ScalarReplAggregates.cpp revision abb25122050349ccf77f8afb198a985e7a4d95ee
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 if (!VectorTy) { 347 VectorTy = VectorType::get(In, AllocaSize/EltSize); 348 return; 349 } 350 351 unsigned CurrentEltSize = cast<VectorType>(VectorTy)->getElementType() 352 ->getPrimitiveSizeInBits()/8; 353 if (EltSize == CurrentEltSize) 354 return; 355 356 if (In->isIntegerTy() && isPowerOf2_32(AllocaSize / EltSize)) 357 return; 358 } 359 } 360 361 // Otherwise, we have a case that we can't handle with an optimized vector 362 // form. We can still turn this into a large integer. 363 VectorTy = Type::getVoidTy(In->getContext()); 364} 365 366/// MergeInVectorType - Handles the vector case of MergeInType, returning true 367/// if the type was successfully merged and false otherwise. 368bool ConvertToScalarInfo::MergeInVectorType(const VectorType *VInTy, 369 uint64_t Offset) { 370 // Remember if we saw a vector type. 371 HadAVector = true; 372 373 // TODO: Support nonzero offsets? 374 if (Offset != 0) 375 return false; 376 377 // Only allow vectors that are a power-of-2 away from the size of the alloca. 378 if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8))) 379 return false; 380 381 // If this the first vector we see, remember the type so that we know the 382 // element size. 383 if (!VectorTy) { 384 VectorTy = VInTy; 385 return true; 386 } 387 388 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth(); 389 unsigned InBitWidth = VInTy->getBitWidth(); 390 391 // Vectors of the same size can be converted using a simple bitcast. 392 if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8)) 393 return true; 394 395 const Type *ElementTy = cast<VectorType>(VectorTy)->getElementType(); 396 const Type *InElementTy = cast<VectorType>(VInTy)->getElementType(); 397 398 // Do not allow mixed integer and floating-point accesses from vectors of 399 // different sizes. 400 if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy()) 401 return false; 402 403 if (ElementTy->isFloatingPointTy()) { 404 // Only allow floating-point vectors of different sizes if they have the 405 // same element type. 406 // TODO: This could be loosened a bit, but would anything benefit? 407 if (ElementTy != InElementTy) 408 return false; 409 410 // There are no arbitrary-precision floating-point types, which limits the 411 // number of legal vector types with larger element types that we can form 412 // to bitcast and extract a subvector. 413 // TODO: We could support some more cases with mixed fp128 and double here. 414 if (!(BitWidth == 64 || BitWidth == 128) || 415 !(InBitWidth == 64 || InBitWidth == 128)) 416 return false; 417 } else { 418 assert(ElementTy->isIntegerTy() && "Vector elements must be either integer " 419 "or floating-point."); 420 unsigned BitWidth = ElementTy->getPrimitiveSizeInBits(); 421 unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits(); 422 423 // Do not allow integer types smaller than a byte or types whose widths are 424 // not a multiple of a byte. 425 if (BitWidth < 8 || InBitWidth < 8 || 426 BitWidth % 8 != 0 || InBitWidth % 8 != 0) 427 return false; 428 } 429 430 // Pick the largest of the two vector types. 431 if (InBitWidth > BitWidth) 432 VectorTy = VInTy; 433 434 return true; 435} 436 437/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 438/// its accesses to a single vector type, return true and set VecTy to 439/// the new type. If we could convert the alloca into a single promotable 440/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 441/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 442/// is the current offset from the base of the alloca being analyzed. 443/// 444/// If we see at least one access to the value that is as a vector type, set the 445/// SawVec flag. 446bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 447 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 448 Instruction *User = cast<Instruction>(*UI); 449 450 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 451 // Don't break volatile loads. 452 if (LI->isVolatile()) 453 return false; 454 // Don't touch MMX operations. 455 if (LI->getType()->isX86_MMXTy()) 456 return false; 457 HadNonMemTransferAccess = true; 458 MergeInType(LI->getType(), Offset, true); 459 continue; 460 } 461 462 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 463 // Storing the pointer, not into the value? 464 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 465 // Don't touch MMX operations. 466 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 467 return false; 468 HadNonMemTransferAccess = true; 469 MergeInType(SI->getOperand(0)->getType(), Offset, true); 470 continue; 471 } 472 473 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 474 IsNotTrivial = true; // Can't be mem2reg'd. 475 if (!CanConvertToScalar(BCI, Offset)) 476 return false; 477 continue; 478 } 479 480 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 481 // If this is a GEP with a variable indices, we can't handle it. 482 if (!GEP->hasAllConstantIndices()) 483 return false; 484 485 // Compute the offset that this GEP adds to the pointer. 