 // Otherwise, we cannot handle this! 524 return false; 525 } 526 527 return true; 528} 529 530/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 531/// directly. This happens when we are converting an "integer union" to a 532/// single integer scalar, or when we are converting a "vector union" to a 533/// vector with insert/extractelement instructions. 534/// 535/// Offset is an offset from the original alloca, in bits that need to be 536/// shifted to the right. By the end of this, there should be no uses of Ptr. 537void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 538 uint64_t Offset) { 539 while (!Ptr->use_empty()) { 540 Instruction *User = cast<Instruction>(Ptr->use_back()); 541 542 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 543 ConvertUsesToScalar(CI, NewAI, Offset); 544 CI->eraseFromParent(); 545 continue; 546 } 547 548 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 549 // Compute the offset that this GEP adds to the pointer. 550 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 551 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 552 &Indices[0], Indices.size()); 553 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 554 GEP->eraseFromParent(); 555 continue; 556 } 557 558 IRBuilder<> Builder(User); 559 560 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 561 // The load is a bit extract from NewAI shifted right by Offset bits. 562 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 563 Value *NewLoadVal 564 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 565 LI->replaceAllUsesWith(NewLoadVal); 566 LI->eraseFromParent(); 567 continue; 568 } 569 570 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 571 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 572 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 573 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 574 Builder); 575 Builder.CreateStore(New, NewAI); 576 SI->eraseFromParent(); 577 578 // If the load we just inserted is now dead, then the inserted store 579 // overwrote the entire thing. 580 if (Old->use_empty()) 581 Old->eraseFromParent(); 582 continue; 583 } 584 585 // If this is a constant sized memset of a constant value (e.g. 0) we can 586 // transform it into a store of the expanded constant value. 587 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 588 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 589 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 590 if (NumBytes != 0) { 591 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 592 593 // Compute the value replicated the right number of times. 594 APInt APVal(NumBytes*8, Val); 595 596 // Splat the value if non-zero. 597 if (Val) 598 for (unsigned i = 1; i != NumBytes; ++i) 599 APVal |= APVal << 8; 600 601 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 602 Value *New = ConvertScalar_InsertValue( 603 ConstantInt::get(User->getContext(), APVal), 604 Old, Offset, Builder); 605 Builder.CreateStore(New, NewAI); 606 607 // If the load we just inserted is now dead, then the memset overwrote 608 // the entire thing. 609 if (Old->use_empty()) 610 Old->eraseFromParent(); 611 } 612 MSI->eraseFromParent(); 613 continue; 614 } 615 616 // If this is a memcpy or memmove into or out of the whole allocation, we 617 // can handle it like a load or store of the scalar type. 618 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 619 assert(Offset == 0 && "must be store to start of alloca"); 620 621 // If the source and destination are both to the same alloca, then this is 622 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 623 // as appropriate. 624 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 625 626 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 627 // Dest must be OrigAI, change this to be a load from the original 628 // pointer (bitcasted), then a store to our new alloca. 629 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 630 Value *SrcPtr = MTI->getSource(); 631 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 632 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 633 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 634 AIPTy = PointerType::get(AIPTy->getElementType(), 635 SPTy->getAddressSpace()); 636 } 637 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 638 639 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 640 SrcVal->setAlignment(MTI->getAlignment()); 641 Builder.CreateStore(SrcVal, NewAI); 642 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 643 // Src must be OrigAI, change this to be a load from NewAI then a store 644 // through the original dest pointer (bitcasted). 645 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 646 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 647 648 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 649 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 650 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 651 AIPTy = PointerType::get(AIPTy->getElementType(), 652 DPTy->getAddressSpace()); 653 } 654 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 655 656 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 657 NewStore->setAlignment(MTI->getAlignment()); 658 } else { 659 // Noop transfer. Src == Dst 660 } 661 662 MTI->eraseFromParent(); 663 continue; 664 } 665 666 llvm_unreachable("Unsupported operation!"); 667 } 668} 669 670/// getScaledElementType - Gets a scaled element type for a partial vector 671/// access of an alloca. The input types must be integer or floating-point 672/// scalar or vector types, and the resulting type is an integer, float or 673/// double. 674static const Type *getScaledElementType(const Type *Ty1, const Type *Ty2, 675 unsigned NewBitWidth) { 676 bool IsFP1 = Ty1->isFloatingPointTy() || 677 (Ty1->isVectorTy() && 678 cast<VectorType>(Ty1)->getElementType()->isFloatingPointTy()); 679 bool IsFP2 = Ty2->isFloatingPointTy() || 680 (Ty2->isVectorTy() && 681 cast<VectorType>(Ty2)->getElementType()->isFloatingPointTy()); 682 683 LLVMContext &Context = Ty1->getContext(); 684 685 // Prefer floating-point types over integer types, as integer types may have 686 // been created by earlier scalar replacement. 687 if (IsFP1 || IsFP2) { 688 if (NewBitWidth == 32) 689 return Type::getFloatTy(Context); 690 if (NewBitWidth == 64) 691 return Type::getDoubleTy(Context); 692 } 693 694 return Type::getIntNTy(Context, NewBitWidth); 695} 696 697/// CreateShuffleVectorCast - Creates a shuffle vector to convert one vector 698/// to another vector of the same element type which has the same allocation 699/// size but different primitive sizes (e.g. <3 x i32> and <4 x i32>). 700static Value *CreateShuffleVectorCast(Value *FromVal, const Type *ToType, 701 IRBuilder<> &Builder) { 702 const Type *FromType = FromVal->getType(); 703 const VectorType *FromVTy = cast<VectorType>(FromType); 704 const VectorType *ToVTy = cast<VectorType>(ToType); 705 assert((ToVTy->getElementType() == FromVTy->getElementType()) && 706 "Vectors must have the same element type"); 707 Value *UnV = UndefValue::get(FromType); 708 unsigned numEltsFrom = FromVTy->getNumElements(); 709 unsigned numEltsTo = ToVTy->getNumElements(); 710 711 SmallVector<Constant*, 3> Args; 712 const Type* Int32Ty = Builder.getInt32Ty(); 713 unsigned minNumElts = std::min(numEltsFrom, numEltsTo); 714 unsigned i; 715 for (i=0; i != minNumElts; ++i) 716 Args.push_back(ConstantInt::get(Int32Ty, i)); 717 718 if (i < numEltsTo) { 719 Constant* UnC = UndefValue::get(Int32Ty); 720 for (; i != numEltsTo; ++i) 721 Args.push_back(UnC); 722 } 723 Constant *Mask = ConstantVector::get(Args); 724 return Builder.CreateShuffleVector(FromVal, UnV, Mask, "tmpV"); 725} 726 727/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 728/// or vector value FromVal, extracting the bits from the offset specified by 729/// Offset. This returns the value, which is of type ToType. 730/// 731/// This happens when we are converting an "integer union" to a single 732/// integer scalar, or when we are converting a "vector union" to a vector with 733/// insert/extractelement instructions. 734/// 735/// Offset is an offset from the original alloca, in bits that need to be 736/// shifted to the right. 737Value *ConvertToScalarInfo:: 738ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 739 uint64_t Offset, IRBuilder<> &Builder) { 740 // If the load is of the whole new alloca, no conversion is needed. 741 const Type *FromType = FromVal->getType(); 742 if (FromType == ToType && Offset == 0) 743 return FromVal; 744 745 // If the result alloca is a vector type, this is either an element 746 // access or a bitcast to another vector type of the same size. 747 if (const VectorType *VTy = dyn_cast<VectorType>(FromType)) { 748 unsigned FromTypeSize = TD.getTypeAllocSize(FromType); 749 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 750 if (FromTypeSize == ToTypeSize) { 751 // If the two types have the same primitive size, use a bit cast. 752 // Otherwise, it is two vectors with the same element type that has 753 // the same allocation size but different number of elements so use 754 // a shuffle vector. 755 if (FromType->getPrimitiveSizeInBits() == 756 ToType->getPrimitiveSizeInBits()) 757 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 758 else 759 return CreateShuffleVectorCast(FromVal, ToType, Builder); 760 } 761 762 if (isPowerOf2_64(FromTypeSize / ToTypeSize)) { 763 assert(!