486 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 487 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 488 &Indices[0], Indices.size()); 489 // See if all uses can be converted. 490 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 491 return false; 492 IsNotTrivial = true; // Can't be mem2reg'd. 493 HadNonMemTransferAccess = true; 494 continue; 495 } 496 497 // If this is a constant sized memset of a constant value (e.g. 0) we can 498 // handle it. 499 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 500 // Store of constant value and constant size. 501 if (!isa<ConstantInt>(MSI->getValue()) || 502 !isa<ConstantInt>(MSI->getLength())) 503 return false; 504 IsNotTrivial = true; // Can't be mem2reg'd. 505 HadNonMemTransferAccess = true; 506 continue; 507 } 508 509 // If this is a memcpy or memmove into or out of the whole allocation, we 510 // can handle it like a load or store of the scalar type. 511 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 512 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 513 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 514 return false; 515 516 IsNotTrivial = true; // Can't be mem2reg'd. 517 continue; 518 } 519 520 // Otherwise, we cannot handle this! 521 return false; 522 } 523 524 return true; 525} 526 527/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 528/// directly. This happens when we are converting an "integer union" to a 529/// single integer scalar, or when we are converting a "vector union" to a 530/// vector with insert/extractelement instructions. 531/// 532/// Offset is an offset from the original alloca, in bits that need to be 533/// shifted to the right. By the end of this, there should be no uses of Ptr. 534void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 535 uint64_t Offset) { 536 while (!Ptr->use_empty()) { 537 Instruction *User = cast<Instruction>(Ptr->use_back()); 538 539 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 540 ConvertUsesToScalar(CI, NewAI, Offset); 541 CI->eraseFromParent(); 542 continue; 543 } 544 545 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 546 // Compute the offset that this GEP adds to the pointer. 547 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 548 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 549 &Indices[0], Indices.size()); 550 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 551 GEP->eraseFromParent(); 552 continue; 553 } 554 555 IRBuilder<> Builder(User); 556 557 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 558 // The load is a bit extract from NewAI shifted right by Offset bits. 559 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 560 Value *NewLoadVal 561 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 562 LI->replaceAllUsesWith(NewLoadVal); 563 LI->eraseFromParent(); 564 continue; 565 } 566 567 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 568 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 569 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 570 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 571 Builder); 572 Builder.CreateStore(New, NewAI); 573 SI->eraseFromParent(); 574 575 // If the load we just inserted is now dead, then the inserted store 576 // overwrote the entire thing. 577 if (Old->use_empty()) 578 Old->eraseFromParent(); 579 continue; 580 } 581 582 // If this is a constant sized memset of a constant value (e.g. 0) we can 583 // transform it into a store of the expanded constant value. 584 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 585 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 586 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 587 if (NumBytes != 0) { 588 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 589 590 // Compute the value replicated the right number of times. 591 APInt APVal(NumBytes*8, Val); 592 593 // Splat the value if non-zero. 594 if (Val) 595 for (unsigned i = 1; i != NumBytes; ++i) 596 APVal |= APVal << 8; 597 598 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 599 Value *New = ConvertScalar_InsertValue( 600 ConstantInt::get(User->getContext(), APVal), 601 Old, Offset, Builder); 602 Builder.CreateStore(New, NewAI); 603 604 // If the load we just inserted is now dead, then the memset overwrote 605 // the entire thing. 606 if (Old->use_empty()) 607 Old->eraseFromParent(); 608 } 609 MSI->eraseFromParent(); 610 continue; 611 } 612 613 // If this is a memcpy or memmove into or out of the whole allocation, we 614 // can handle it like a load or store of the scalar type. 615 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 616 assert(Offset == 0 && "must be store to start of alloca"); 617 618 // If the source and destination are both to the same alloca, then this is 619 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 620 // as appropriate. 621 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 622 623 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 624 // Dest must be OrigAI, change this to be a load from the original 625 // pointer (bitcasted), then a store to our new alloca. 626 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 627 Value *SrcPtr = MTI->getSource(); 628 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 629 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 630 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 631 AIPTy = PointerType::get(AIPTy->getElementType(), 632 SPTy->getAddressSpace()); 633 } 634 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 635 636 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 637 SrcVal->setAlignment(MTI->getAlignment()); 638 Builder.CreateStore(SrcVal, NewAI); 639 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 640 // Src must be OrigAI, change this to be a load from NewAI then a store 641 // through the original dest pointer (bitcasted). 642 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 643 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 644 645 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 646 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 647 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 648 AIPTy = PointerType::get(AIPTy->getElementType(), 649 DPTy->getAddressSpace()); 650 } 651 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 652 653 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 654 NewStore->setAlignment(MTI->getAlignment()); 655 } else { 656 // Noop transfer. Src == Dst 657 } 658 659 MTI->eraseFromParent(); 660 continue; 661 } 662 663 llvm_unreachable("Unsupported operation!"); 664 } 665} 666 667/// getScaledElementType - Gets a scaled element type for a partial vector 668/// access of an alloca. The input types must be integer or floating-point 669/// scalar or vector types, and the resulting type is an integer, float or 670/// double. 671static const Type *getScaledElementType(const Type *Ty1, const Type *Ty2, 672 unsigned NewBitWidth) { 673 bool IsFP1 = Ty1->isFloatingPointTy() || 674 (Ty1->isVectorTy() && 675 cast<VectorType>(Ty1)->getElementType()->isFloatingPointTy()); 676 bool IsFP2 = Ty2->isFloatingPointTy() || 677 (Ty2->isVectorTy() && 678 cast<VectorType>(Ty2)->getElementType()->isFloatingPointTy()); 679 680 LLVMContext &Context = Ty1->getContext(); 681 682 // Prefer floating-point types over integer types, as integer types may have 683 // been created by earlier scalar replacement. 684 if (IsFP1 || IsFP2) { 685 if (NewBitWidth == 32) 686 return Type::getFloatTy(Context); 687 if (NewBitWidth == 64) 688 return Type::getDoubleTy(Context); 689 } 690 691 return Type::getIntNTy(Context, NewBitWidth); 692} 693 694/// CreateShuffleVectorCast - Creates a shuffle vector to convert one vector 695/// to another vector of the same element type which has the same allocation 696/// size but different primitive sizes (e.g. <3 x i32> and <4 x i32>). 697static Value *CreateShuffleVectorCast(Value *FromVal, const Type *ToType, 698 IRBuilder<> &Builder) { 699 const Type *FromType = FromVal->getType(); 700 const VectorType *FromVTy = cast<VectorType>(FromType); 701 const VectorType *ToVTy = cast<VectorType>(ToType); 702 assert((ToVTy->getElementType() == FromVTy->getElementType()) && 703 "Vectors must have the same element type"); 704 Value *UnV = UndefValue::get(FromType); 705 unsigned numEltsFrom = FromVTy->getNumElements(); 706 unsigned numEltsTo = ToVTy->getNumElements(); 707 708 SmallVector<Constant*, 3> Args; 709 const Type* Int32Ty = Builder.getInt32Ty(); 710 unsigned minNumElts = std::min(numEltsFrom, numEltsTo); 711 unsigned i; 712 for (i=0; i != minNumElts; ++i) 713 Args.push_back(ConstantInt::get(Int32Ty, i)); 714 715 if (i < numEltsTo) { 716 Constant* UnC = UndefValue::get(Int32Ty); 717 for (; i != numEltsTo; ++i) 718 Args.push_back(UnC); 719 } 720 Constant *Mask = ConstantVector::get(Args); 721 return Builder.CreateShuffleVector(FromVal, UnV, Mask, "tmpV"); 722} 723 724/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 725/// or vector value FromVal, extracting the bits from the offset specified by 726/// Offset. This returns the value, which is of type ToType. 727/// 728/// This happens when we are converting an "integer union" to a single 729/// integer scalar, or when we are converting a "vector union" to a vector with 730/// insert/extractelement instructions. 731/// 732/// Offset is an offset from the original alloca, in bits that need to be 733/// shifted to the right. 734Value *ConvertToScalarInfo:: 735ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 736 uint64_t Offset, IRBuilder<> &Builder) { 737 // If the load is of the whole new alloca, no conversion is needed. 