(ToType->isVectorTy() && Offset != 0) && "Can't extract a value " 764 "of a smaller vector type at a nonzero offset."); 765 766 const Type *CastElementTy = getScaledElementType(FromType, ToType, 767 ToTypeSize * 8); 768 unsigned NumCastVectorElements = FromTypeSize / ToTypeSize; 769 770 LLVMContext &Context = FromVal->getContext(); 771 const Type *CastTy = VectorType::get(CastElementTy, 772 NumCastVectorElements); 773 Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp"); 774 775 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); 776 unsigned Elt = Offset/EltSize; 777 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 778 Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get( 779 Type::getInt32Ty(Context), Elt), "tmp"); 780 return Builder.CreateBitCast(Extract, ToType, "tmp"); 781 } 782 783 // Otherwise it must be an element access. 784 unsigned Elt = 0; 785 if (Offset) { 786 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 787 Elt = Offset/EltSize; 788 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 789 } 790 // Return the element extracted out of it. 791 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 792 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 793 if (V->getType() != ToType) 794 V = Builder.CreateBitCast(V, ToType, "tmp"); 795 return V; 796 } 797 798 // If ToType is a first class aggregate, extract out each of the pieces and 799 // use insertvalue's to form the FCA. 800 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 801 const StructLayout &Layout = *TD.getStructLayout(ST); 802 Value *Res = UndefValue::get(ST); 803 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 804 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 805 Offset+Layout.getElementOffsetInBits(i), 806 Builder); 807 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 808 } 809 return Res; 810 } 811 812 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 813 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 814 Value *Res = UndefValue::get(AT); 815 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 816 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 817 Offset+i*EltSize, Builder); 818 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 819 } 820 return Res; 821 } 822 823 // Otherwise, this must be a union that was converted to an integer value. 824 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 825 826 // If this is a big-endian system and the load is narrower than the 827 // full alloca type, we need to do a shift to get the right bits. 828 int ShAmt = 0; 829 if (TD.isBigEndian()) { 830 // On big-endian machines, the lowest bit is stored at the bit offset 831 // from the pointer given by getTypeStoreSizeInBits. This matters for 832 // integers with a bitwidth that is not a multiple of 8. 833 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 834 TD.getTypeStoreSizeInBits(ToType) - Offset; 835 } else { 836 ShAmt = Offset; 837 } 838 839 // Note: we support negative bitwidths (with shl) which are not defined. 840 // We do this to support (f.e.) loads off the end of a structure where 841 // only some bits are used. 842 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 843 FromVal = Builder.CreateLShr(FromVal, 844 ConstantInt::get(FromVal->getType(), 845 ShAmt), "tmp"); 846 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 847 FromVal = Builder.CreateShl(FromVal, 848 ConstantInt::get(FromVal->getType(), 849 -ShAmt), "tmp"); 850 851 // Finally, unconditionally truncate the integer to the right width. 852 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 853 if (LIBitWidth < NTy->getBitWidth()) 854 FromVal = 855 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 856 LIBitWidth), "tmp"); 857 else if (LIBitWidth > NTy->getBitWidth()) 858 FromVal = 859 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 860 LIBitWidth), "tmp"); 861 862 // If the result is an integer, this is a trunc or bitcast. 863 if (ToType->isIntegerTy()) { 864 // Should be done. 865 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 866 // Just do a bitcast, we know the sizes match up. 867 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 868 } else { 869 // Otherwise must be a pointer. 870 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 871 } 872 assert(FromVal->getType() == ToType && "Didn't convert right?"); 873 return FromVal; 874} 875 876/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 877/// or vector value "Old" at the offset specified by Offset. 878/// 879/// This happens when we are converting an "integer union" to a 880/// single integer scalar, or when we are converting a "vector union" to a 881/// vector with insert/extractelement instructions. 882/// 883/// Offset is an offset from the original alloca, in bits that need to be 884/// shifted to the right. 885Value *ConvertToScalarInfo:: 886ConvertScalar_InsertValue(Value *SV, Value *Old, 887 uint64_t Offset, IRBuilder<> &Builder) { 888 // Convert the stored type to the actual type, shift it left to insert 889 // then 'or' into place. 890 const Type *AllocaType = Old->getType(); 891 LLVMContext &Context = Old->getContext(); 892 893 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 894 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 895 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 896 897 // Changing the whole vector with memset or with an access of a different 898 // vector type? 899 if (ValSize == VecSize) { 900 // If the two types have the same primitive size, use a bit cast. 901 // Otherwise, it is two vectors with the same element type that has 902 // the same allocation size but different number of elements so use 903 // a shuffle vector. 904 if (VTy->getPrimitiveSizeInBits() == 905 SV->getType()->getPrimitiveSizeInBits()) 906 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 907 else 908 return CreateShuffleVectorCast(SV, VTy, Builder); 909 } 910 911 if (isPowerOf2_64(VecSize / ValSize)) { 912 assert(!(SV->getType()->isVectorTy() && Offset != 0) && "Can't insert a " 913 "value of a smaller vector type at a nonzero offset."); 914 915 const Type *CastElementTy = getScaledElementType(VTy, SV->getType(), 916 ValSize); 917 unsigned NumCastVectorElements = VecSize / ValSize; 918 919 LLVMContext &Context = SV->getContext(); 920 const Type *OldCastTy = VectorType::get(CastElementTy, 921 NumCastVectorElements); 922 Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp"); 923 924 Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp"); 925 926 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); 927 unsigned Elt = Offset/EltSize; 928 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 929 Value *Insert = 930 Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get( 931 Type::getInt32Ty(Context), Elt), "tmp"); 932 return Builder.CreateBitCast(Insert, AllocaType, "tmp"); 933 } 934 935 // Must be an element insertion. 936 assert(SV->getType() == VTy->getElementType()); 937 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 938 unsigned Elt = Offset/EltSize; 939 return Builder.CreateInsertElement(Old, SV, 940 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 941 "tmp"); 942 } 943 944 // If SV is a first-class aggregate value, insert each value recursively. 945 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 946 const StructLayout &Layout = *TD.getStructLayout(ST); 947 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 948 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 949 Old = ConvertScalar_InsertValue(Elt, Old, 950 Offset+Layout.getElementOffsetInBits(i), 951 Builder); 952 } 953 return Old; 954 } 955 956 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 957 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 958 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 959 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 960 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 961 } 962 return Old; 963 } 964 965 // If SV is a float, convert it to the appropriate integer type. 966 // If it is a pointer, do the same. 967 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 968 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 969 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 970 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 971 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 972 SV = Builder.CreateBitCast(SV, 973 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 974 else if (SV->getType()->isPointerTy()) 975 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 976 977 // Zero extend or truncate the value if needed. 978 if (SV->getType() != AllocaType) { 979 if (SV->getType()->getPrimitiveSizeInBits() < 980 AllocaType->getPrimitiveSizeInBits()) 981 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 982 else { 983 // Truncation may be needed if storing more than the alloca can hold 984 // (undefined behavior). 985 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 986 SrcWidth = DestWidth; 987 SrcStoreWidth = DestStoreWidth; 988 } 989 } 990 991 // If this is a big-endian system and the store is narrower than the 992 // full alloca type, we need to do a shift to get the right bits. 993 int ShAmt = 0; 994 if (TD.isBigEndian()) { 995 // On big-endian machines, the lowest bit is stored at the bit offset 996 // from the pointer given by getTypeStoreSizeInBits. This matters for 997 // integers with a bitwidth that is not a multiple of 8. 998 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 999 } else { 1000 ShAmt = Offset; 1001 } 1002 1003 // Note: we support negative bitwidths (with shr) which are not defined. 