738 const Type *FromType = FromVal->getType(); 739 if (FromType == ToType && Offset == 0) 740 return FromVal; 741 742 // If the result alloca is a vector type, this is either an element 743 // access or a bitcast to another vector type of the same size. 744 if (const VectorType *VTy = dyn_cast<VectorType>(FromType)) { 745 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 746 if (ToTypeSize == AllocaSize) { 747 // If the two types have the same primitive size, use a bit cast. 748 // Otherwise, it is two vectors with the same element type that has 749 // the same allocation size but different number of elements so use 750 // a shuffle vector. 751 if (FromType->getPrimitiveSizeInBits() == 752 ToType->getPrimitiveSizeInBits()) 753 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 754 else 755 return CreateShuffleVectorCast(FromVal, ToType, Builder); 756 } 757 758 if (isPowerOf2_64(AllocaSize / ToTypeSize)) { 759 assert(!(ToType->isVectorTy() && Offset != 0) && "Can't extract a value " 760 "of a smaller vector type at a nonzero offset."); 761 762 const Type *CastElementTy = getScaledElementType(FromType, ToType, 763 ToTypeSize * 8); 764 unsigned NumCastVectorElements = AllocaSize / ToTypeSize; 765 766 LLVMContext &Context = FromVal->getContext(); 767 const Type *CastTy = VectorType::get(CastElementTy, 768 NumCastVectorElements); 769 Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp"); 770 771 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); 772 unsigned Elt = Offset/EltSize; 773 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 774 Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get( 775 Type::getInt32Ty(Context), Elt), "tmp"); 776 return Builder.CreateBitCast(Extract, ToType, "tmp"); 777 } 778 779 // Otherwise it must be an element access. 780 unsigned Elt = 0; 781 if (Offset) { 782 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 783 Elt = Offset/EltSize; 784 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 785 } 786 // Return the element extracted out of it. 787 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 788 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 789 if (V->getType() != ToType) 790 V = Builder.CreateBitCast(V, ToType, "tmp"); 791 return V; 792 } 793 794 // If ToType is a first class aggregate, extract out each of the pieces and 795 // use insertvalue's to form the FCA. 796 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 797 const StructLayout &Layout = *TD.getStructLayout(ST); 798 Value *Res = UndefValue::get(ST); 799 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 800 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 801 Offset+Layout.getElementOffsetInBits(i), 802 Builder); 803 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 804 } 805 return Res; 806 } 807 808 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 809 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 810 Value *Res = UndefValue::get(AT); 811 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 812 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 813 Offset+i*EltSize, Builder); 814 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 815 } 816 return Res; 817 } 818 819 // Otherwise, this must be a union that was converted to an integer value. 820 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 821 822 // If this is a big-endian system and the load is narrower than the 823 // full alloca type, we need to do a shift to get the right bits. 824 int ShAmt = 0; 825 if (TD.isBigEndian()) { 826 // On big-endian machines, the lowest bit is stored at the bit offset 827 // from the pointer given by getTypeStoreSizeInBits. This matters for 828 // integers with a bitwidth that is not a multiple of 8. 829 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 830 TD.getTypeStoreSizeInBits(ToType) - Offset; 831 } else { 832 ShAmt = Offset; 833 } 834 835 // Note: we support negative bitwidths (with shl) which are not defined. 836 // We do this to support (f.e.) loads off the end of a structure where 837 // only some bits are used. 838 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 839 FromVal = Builder.CreateLShr(FromVal, 840 ConstantInt::get(FromVal->getType(), 841 ShAmt), "tmp"); 842 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 843 FromVal = Builder.CreateShl(FromVal, 844 ConstantInt::get(FromVal->getType(), 845 -ShAmt), "tmp"); 846 847 // Finally, unconditionally truncate the integer to the right width. 848 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 849 if (LIBitWidth < NTy->getBitWidth()) 850 FromVal = 851 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 852 LIBitWidth), "tmp"); 853 else if (LIBitWidth > NTy->getBitWidth()) 854 FromVal = 855 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 856 LIBitWidth), "tmp"); 857 858 // If the result is an integer, this is a trunc or bitcast. 859 if (ToType->isIntegerTy()) { 860 // Should be done. 861 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 862 // Just do a bitcast, we know the sizes match up. 863 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 864 } else { 865 // Otherwise must be a pointer. 