1004 // We do this to support (f.e.) stores off the end of a structure where 1005 // only some bits in the structure are set. 1006 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1007 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1008 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1009 ShAmt), "tmp"); 1010 Mask <<= ShAmt; 1011 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1012 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1013 -ShAmt), "tmp"); 1014 Mask = Mask.lshr(-ShAmt); 1015 } 1016 1017 // Mask out the bits we are about to insert from the old value, and or 1018 // in the new bits. 1019 if (SrcWidth != DestWidth) { 1020 assert(DestWidth > SrcWidth); 1021 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1022 SV = Builder.CreateOr(Old, SV, "ins"); 1023 } 1024 return SV; 1025} 1026 1027 1028//===----------------------------------------------------------------------===// 1029// SRoA Driver 1030//===----------------------------------------------------------------------===// 1031 1032 1033bool SROA::runOnFunction(Function &F) { 1034 TD = getAnalysisIfAvailable<TargetData>(); 1035 1036 bool Changed = performPromotion(F); 1037 1038 // FIXME: ScalarRepl currently depends on TargetData more than it 1039 // theoretically needs to. It should be refactored in order to support 1040 // target-independent IR. Until this is done, just skip the actual 1041 // scalar-replacement portion of this pass. 1042 if (!TD) return Changed; 1043 1044 while (1) { 1045 bool LocalChange = performScalarRepl(F); 1046 if (!LocalChange) break; // No need to repromote if no scalarrepl 1047 Changed = true; 1048 LocalChange = performPromotion(F); 1049 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 1050 } 1051 1052 return Changed; 1053} 1054 1055namespace { 1056class AllocaPromoter : public LoadAndStorePromoter { 1057 AllocaInst *AI; 1058public: 1059 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 1060 DbgDeclareInst *DD, DIBuilder *&DB) 1061 : LoadAndStorePromoter(Insts, S, DD, DB), AI(0) {} 1062 1063 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 1064 // Remember which alloca we're promoting (for isInstInList). 1065 this->AI = AI; 1066 LoadAndStorePromoter::run(Insts); 1067 AI->eraseFromParent(); 1068 } 1069 1070 virtual bool isInstInList(Instruction *I, 1071 const SmallVectorImpl<Instruction*> &Insts) const { 1072 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1073 return LI->getOperand(0) == AI; 1074 return cast<StoreInst>(I)->getPointerOperand() == AI; 1075 } 1076}; 1077} // end anon namespace 1078 1079/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1080/// subsequently loaded can be rewritten to load both input pointers and then 1081/// select between the result, allowing the load of the alloca to be promoted. 1082/// From this: 1083/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1084/// %V = load i32* %P2 1085/// to: 1086/// %V1 = load i32* %Alloca -> will be mem2reg'd 1087/// %V2 = load i32* %Other 1088/// %V = select i1 %cond, i32 %V1, i32 %V2 1089/// 1090/// We can do this to a select if its only uses are loads and if the operand to 1091/// the select can be loaded unconditionally. 1092static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1093 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1094 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1095 1096 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1097 UI != UE; ++UI) { 1098 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1099 if (LI == 0 || LI->isVolatile()) return false; 1100 1101 // Both operands to the select need to be dereferencable, either absolutely 1102 // (e.g. allocas) or at this point because we can see other accesses to it. 1103 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1104 LI->getAlignment(), TD)) 1105 return false; 1106 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1107 LI->getAlignment(), TD)) 1108 return false; 1109 } 1110 1111 return true; 1112} 1113 1114/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1115/// subsequently loaded can be rewritten to load both input pointers in the pred 1116/// blocks and then PHI the results, allowing the load of the alloca to be 1117/// promoted. 1118/// From this: 1119/// %P2 = phi [i32* %Alloca, i32* %Other] 1120/// %V = load i32* %P2 1121/// to: 1122/// %V1 = load i32* %Alloca -> will be mem2reg'd 1123/// ... 1124/// %V2 = load i32* %Other 1125/// ... 1126/// %V = phi [i32 %V1, i32 %V2] 1127/// 1128/// We can do this to a select if its only uses are loads and if the operand to 1129/// the select can be loaded unconditionally. 1130static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1131 // For now, we can only do this promotion if the load is in the same block as 1132 // the PHI, and if there are no stores between the phi and load. 1133 // TODO: Allow recursive phi users. 1134 // TODO: Allow stores. 1135 BasicBlock *BB = PN->getParent(); 1136 unsigned MaxAlign = 0; 1137 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1138 UI != UE; ++UI) { 1139 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1140 if (LI == 0 || LI->isVolatile()) return false; 1141 1142 // For now we only allow loads in the same block as the PHI. This is a 1143 // common case that happens when instcombine merges two loads through a PHI. 1144 if (LI->getParent() != BB) return false; 1145 1146 // Ensure that there are no instructions between the PHI and the load that 1147 // could store. 1148 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1149 if (BBI->mayWriteToMemory()) 1150 return false; 1151 1152 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1153 } 1154 1155 // Okay, we know that we have one or more loads in the same block as the PHI. 1156 // We can transform this if it is safe to push the loads into the predecessor 1157 // blocks. The only thing to watch out for is that we can't put a possibly 1158 // trapping load in the predecessor if it is a critical edge. 1159 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1160 BasicBlock *Pred = PN->getIncomingBlock(i); 1161 1162 // If the predecessor has a single successor, then the edge isn't critical. 1163 if (Pred->getTerminator()->getNumSuccessors() == 1) 1164 continue; 1165 1166 Value *InVal = PN->getIncomingValue(i); 1167 1168 // If the InVal is an invoke in the pred, we can't put a load on the edge. 1169 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 1170 if (II->getParent() == Pred) 1171 return false; 1172 1173 // If this pointer is always safe to load, or if we can prove that there is 1174 // already a load in the block, then we can move the load to the pred block. 1175 if (InVal->isDereferenceablePointer() || 1176 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1177 continue; 1178 1179 return false; 1180 } 1181 1182 return true; 1183} 1184 1185 1186/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1187/// direct (non-volatile) loads and stores to it. If the alloca is close but 1188/// not quite there, this will transform the code to allow promotion. As such, 1189/// it is a non-pure predicate. 1190static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1191 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1192 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1193 1194 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1195 UI != UE; ++UI) { 1196 User *U = *UI; 1197 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1198 if (LI->isVolatile()) 1199 return false; 1200 continue; 1201 } 1202 1203 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1204 if (SI->getOperand(0) == AI || SI->isVolatile()) 1205 return false; // Don't allow a store OF the AI, only INTO the AI. 1206 continue; 1207 } 1208 1209 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1210 // If the condition being selected on is a constant, fold the select, yes 1211 // this does (rarely) happen early on. 1212 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1213 Value *Result = SI->getOperand(1+CI->isZero()); 1214 SI->replaceAllUsesWith(Result); 1215 SI->eraseFromParent(); 1216 1217 // This is very rare and we just scrambled the use list of AI, start 1218 // over completely. 1219 return tryToMakeAllocaBePromotable(AI, TD); 1220 } 1221 1222 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1223 // loads, then we can transform this by rewriting the select. 1224 if (!isSafeSelectToSpeculate(SI, TD)) 1225 return false; 1226 1227 InstsToRewrite.insert(SI); 1228 continue; 1229 } 1230 1231 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1232 if (PN->use_empty()) { // Dead PHIs can be stripped. 1233 InstsToRewrite.insert(PN); 1234 continue; 1235 } 1236 1237 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1238 // in the pred blocks, then we can transform this by rewriting the PHI. 1239 if (!isSafePHIToSpeculate(PN, TD)) 1240 return false; 1241 1242 InstsToRewrite.insert(PN); 1243 continue; 1244 } 1245 1246 return false; 1247 } 1248 1249 // If there are no instructions to rewrite, then all uses are load/stores and 1250 // we're done! 1251 if (InstsToRewrite.empty()) 1252 return true; 1253 1254 // If we have instructions that need to be rewritten for this to be promotable 1255 // take care of it now. 1256 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1257 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1258 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1259 // loads with a new select. 1260 while (!SI->use_empty()) { 1261 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1262 1263 IRBuilder<> Builder(LI); 1264 LoadInst *TrueLoad = 1265 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1266 LoadInst *FalseLoad = 1267 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".t"); 1268 1269 // Transfer alignment and TBAA info if present. 1270 TrueLoad->setAlignment(LI->getAlignment()); 1271 FalseLoad->setAlignment(LI->getAlignment()); 1272 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1273 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1274 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1275 } 1276 1277 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1278 V->takeName(LI); 1279 LI->replaceAllUsesWith(V); 1280 LI->eraseFromParent(); 1281 } 1282 1283 // Now that all the loads are gone, the select is gone too. 1284 SI->eraseFromParent(); 1285 continue; 1286 } 1287 1288 // Otherwise, we have a PHI node which allows us to push the loads into the 1289 // predecessors. 1290 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1291 if (PN->use_empty()) { 1292 PN->eraseFromParent(); 1293 continue; 1294 } 1295 1296 const Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1297 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1298 PN->getName()+".ld", PN); 1299 1300 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1301 // matter which one we get and if any differ, it doesn't matter. 1302 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1303 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1304 unsigned Align = SomeLoad->getAlignment(); 1305 1306 // Rewrite all loads of the PN to use the new PHI. 1307 while (!PN->use_empty()) { 1308 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1309 LI->replaceAllUsesWith(NewPN); 1310 LI->eraseFromParent(); 1311 } 1312 1313 // Inject loads into all of the pred blocks. Keep track of which blocks we 1314 // insert them into in case we have multiple edges from the same block. 1315 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1316 1317 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1318 BasicBlock *Pred = PN->getIncomingBlock(i); 1319 LoadInst *&Load = InsertedLoads[Pred]; 1320 if (Load == 0) { 1321 Load = new LoadInst(PN->getIncomingValue(i), 1322 PN->getName() + "." + Pred->getName(), 1323 Pred->getTerminator()); 1324 Load->setAlignment(Align); 1325 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1326 } 1327 1328 NewPN->addIncoming(Load, Pred); 1329 } 1330 1331 PN->eraseFromParent(); 1332 } 1333 1334 ++NumAdjusted; 1335 return true; 1336} 1337 1338bool SROA::performPromotion(Function &F) { 1339 std::vector<AllocaInst*> Allocas; 1340 DominatorTree *DT = 0; 1341 if (HasDomTree) 1342 DT = &getAnalysis<DominatorTree>(); 1343 1344 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1345 1346 bool Changed = false; 1347 SmallVector<Instruction*, 64> Insts; 1348 DIBuilder *DIB = 0; 1349 while (1) { 1350 Allocas.clear(); 1351 1352 // Find allocas that are safe to promote, by looking at all instructions in 1353 // the entry node 1354 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1355 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1356 if (tryToMakeAllocaBePromotable(AI, TD)) 1357 Allocas.push_back(AI); 1358 1359 if (Allocas.empty()) break; 1360 1361 if (HasDomTree) 1362 PromoteMemToReg(Allocas, *DT); 1363 else { 1364 SSAUpdater SSA; 1365 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1366 AllocaInst *AI = Allocas[i]; 1367 1368 // Build list of instructions to promote. 1369 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1370 UI != E; ++UI) 1371 Insts.push_back(cast<Instruction>(*UI)); 1372 1373 DbgDeclareInst *DDI = FindAllocaDbgDeclare(AI); 1374 if (DDI && !DIB) 1375 DIB = new DIBuilder(*AI->getParent()->getParent()->getParent()); 1376 AllocaPromoter(Insts, SSA, DDI, DIB).run(AI, Insts); 1377 Insts.clear(); 1378 } 1379 } 1380 NumPromoted += Allocas.size(); 1381 Changed = true; 1382 } 1383 1384 // FIXME: Is there a better way to handle the lazy initialization of DIB 1385 // so that there doesn't need to be an explicit delete? 1386 delete DIB; 1387 1388 return Changed; 1389} 1390 1391 1392/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1393/// SROA. It must be a struct or array type with a small number of elements. 1394static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1395 const Type *T = AI->getAllocatedType(); 1396 // Do not promote any struct into more than 32 separate vars. 1397 if (const StructType *ST = dyn_cast<StructType>(T)) 1398 return ST->getNumElements() <= 32; 1399 // Arrays are much less likely to be safe for SROA; only consider 1400 // them if they are very small. 1401 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 1402 return AT->getNumElements() <= 8; 1403 return false; 1404} 1405 1406 1407// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1408// which runs on all of the malloc/alloca instructions in the function, removing 1409// them if they are only used by getelementptr instructions. 1410// 1411bool SROA::performScalarRepl(Function &F) { 1412 std::vector<AllocaInst*> WorkList; 1413 1414 // Scan the entry basic block, adding allocas to the worklist. 1415 BasicBlock &BB = F.getEntryBlock(); 1416 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1417 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1418 WorkList.push_back(A); 1419 1420 // Process the worklist 1421 bool Changed = false; 1422 while (!WorkList.empty()) { 1423 AllocaInst *AI = WorkList.back(); 1424 WorkList.pop_back(); 1425 1426 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1427 // with unused elements. 1428 if (AI->use_empty()) { 1429 AI->eraseFromParent(); 1430 Changed = true; 1431 continue; 1432 } 1433 1434 // If this alloca is impossible for us to promote, reject it early. 1435 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1436 continue; 1437 1438 // Check to see if this allocation is only modified by a memcpy/memmove from 1439 // a constant global. If this is the case, we can change all users to use 1440 // the constant global instead. This is commonly produced by the CFE by 1441 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1442 // is only subsequently read. 1443 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 1444 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1445 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 1446 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 1447 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1448 TheCopy->eraseFromParent(); // Don't mutate the global. 1449 AI->eraseFromParent(); 1450 ++NumGlobals; 1451 Changed = true; 1452 continue; 1453 } 1454 1455 // Check to see if we can perform the core SROA transformation. We cannot 1456 // transform the allocation instruction if it is an array allocation 1457 // (allocations OF arrays are ok though), and an allocation of a scalar 1458 // value cannot be decomposed at all. 1459 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1460 1461 // Do not promote [0 x %struct]. 1462 if (AllocaSize == 0) continue; 1463 1464 // Do not promote any struct whose size is too big. 1465 if (AllocaSize > SRThreshold) continue; 1466 1467 // If the alloca looks like a good candidate for scalar replacement, and if 1468 // all its users can be transformed, then split up the aggregate into its 1469 // separate elements. 1470 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1471 DoScalarReplacement(AI, WorkList); 1472 Changed = true; 1473 continue; 1474 } 1475 1476 // If we can turn this aggregate value (potentially with casts) into a 1477 // simple scalar value that can be mem2reg'd into a register value. 1478 // IsNotTrivial tracks whether this is something that mem2reg could have 1479 // promoted itself. If so, we don't want to transform it needlessly. Note 1480 // that we can't just check based on the type: the alloca may be of an i32 1481 // but that has pointer arithmetic to set byte 3 of it or something. 1482 if (AllocaInst *NewAI = 1483 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1484 NewAI->takeName(AI); 1485 AI->eraseFromParent(); 1486 ++NumConverted; 1487 Changed = true; 1488 continue; 1489 } 1490 1491 // Otherwise, couldn't process this alloca. 1492 } 1493 1494 return Changed; 1495} 1496 1497/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1498/// predicate, do SROA now. 1499void SROA::DoScalarReplacement(AllocaInst *AI, 1500 std::vector<AllocaInst*> &WorkList) { 1501 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1502 SmallVector<AllocaInst*, 32> ElementAllocas; 1503 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1504 ElementAllocas.reserve(ST->getNumContainedTypes()); 1505 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1506 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1507 AI->getAlignment(), 1508 AI->getName() + "." + Twine(i), AI); 1509 ElementAllocas.push_back(NA); 1510 WorkList.push_back(NA); // Add to worklist for recursive processing 1511 } 1512 } else { 1513 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1514 ElementAllocas.reserve(AT->getNumElements()); 1515 const Type *ElTy = AT->getElementType(); 1516 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1517 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1518 AI->getName() + "." + Twine(i), AI); 1519 ElementAllocas.push_back(NA); 1520 WorkList.push_back(NA); // Add to worklist for recursive processing 1521 } 1522 } 1523 1524 // Now that we have created the new alloca instructions, rewrite all the 1525 // uses of the old alloca. 1526 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1527 1528 // Now erase any instructions that were made dead while rewriting the alloca. 1529 DeleteDeadInstructions(); 1530 AI->eraseFromParent(); 1531 1532 ++NumReplaced; 1533} 1534 1535/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1536/// recursively including all their operands that become trivially dead. 1537void SROA::DeleteDeadInstructions() { 1538 while (!DeadInsts.empty()) { 1539 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1540 1541 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1542 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1543 // Zero out the operand and see if it becomes trivially dead. 1544 // (But, don't add allocas to the dead instruction list -- they are 1545 // already on the worklist and will be deleted separately.) 