866 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 867 } 868 assert(FromVal->getType() == ToType && "Didn't convert right?"); 869 return FromVal; 870} 871 872/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 873/// or vector value "Old" at the offset specified by Offset. 874/// 875/// This happens when we are converting an "integer union" to a 876/// single integer scalar, or when we are converting a "vector union" to a 877/// vector with insert/extractelement instructions. 878/// 879/// Offset is an offset from the original alloca, in bits that need to be 880/// shifted to the right. 881Value *ConvertToScalarInfo:: 882ConvertScalar_InsertValue(Value *SV, Value *Old, 883 uint64_t Offset, IRBuilder<> &Builder) { 884 // Convert the stored type to the actual type, shift it left to insert 885 // then 'or' into place. 886 const Type *AllocaType = Old->getType(); 887 LLVMContext &Context = Old->getContext(); 888 889 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 890 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 891 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 892 893 // Changing the whole vector with memset or with an access of a different 894 // vector type? 895 if (ValSize == VecSize) { 896 // If the two types have the same primitive size, use a bit cast. 897 // Otherwise, it is two vectors with the same element type that has 898 // the same allocation size but different number of elements so use 899 // a shuffle vector. 900 if (VTy->getPrimitiveSizeInBits() == 901 SV->getType()->getPrimitiveSizeInBits()) 902 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 903 else 904 return CreateShuffleVectorCast(SV, VTy, Builder); 905 } 906 907 if (isPowerOf2_64(VecSize / ValSize)) { 908 assert(!(SV->getType()->isVectorTy() && Offset != 0) && "Can't insert a " 909 "value of a smaller vector type at a nonzero offset."); 910 911 const Type *CastElementTy = getScaledElementType(VTy, SV->getType(), 912 ValSize); 913 unsigned NumCastVectorElements = VecSize / ValSize; 914 915 LLVMContext &Context = SV->getContext(); 916 const Type *OldCastTy = VectorType::get(CastElementTy, 917 NumCastVectorElements); 918 Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp"); 919 920 Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp"); 921 922 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); 923 unsigned Elt = Offset/EltSize; 924 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 925 Value *Insert = 926 Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get( 927 Type::getInt32Ty(Context), Elt), "tmp"); 928 return Builder.CreateBitCast(Insert, AllocaType, "tmp"); 929 } 930 931 // Must be an element insertion. 932 assert(SV->getType() == VTy->getElementType()); 933 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 934 unsigned Elt = Offset/EltSize; 935 return Builder.CreateInsertElement(Old, SV, 936 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 937 "tmp"); 938 } 939 940 // If SV is a first-class aggregate value, insert each value recursively. 941 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 942 const StructLayout &Layout = *TD.getStructLayout(ST); 943 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 944 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 945 Old = ConvertScalar_InsertValue(Elt, Old, 946 Offset+Layout.getElementOffsetInBits(i), 947 Builder); 948 } 949 return Old; 950 } 951 952 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 953 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 954 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 955 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 956 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 957 } 958 return Old; 959 } 960 961 // If SV is a float, convert it to the appropriate integer type. 962 // If it is a pointer, do the same. 963 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 964 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 965 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 966 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 967 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 968 SV = Builder.CreateBitCast(SV, 969 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 970 else if (SV->getType()->isPointerTy()) 971 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 972 973 // Zero extend or truncate the value if needed. 974 if (SV->getType() != AllocaType) { 975 if (SV->getType()->getPrimitiveSizeInBits() < 976 AllocaType->getPrimitiveSizeInBits()) 977 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 978 else { 979 // Truncation may be needed if storing more than the alloca can hold 980 // (undefined behavior). 981 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 982 SrcWidth = DestWidth; 983 SrcStoreWidth = DestStoreWidth; 984 } 985 } 986 987 // If this is a big-endian system and the store is narrower than the 988 // full alloca type, we need to do a shift to get the right bits. 989 int ShAmt = 0; 990 if (TD.isBigEndian()) { 991 // On big-endian machines, the lowest bit is stored at the bit offset 992 // from the pointer given by getTypeStoreSizeInBits. This matters for 993 // integers with a bitwidth that is not a multiple of 8. 