1546 *OI = 0; 1547 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1548 DeadInsts.push_back(U); 1549 } 1550 1551 I->eraseFromParent(); 1552 } 1553} 1554 1555/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1556/// performing scalar replacement of alloca AI. The results are flagged in 1557/// the Info parameter. Offset indicates the position within AI that is 1558/// referenced by this instruction. 1559void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1560 AllocaInfo &Info) { 1561 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1562 Instruction *User = cast<Instruction>(*UI); 1563 1564 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1565 isSafeForScalarRepl(BC, Offset, Info); 1566 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1567 uint64_t GEPOffset = Offset; 1568 isSafeGEP(GEPI, GEPOffset, Info); 1569 if (!Info.isUnsafe) 1570 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1571 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1572 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1573 if (Length == 0) 1574 return MarkUnsafe(Info, User); 1575 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1576 UI.getOperandNo() == 0, Info, MI, 1577 true /*AllowWholeAccess*/); 1578 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1579 if (LI->isVolatile()) 1580 return MarkUnsafe(Info, User); 1581 const Type *LIType = LI->getType(); 1582 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1583 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1584 Info.hasALoadOrStore = true; 1585 1586 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1587 // Store is ok if storing INTO the pointer, not storing the pointer 1588 if (SI->isVolatile() || SI->getOperand(0) == I) 1589 return MarkUnsafe(Info, User); 1590 1591 const Type *SIType = SI->getOperand(0)->getType(); 1592 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1593 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1594 Info.hasALoadOrStore = true; 1595 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1596 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1597 } else { 1598 return MarkUnsafe(Info, User); 1599 } 1600 if (Info.isUnsafe) return; 1601 } 1602} 1603 1604 1605/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1606/// derived from the alloca, we can often still split the alloca into elements. 1607/// This is useful if we have a large alloca where one element is phi'd 1608/// together somewhere: we can SRoA and promote all the other elements even if 1609/// we end up not being able to promote this one. 1610/// 1611/// All we require is that the uses of the PHI do not index into other parts of 1612/// the alloca. The most important use case for this is single load and stores 1613/// that are PHI'd together, which can happen due to code sinking. 1614void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1615 AllocaInfo &Info) { 1616 // If we've already checked this PHI, don't do it again. 1617 if (PHINode *PN = dyn_cast<PHINode>(I)) 1618 if (!Info.CheckedPHIs.insert(PN)) 1619 return; 1620 1621 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1622 Instruction *User = cast<Instruction>(*UI); 1623 1624 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1625 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1626 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1627 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1628 // but would have to prove that we're staying inside of an element being 1629 // promoted. 1630 if (!GEPI->hasAllZeroIndices()) 1631 return MarkUnsafe(Info, User); 1632 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1633 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1634 if (LI->isVolatile()) 1635 return MarkUnsafe(Info, User); 1636 const Type *LIType = LI->getType(); 1637 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1638 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1639 Info.hasALoadOrStore = true; 1640 1641 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1642 // Store is ok if storing INTO the pointer, not storing the pointer 1643 if (SI->isVolatile() || SI->getOperand(0) == I) 1644 return MarkUnsafe(Info, User); 1645 1646 const Type *SIType = SI->getOperand(0)->getType(); 1647 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1648 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1649 Info.hasALoadOrStore = true; 1650 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1651 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1652 } else { 1653 return MarkUnsafe(Info, User); 1654 } 1655 if (Info.isUnsafe) return; 1656 } 1657} 1658 1659/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1660/// replacement. It is safe when all the indices are constant, in-bounds 1661/// references, and when the resulting offset corresponds to an element within 1662/// the alloca type. The results are flagged in the Info parameter. Upon 1663/// return, Offset is adjusted as specified by the GEP indices. 1664void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1665 uint64_t &Offset, AllocaInfo &Info) { 1666 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1667 if (GEPIt == E) 1668 return; 1669 1670 // Walk through the GEP type indices, checking the types that this indexes 1671 // into. 1672 for (; GEPIt != E; ++GEPIt) { 1673 // Ignore struct elements, no extra checking needed for these. 1674 if ((*GEPIt)->isStructTy()) 1675 continue; 1676 1677 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1678 if (!IdxVal) 1679 return MarkUnsafe(Info, GEPI); 1680 } 1681 1682 // Compute the offset due to this GEP and check if the alloca has a 1683 // component element at that offset. 1684 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1685 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1686 &Indices[0], Indices.size()); 1687 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1688 MarkUnsafe(Info, GEPI); 1689} 1690 1691/// isHomogeneousAggregate - Check if type T is a struct or array containing 1692/// elements of the same type (which is always true for arrays). If so, 1693/// return true with NumElts and EltTy set to the number of elements and the 1694/// element type, respectively. 1695static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1696 const Type *&EltTy) { 1697 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1698 NumElts = AT->getNumElements(); 1699 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1700 return true; 1701 } 1702 if (const StructType *ST = dyn_cast<StructType>(T)) { 1703 NumElts = ST->getNumContainedTypes(); 1704 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1705 for (unsigned n = 1; n < NumElts; ++n) { 1706 if (ST->getContainedType(n) != EltTy) 1707 return false; 1708 } 1709 return true; 1710 } 1711 return false; 1712} 1713 1714/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1715/// "homogeneous" aggregates with the same element type and number of elements. 1716static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1717 if (T1 == T2) 1718 return true; 1719 1720 unsigned NumElts1, NumElts2; 1721 const Type *EltTy1, *EltTy2; 1722 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1723 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1724 NumElts1 == NumElts2 && 1725 EltTy1 == EltTy2) 1726 return true; 1727 1728 return false; 1729} 1730 1731/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1732/// alloca or has an offset and size that corresponds to a component element 1733/// within it. The offset checked here may have been formed from a GEP with a 1734/// pointer bitcasted to a different type. 1735/// 1736/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1737/// unit. If false, it only allows accesses known to be in a single element. 1738void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1739 const Type *MemOpType, bool isStore, 1740 AllocaInfo &Info, Instruction *TheAccess, 1741 bool AllowWholeAccess) { 1742 // Check if this is a load/store of the entire alloca. 1743 if (Offset == 0 && AllowWholeAccess && 1744 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1745 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1746 // loads/stores (which are essentially the same as the MemIntrinsics with 1747 // regard to copying padding between elements). But, if an alloca is 1748 // flagged as both a source and destination of such operations, we'll need 1749 // to check later for padding between elements. 1750 if (!MemOpType || MemOpType->isIntegerTy()) { 1751 if (isStore) 1752 Info.isMemCpyDst = true; 1753 else 1754 Info.isMemCpySrc = true; 1755 return; 1756 } 1757 // This is also safe for references using a type that is compatible with 1758 // the type of the alloca, so that loads/stores can be rewritten using 1759 // insertvalue/extractvalue. 1760 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1761 Info.hasSubelementAccess = true; 1762 return; 1763 } 1764 } 1765 // Check if the offset/size correspond to a component within the alloca type. 1766 const Type *T = Info.AI->getAllocatedType(); 1767 if (TypeHasComponent(T, Offset, MemSize)) { 1768 Info.hasSubelementAccess = true; 1769 return; 1770 } 1771 1772 return MarkUnsafe(Info, TheAccess); 1773} 1774 1775/// TypeHasComponent - Return true if T has a component type with the 1776/// specified offset and size. If Size is zero, do not check the size. 1777bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1778 const Type *EltTy; 1779 uint64_t EltSize; 1780 if (const StructType *ST = dyn_cast<StructType>(T)) { 1781 const StructLayout *Layout = TD->getStructLayout(ST); 1782 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1783 EltTy = ST->getContainedType(EltIdx); 1784 EltSize = TD->getTypeAllocSize(EltTy); 1785 Offset -= Layout->getElementOffset(EltIdx); 1786 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1787 EltTy = AT->getElementType(); 1788 EltSize = TD->getTypeAllocSize(EltTy); 1789 if (Offset >= AT->getNumElements() * EltSize) 1790 return false; 1791 Offset %= EltSize; 1792 } else { 1793 return false; 1794 } 1795 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1796 return true; 1797 // Check if the component spans multiple elements. 