994 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 995 } else { 996 ShAmt = Offset; 997 } 998 999 // Note: we support negative bitwidths (with shr) which are not defined. 1000 // We do this to support (f.e.) stores off the end of a structure where 1001 // only some bits in the structure are set. 1002 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1003 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1004 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1005 ShAmt), "tmp"); 1006 Mask <<= ShAmt; 1007 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1008 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1009 -ShAmt), "tmp"); 1010 Mask = Mask.lshr(-ShAmt); 1011 } 1012 1013 // Mask out the bits we are about to insert from the old value, and or 1014 // in the new bits. 1015 if (SrcWidth != DestWidth) { 1016 assert(DestWidth > SrcWidth); 1017 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1018 SV = Builder.CreateOr(Old, SV, "ins"); 1019 } 1020 return SV; 1021} 1022 1023 1024//===----------------------------------------------------------------------===// 1025// SRoA Driver 1026//===----------------------------------------------------------------------===// 1027 1028 1029bool SROA::runOnFunction(Function &F) { 1030 TD = getAnalysisIfAvailable<TargetData>(); 1031 1032 bool Changed = performPromotion(F); 1033 1034 // FIXME: ScalarRepl currently depends on TargetData more than it 1035 // theoretically needs to. It should be refactored in order to support 1036 // target-independent IR. Until this is done, just skip the actual 1037 // scalar-replacement portion of this pass. 1038 if (!TD) return Changed; 1039 1040 while (1) { 1041 bool LocalChange = performScalarRepl(F); 1042 if (!LocalChange) break; // No need to repromote if no scalarrepl 1043 Changed = true; 1044 LocalChange = performPromotion(F); 1045 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 1046 } 1047 1048 return Changed; 1049} 1050 1051namespace { 1052class AllocaPromoter : public LoadAndStorePromoter { 1053 AllocaInst *AI; 1054public: 1055 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 1056 DbgDeclareInst *DD, DIBuilder *&DB) 1057 : LoadAndStorePromoter(Insts, S, DD, DB), AI(0) {} 1058 1059 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 1060 // Remember which alloca we're promoting (for isInstInList). 1061 this->AI = AI; 1062 LoadAndStorePromoter::run(Insts); 1063 AI->eraseFromParent(); 1064 } 1065 1066 virtual bool isInstInList(Instruction *I, 1067 const SmallVectorImpl<Instruction*> &Insts) const { 1068 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1069 return LI->getOperand(0) == AI; 1070 return cast<StoreInst>(I)->getPointerOperand() == AI; 1071 } 1072}; 1073} // end anon namespace 1074 1075/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1076/// subsequently loaded can be rewritten to load both input pointers and then 1077/// select between the result, allowing the load of the alloca to be promoted. 1078/// From this: 1079/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1080/// %V = load i32* %P2 1081/// to: 1082/// %V1 = load i32* %Alloca -> will be mem2reg'd 1083/// %V2 = load i32* %Other 1084/// %V = select i1 %cond, i32 %V1, i32 %V2 1085/// 1086/// We can do this to a select if its only uses are loads and if the operand to 1087/// the select can be loaded unconditionally. 1088static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1089 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1090 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1091 1092 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1093 UI != UE; ++UI) { 1094 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1095 if (LI == 0 || LI->isVolatile()) return false; 1096 1097 // Both operands to the select need to be dereferencable, either absolutely 1098 // (e.g. allocas) or at this point because we can see other accesses to it. 1099 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1100 LI->getAlignment(), TD)) 1101 return false; 1102 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1103 LI->getAlignment(), TD)) 1104 return false; 1105 } 1106 1107 return true; 1108} 1109 1110/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1111/// subsequently loaded can be rewritten to load both input pointers in the pred 1112/// blocks and then PHI the results, allowing the load of the alloca to be 1113/// promoted. 1114/// From this: 1115/// %P2 = phi [i32* %Alloca, i32* %Other] 1116/// %V = load i32* %P2 1117/// to: 1118/// %V1 = load i32* %Alloca -> will be mem2reg'd 1119/// ... 1120/// %V2 = load i32* %Other 1121/// ... 1122/// %V = phi [i32 %V1, i32 %V2] 1123/// 1124/// We can do this to a select if its only uses are loads and if the operand to 1125/// the select can be loaded unconditionally. 