1798 if (Offset + Size > EltSize) 1799 return false; 1800 return TypeHasComponent(EltTy, Offset, Size); 1801} 1802 1803/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1804/// the instruction I, which references it, to use the separate elements. 1805/// Offset indicates the position within AI that is referenced by this 1806/// instruction. 1807void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1808 SmallVector<AllocaInst*, 32> &NewElts) { 1809 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1810 Use &TheUse = UI.getUse(); 1811 Instruction *User = cast<Instruction>(*UI++); 1812 1813 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1814 RewriteBitCast(BC, AI, Offset, NewElts); 1815 continue; 1816 } 1817 1818 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1819 RewriteGEP(GEPI, AI, Offset, NewElts); 1820 continue; 1821 } 1822 1823 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1824 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1825 uint64_t MemSize = Length->getZExtValue(); 1826 if (Offset == 0 && 1827 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1828 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1829 // Otherwise the intrinsic can only touch a single element and the 1830 // address operand will be updated, so nothing else needs to be done. 1831 continue; 1832 } 1833 1834 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1835 const Type *LIType = LI->getType(); 1836 1837 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1838 // Replace: 1839 // %res = load { i32, i32 }* %alloc 1840 // with: 1841 // %load.0 = load i32* %alloc.0 1842 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1843 // %load.1 = load i32* %alloc.1 1844 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1845 // (Also works for arrays instead of structs) 1846 Value *Insert = UndefValue::get(LIType); 1847 IRBuilder<> Builder(LI); 1848 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1849 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1850 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1851 } 1852 LI->replaceAllUsesWith(Insert); 1853 DeadInsts.push_back(LI); 1854 } else if (LIType->isIntegerTy() && 1855 TD->getTypeAllocSize(LIType) == 1856 TD->getTypeAllocSize(AI->getAllocatedType())) { 1857 // If this is a load of the entire alloca to an integer, rewrite it. 1858 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1859 } 1860 continue; 1861 } 1862 1863 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1864 Value *Val = SI->getOperand(0); 1865 const Type *SIType = Val->getType(); 1866 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1867 // Replace: 1868 // store { i32, i32 } %val, { i32, i32 }* %alloc 1869 // with: 1870 // %val.0 = extractvalue { i32, i32 } %val, 0 1871 // store i32 %val.0, i32* %alloc.0 1872 // %val.1 = extractvalue { i32, i32 } %val, 1 1873 // store i32 %val.1, i32* %alloc.1 1874 // (Also works for arrays instead of structs) 1875 IRBuilder<> Builder(SI); 1876 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1877 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1878 Builder.CreateStore(Extract, NewElts[i]); 1879 } 1880 DeadInsts.push_back(SI); 1881 } else if (SIType->isIntegerTy() && 1882 TD->getTypeAllocSize(SIType) == 1883 TD->getTypeAllocSize(AI->getAllocatedType())) { 1884 // If this is a store of the entire alloca from an integer, rewrite it. 1885 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1886 } 1887 continue; 1888 } 1889 1890 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1891 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1892 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1893 // the new pointer. 1894 if (!isa<AllocaInst>(I)) continue; 1895 1896 assert(Offset == 0 && NewElts[0] && 1897 "Direct alloca use should have a zero offset"); 1898 1899 // If we have a use of the alloca, we know the derived uses will be 1900 // utilizing just the first element of the scalarized result. Insert a 1901 // bitcast of the first alloca before the user as required. 1902 AllocaInst *NewAI = NewElts[0]; 1903 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1904 NewAI->moveBefore(BCI); 1905 TheUse = BCI; 1906 continue; 1907 } 1908 } 1909} 1910 1911/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1912/// and recursively continue updating all of its uses. 1913void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1914 SmallVector<AllocaInst*, 32> &NewElts) { 1915 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1916 if (BC->getOperand(0) != AI) 1917 return; 1918 1919 // The bitcast references the original alloca. Replace its uses with 1920 // references to the first new element alloca. 1921 Instruction *Val = NewElts[0]; 1922 if (Val->getType() != BC->getDestTy()) { 1923 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1924 Val->takeName(BC); 1925 } 1926 BC->replaceAllUsesWith(Val); 1927 DeadInsts.push_back(BC); 1928} 1929 1930/// FindElementAndOffset - Return the index of the element containing Offset 1931/// within the specified type, which must be either a struct or an array. 1932/// Sets T to the type of the element and Offset to the offset within that 1933/// element. IdxTy is set to the type of the index result to be used in a 1934/// GEP instruction. 1935uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1936 const Type *&IdxTy) { 1937 uint64_t Idx = 0; 1938 if (const StructType *ST = dyn_cast<StructType>(T)) { 1939 const StructLayout *Layout = TD->getStructLayout(ST); 1940 Idx = Layout->getElementContainingOffset(Offset); 1941 T = ST->getContainedType(Idx); 1942 Offset -= Layout->getElementOffset(Idx); 1943 IdxTy = Type::getInt32Ty(T->getContext()); 1944 return Idx; 1945 } 1946 const ArrayType *AT = cast<ArrayType>(T); 1947 T = AT->getElementType(); 1948 uint64_t EltSize = TD->getTypeAllocSize(T); 1949 Idx = Offset / EltSize; 1950 Offset -= Idx * EltSize; 1951 IdxTy = Type::getInt64Ty(T->getContext()); 1952 return Idx; 1953} 1954 1955/// RewriteGEP - Check if this GEP instruction moves the pointer across 1956/// elements of the alloca that are being split apart, and if so, rewrite 1957/// the GEP to be relative to the new element. 1958void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1959 SmallVector<AllocaInst*, 32> &NewElts) { 1960 uint64_t OldOffset = Offset; 1961 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1962 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1963 &Indices[0], Indices.size()); 1964 1965 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1966 1967 const Type *T = AI->getAllocatedType(); 1968 const Type *IdxTy; 1969 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1970 if (GEPI->getOperand(0) == AI) 1971 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1972 1973 T = AI->getAllocatedType(); 1974 uint64_t EltOffset = Offset; 1975 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1976 1977 // If this GEP does not move the pointer across elements of the alloca 1978 // being split, then it does not needs to be rewritten. 1979 if (Idx == OldIdx) 1980 return; 1981 1982 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1983 SmallVector<Value*, 8> NewArgs; 1984 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1985 while (EltOffset != 0) { 1986 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1987 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1988 } 1989 Instruction *Val = NewElts[Idx]; 1990 if (NewArgs.size() > 1) { 1991 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1992 NewArgs.end(), "", GEPI); 1993 Val->takeName(GEPI); 1994 } 1995 if (Val->getType() != GEPI->getType()) 1996 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1997 GEPI->replaceAllUsesWith(Val); 1998 DeadInsts.push_back(GEPI); 1999} 2000 2001/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2002/// Rewrite it to copy or set the elements of the scalarized memory. 2003void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2004 AllocaInst *AI, 2005 SmallVector<AllocaInst*, 32> &NewElts) { 2006 // If this is a memcpy/memmove, construct the other pointer as the 2007 // appropriate type. The "Other" pointer is the pointer that goes to memory 2008 // that doesn't have anything to do with the alloca that we are promoting. For 2009 // memset, this Value* stays null. 2010 Value *OtherPtr = 0; 2011 unsigned MemAlignment = MI->getAlignment(); 2012 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2013 if (Inst == MTI->getRawDest()) 2014 OtherPtr = MTI->getRawSource(); 2015 else { 2016 assert(Inst == MTI->getRawSource()); 2017 OtherPtr = MTI->getRawDest(); 2018 } 2019 } 2020 2021 // If there is an other pointer, we want to convert it to the same pointer 2022 // type as AI has, so we can GEP through it safely. 2023 if (OtherPtr) { 2024 unsigned AddrSpace = 2025 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2026 2027 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2028 // optimization, but it's also required to detect the corner case where 2029 // both pointer operands are referencing the same memory, and where 2030 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2031 // function is only called for mem intrinsics that access the whole 2032 // aggregate, so non-zero GEPs are not an issue here.) 2033 OtherPtr = OtherPtr->stripPointerCasts(); 2034 2035 // Copying the alloca to itself is a no-op: just delete it. 2036 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2037 // This code will run twice for a no-op memcpy -- once for each operand. 2038 // Put only one reference to MI on the DeadInsts list. 2039 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2040 E = DeadInsts.end(); I != E; ++I) 2041 if (*I == MI) return; 2042 DeadInsts.push_back(MI); 2043 return; 2044 } 2045 2046 // If the pointer is not the right type, insert a bitcast to the right 2047 // type. 2048 const Type *NewTy = 2049 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2050 2051 if (OtherPtr->getType() != NewTy) 2052 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2053 } 2054 2055 // Process each element of the aggregate. 2056 bool SROADest = MI->getRawDest() == Inst; 2057 2058 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2059 2060 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2061 // If this is a memcpy/memmove, emit a GEP of the other element address. 2062 Value *OtherElt = 0; 2063 unsigned OtherEltAlign = MemAlignment; 2064 2065 if (OtherPtr) { 2066 Value *Idx[2] = { Zero, 2067 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2068 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 2069 OtherPtr->getName()+"."+Twine(i), 2070 MI); 2071 uint64_t EltOffset; 2072 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2073 const Type *OtherTy = OtherPtrTy->getElementType(); 2074 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 2075 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2076 } else { 2077 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2078 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2079 } 2080 2081 // The alignment of the other pointer is the guaranteed alignment of the 2082 // element, which is affected by both the known alignment of the whole 2083 // mem intrinsic and the alignment of the element. If the alignment of 2084 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2085 // known alignment is just 4 bytes. 2086 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2087 } 2088 2089 Value *EltPtr = NewElts[i]; 2090 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2091 2092 // If we got down to a scalar, insert a load or store as appropriate. 2093 if (EltTy->isSingleValueType()) { 2094 if (isa<MemTransferInst>(MI)) { 2095 if (SROADest) { 2096 // From Other to Alloca. 2097 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2098 new StoreInst(Elt, EltPtr, MI); 2099 } else { 2100 // From Alloca to Other. 2101 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2102 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2103 } 2104 continue; 2105 } 2106 assert(isa<MemSetInst>(MI)); 2107 2108 // If the stored element is zero (common case), just store a null 2109 // constant. 2110 Constant *StoreVal; 2111 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2112 if (CI->isZero()) { 2113 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2114 } else { 2115 // If EltTy is a vector type, get the element type. 2116 const Type *ValTy = EltTy->getScalarType(); 2117 2118 // Construct an integer with the right value. 2119 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2120 APInt OneVal(EltSize, CI->getZExtValue()); 2121 APInt TotalVal(OneVal); 2122 // Set each byte. 2123 for (unsigned i = 0; 8*i < EltSize; ++i) { 2124 TotalVal = TotalVal.shl(8); 2125 TotalVal |= OneVal; 2126 } 2127 2128 // Convert the integer value to the appropriate type. 2129 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2130 if (ValTy->isPointerTy()) 2131 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2132 else if (ValTy->isFloatingPointTy()) 2133 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2134 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2135 2136 // If the requested value was a vector constant, create it. 2137 if (EltTy != ValTy) { 2138 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 2139 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2140 StoreVal = ConstantVector::get(Elts); 2141 } 2142 } 2143 new StoreInst(StoreVal, EltPtr, MI); 2144 continue; 2145 } 2146 // Otherwise, if we're storing a byte variable, use a memset call for 2147 // this element. 2148 } 2149 2150 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2151 2152 IRBuilder<> Builder(MI); 2153 2154 // Finally, insert the meminst for this element. 2155 if (isa<MemSetInst>(MI)) { 2156 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2157 MI->isVolatile()); 2158 } else { 2159 assert(isa<MemTransferInst>(MI)); 2160 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2161 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2162 2163 if (isa<MemCpyInst>(MI)) 2164 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2165 else 2166 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2167 } 2168 } 2169 DeadInsts.push_back(MI); 2170} 2171 2172/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2173/// overwrites the entire allocation. Extract out the pieces of the stored 2174/// integer and store them individually. 2175void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2176 SmallVector<AllocaInst*, 32> &NewElts){ 2177 // Extract each element out of the integer according to its structure offset 2178 // and store the element value to the individual alloca. 2179 Value *SrcVal = SI->getOperand(0); 2180 const Type *AllocaEltTy = AI->getAllocatedType(); 2181 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2182 2183 IRBuilder<> Builder(SI); 2184 2185 // Handle tail padding by extending the operand 2186 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2187 SrcVal = Builder.CreateZExt(SrcVal, 2188 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2189 2190 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2191 << '\n'); 2192 2193 // There are two forms here: AI could be an array or struct. Both cases 2194 // have different ways to compute the element offset. 2195 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2196 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2197 2198 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2199 // Get the number of bits to shift SrcVal to get the value. 2200 const Type *FieldTy = EltSTy->getElementType(i); 2201 uint64_t Shift = Layout->getElementOffsetInBits(i); 2202 2203 if (TD->isBigEndian()) 2204 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2205 2206 Value *EltVal = SrcVal; 2207 if (Shift) { 2208 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2209 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2210 } 2211 2212 // Truncate down to an integer of the right size. 2213 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2214 2215 // Ignore zero sized fields like {}, they obviously contain no data. 2216 if (FieldSizeBits == 0) continue; 2217 2218 if (FieldSizeBits != AllocaSizeBits) 2219 EltVal = Builder.CreateTrunc(EltVal, 2220 IntegerType::get(SI->getContext(), FieldSizeBits)); 2221 Value *DestField = NewElts[i]; 2222 if (EltVal->getType() == FieldTy) { 2223 // Storing to an integer field of this size, just do it. 2224 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2225 // Bitcast to the right element type (for fp/vector values). 2226 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2227 } else { 2228 // Otherwise, bitcast the dest pointer (for aggregates). 2229 DestField = Builder.CreateBitCast(DestField, 2230 PointerType::getUnqual(EltVal->getType())); 2231 } 2232 new StoreInst(EltVal, DestField, SI); 2233 } 2234 2235 } else { 2236 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2237 const Type *ArrayEltTy = ATy->getElementType(); 2238 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2239 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2240 2241 uint64_t Shift; 2242 2243 if (TD->isBigEndian()) 2244 Shift = AllocaSizeBits-ElementOffset; 2245 else 2246 Shift = 0; 2247 2248 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2249 // Ignore zero sized fields like {}, they obviously contain no data. 2250 if (ElementSizeBits == 0) continue; 2251 2252 Value *EltVal = SrcVal; 2253 if (Shift) { 2254 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2255 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2256 } 2257 2258 // Truncate down to an integer of the right size. 2259 if (ElementSizeBits != AllocaSizeBits) 2260 EltVal = Builder.CreateTrunc(EltVal, 2261 IntegerType::get(SI->getContext(), 2262 ElementSizeBits)); 2263 Value *DestField = NewElts[i]; 2264 if (EltVal->getType() == ArrayEltTy) { 2265 // Storing to an integer field of this size, just do it. 2266 } else if (ArrayEltTy->isFloatingPointTy() || 2267 ArrayEltTy->isVectorTy()) { 2268 // Bitcast to the right element type (for fp/vector values). 2269 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2270 } else { 2271 // Otherwise, bitcast the dest pointer (for aggregates). 2272 DestField = Builder.CreateBitCast(DestField, 2273 PointerType::getUnqual(EltVal->getType())); 2274 } 2275 new StoreInst(EltVal, DestField, SI); 2276 2277 if (TD->isBigEndian()) 2278 Shift -= ElementOffset; 2279 else 2280 Shift += ElementOffset; 2281 } 2282 } 2283 2284 DeadInsts.push_back(SI); 2285} 2286 2287/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2288/// an integer. Load the individual pieces to form the aggregate value. 2289void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2290 SmallVector<AllocaInst*, 32> &NewElts) { 2291 // Extract each element out of the NewElts according to its structure offset 2292 // and form the result value. 2293 const Type *AllocaEltTy = AI->getAllocatedType(); 2294 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2295 2296 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2297 << '\n'); 2298 2299 // There are two forms here: AI could be an array or struct. Both cases 2300 // have different ways to compute the element offset. 2301 const StructLayout *Layout = 0; 2302 uint64_t ArrayEltBitOffset = 0; 2303 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2304 Layout = TD->getStructLayout(EltSTy); 2305 } else { 2306 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2307 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2308 } 2309 2310 Value *ResultVal = 2311 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2312 2313 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2314 // Load the value from the alloca. If the NewElt is an aggregate, cast 2315 // the pointer to an integer of the same size before doing the load. 2316 Value *SrcField = NewElts[i]; 2317 const Type *FieldTy = 2318 cast<PointerType>(SrcField->getType())->getElementType(); 2319 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2320 2321 // Ignore zero sized fields like {}, they obviously contain no data. 2322 if (FieldSizeBits == 0) continue; 2323 2324 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2325 FieldSizeBits); 2326 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2327 !FieldTy->isVectorTy()) 2328 SrcField = new BitCastInst(SrcField, 2329 PointerType::getUnqual(FieldIntTy), 2330 "", LI); 2331 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2332 2333 // If SrcField is a fp or vector of the right size but that isn't an 2334 // integer type, bitcast to an integer so we can shift it. 2335 if (SrcField->getType() != FieldIntTy) 2336 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2337 2338 // Zero extend the field to be the same size as the final alloca so that 2339 // we can shift and insert it. 2340 if (SrcField->getType() != ResultVal->getType()) 2341 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2342 2343 // Determine the number of bits to shift SrcField. 2344 uint64_t Shift; 2345 if (Layout) // Struct case. 2346 Shift = Layout->getElementOffsetInBits(i); 2347 else // Array case. 2348 Shift = i*ArrayEltBitOffset; 2349 2350 if (TD->isBigEndian()) 2351 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2352 2353 if (Shift) { 2354 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2355 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2356 } 2357 2358 // Don't create an 'or x, 0' on the first iteration. 2359 if (!isa<Constant>(ResultVal) || 2360 !cast<Constant>(ResultVal)->isNullValue()) 2361 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2362 else 2363 ResultVal = SrcField; 2364 } 2365 2366 // Handle tail padding by truncating the result 2367 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2368 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2369 2370 LI->replaceAllUsesWith(ResultVal); 2371 DeadInsts.push_back(LI); 2372} 2373 2374/// HasPadding - Return true if the specified type has any structure or 2375/// alignment padding in between the elements that would be split apart 2376/// by SROA; return false otherwise. 2377static bool HasPadding(const Type *Ty, const TargetData &TD) { 2378 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2379 Ty = ATy->getElementType(); 2380 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2381 } 2382 2383 // SROA currently handles only Arrays and Structs. 2384 const StructType *STy = cast<StructType>(Ty); 2385 const StructLayout *SL = TD.getStructLayout(STy); 2386 unsigned PrevFieldBitOffset = 0; 2387 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2388 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2389 2390 // Check to see if there is any padding between this element and the 2391 // previous one. 2392 if (i) { 2393 unsigned PrevFieldEnd = 2394 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2395 if (PrevFieldEnd < FieldBitOffset) 2396 return true; 2397 } 2398 PrevFieldBitOffset = FieldBitOffset; 2399 } 2400 // Check for tail padding. 2401 if (unsigned EltCount = STy->getNumElements()) { 2402 unsigned PrevFieldEnd = PrevFieldBitOffset + 2403 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2404 if (PrevFieldEnd < SL->getSizeInBits()) 2405 return true; 2406 } 2407 return false; 2408} 2409 2410/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2411/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2412/// or 1 if safe after canonicalization has been performed. 2413bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2414 // Loop over the use list of the alloca. We can only transform it if all of 2415 // the users are safe to transform. 2416 AllocaInfo Info(AI); 2417 2418 isSafeForScalarRepl(AI, 0, Info); 2419 if (Info.isUnsafe) { 2420 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2421 return false; 2422 } 2423 2424 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2425 // source and destination, we have to be careful. In particular, the memcpy 2426 // could be moving around elements that live in structure padding of the LLVM 2427 // types, but may actually be used. In these cases, we refuse to promote the 2428 // struct. 2429 if (Info.isMemCpySrc && Info.isMemCpyDst && 2430 HasPadding(AI->getAllocatedType(), *TD)) 2431 return false; 2432 2433 // If the alloca never has an access to just *part* of it, but is accessed 2434 // via loads and stores, then we should use ConvertToScalarInfo to promote 2435 // the alloca instead of promoting each piece at a time and inserting fission 2436 // and fusion code. 2437 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2438 // If the struct/array just has one element, use basic SRoA. 2439 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2440 if (ST->getNumElements() > 1) return false; 2441 } else { 2442 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2443 return false; 2444 } 2445 } 2446 2447 return true; 2448} 2449 2450 2451 2452/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2453/// some part of a constant global variable. This intentionally only accepts 2454/// constant expressions because we don't can't rewrite arbitrary instructions. 2455static bool PointsToConstantGlobal(Value *V) { 2456 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2457 return GV->isConstant(); 2458 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2459 if (CE->getOpcode() == Instruction::BitCast || 2460 CE->getOpcode() == Instruction::GetElementPtr) 2461 return PointsToConstantGlobal(CE->getOperand(0)); 2462 return false; 2463} 2464 2465/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2466/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2467/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2468/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2469/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2470/// the alloca, and if the source pointer is a pointer to a constant global, we 2471/// can optimize this. 2472static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2473 bool isOffset) { 2474 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2475 User *U = cast<Instruction>(*UI); 2476 2477 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2478 // Ignore non-volatile loads, they are always ok. 2479 if (LI->isVolatile()) return false; 2480 continue; 2481 } 2482 2483 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2484 // If uses of the bitcast are ok, we are ok. 2485 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 2486 return false; 2487 continue; 2488 } 2489 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2490 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2491 // doesn't, it does. 2492 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2493 isOffset || !GEP->hasAllZeroIndices())) 2494 return false; 2495 continue; 2496 } 2497 2498 if (CallSite CS = U) { 2499 // If this is the function being called then we treat it like a load and 2500 // ignore it. 2501 if (CS.isCallee(UI)) 2502 continue; 2503 2504 // If this is a readonly/readnone call site, then we know it is just a 2505 // load (but one that potentially returns the value itself), so we can 2506 // ignore it if we know that the value isn't captured. 2507 unsigned ArgNo = CS.getArgumentNo(UI); 2508 if (CS.onlyReadsMemory() && 2509 (CS.getInstruction()->use_empty() || 2510 CS.paramHasAttr(ArgNo+1, Attribute::NoCapture))) 2511 continue; 2512 2513 // If this is being passed as a byval argument, the caller is making a 2514 // copy, so it is only a read of the alloca. 2515 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2516 continue; 2517 } 2518 2519 // If this is isn't our memcpy/memmove, reject it as something we can't 2520 // handle. 2521 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2522 if (MI == 0) 2523 return false; 2524 2525 // If the transfer is using the alloca as a source of the transfer, then 2526 // ignore it since it is a load (unless the transfer is volatile). 2527 if (UI.getOperandNo() == 1) { 2528 if (MI->isVolatile()) return false; 2529 continue; 2530 } 2531 2532 // If we already have seen a copy, reject the second one. 2533 if (TheCopy) return false; 2534 2535 // If the pointer has been offset from the start of the alloca, we can't 2536 // safely handle this. 2537 if (isOffset) return false; 2538 2539 // If the memintrinsic isn't using the alloca as the dest, reject it. 2540 if (UI.getOperandNo() != 0) return false; 2541 2542 // If the source of the memcpy/move is not a constant global, reject it. 2543 if (!PointsToConstantGlobal(MI->getSource())) 2544 return false; 2545 2546 // Otherwise, the transform is safe. Remember the copy instruction. 2547 TheCopy = MI; 2548 } 2549 return true; 2550} 2551 2552/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2553/// modified by a copy from a constant global. If we can prove this, we can 2554/// replace any uses of the alloca with uses of the global directly. 2555MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 2556 MemTransferInst *TheCopy = 0; 2557 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 2558 return TheCopy; 2559 return 0; 2560} 2561