1126static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1127 // For now, we can only do this promotion if the load is in the same block as 1128 // the PHI, and if there are no stores between the phi and load. 1129 // TODO: Allow recursive phi users. 1130 // TODO: Allow stores. 1131 BasicBlock *BB = PN->getParent(); 1132 unsigned MaxAlign = 0; 1133 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1134 UI != UE; ++UI) { 1135 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1136 if (LI == 0 || LI->isVolatile()) return false; 1137 1138 // For now we only allow loads in the same block as the PHI. This is a 1139 // common case that happens when instcombine merges two loads through a PHI. 1140 if (LI->getParent() != BB) return false; 1141 1142 // Ensure that there are no instructions between the PHI and the load that 1143 // could store. 1144 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1145 if (BBI->mayWriteToMemory()) 1146 return false; 1147 1148 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1149 } 1150 1151 // Okay, we know that we have one or more loads in the same block as the PHI. 1152 // We can transform this if it is safe to push the loads into the predecessor 1153 // blocks. The only thing to watch out for is that we can't put a possibly 1154 // trapping load in the predecessor if it is a critical edge. 1155 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1156 BasicBlock *Pred = PN->getIncomingBlock(i); 1157 1158 // If the predecessor has a single successor, then the edge isn't critical. 1159 if (Pred->getTerminator()->getNumSuccessors() == 1) 1160 continue; 1161 1162 Value *InVal = PN->getIncomingValue(i); 1163 1164 // If the InVal is an invoke in the pred, we can't put a load on the edge. 1165 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 1166 if (II->getParent() == Pred) 1167 return false; 1168 1169 // If this pointer is always safe to load, or if we can prove that there is 1170 // already a load in the block, then we can move the load to the pred block. 1171 if (InVal->isDereferenceablePointer() || 1172 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1173 continue; 1174 1175 return false; 1176 } 1177 1178 return true; 1179} 1180 1181 1182/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1183/// direct (non-volatile) loads and stores to it. If the alloca is close but 1184/// not quite there, this will transform the code to allow promotion. As such, 1185/// it is a non-pure predicate. 1186static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1187 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1188 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1189 1190 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1191 UI != UE; ++UI) { 1192 User *U = *UI; 1193 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1194 if (LI->isVolatile()) 1195 return false; 1196 continue; 1197 } 1198 1199 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1200 if (SI->getOperand(0) == AI || SI->isVolatile()) 1201 return false; // Don't allow a store OF the AI, only INTO the AI. 1202 continue; 1203 } 1204 1205 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1206 // If the condition being selected on is a constant, fold the select, yes 1207 // this does (rarely) happen early on. 1208 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1209 Value *Result = SI->getOperand(1+CI->isZero()); 1210 SI->replaceAllUsesWith(Result); 1211 SI->eraseFromParent(); 1212 1213 // This is very rare and we just scrambled the use list of AI, start 1214 // over completely. 1215 return tryToMakeAllocaBePromotable(AI, TD); 1216 } 1217 1218 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1219 // loads, then we can transform this by rewriting the select. 1220 if (!isSafeSelectToSpeculate(SI, TD)) 1221 return false; 1222 1223 InstsToRewrite.insert(SI); 1224 continue; 1225 } 1226 1227 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1228 if (PN->use_empty()) { // Dead PHIs can be stripped. 1229 InstsToRewrite.insert(PN); 1230 continue; 1231 } 1232 1233 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1234 // in the pred blocks, then we can transform this by rewriting the PHI. 1235 if (!isSafePHIToSpeculate(PN, TD)) 1236 return false; 1237 1238 InstsToRewrite.insert(PN); 1239 continue; 1240 } 1241 1242 return false; 1243 } 1244 1245 // If there are no instructions to rewrite, then all uses are load/stores and 1246 // we're done! 1247 if (InstsToRewrite.empty()) 1248 return true; 1249 1250 // If we have instructions that need to be rewritten for this to be promotable 1251 // take care of it now. 1252 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1253 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1254 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1255 // loads with a new select. 1256 while (!SI->use_empty()) { 1257 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1258 1259 IRBuilder<> Builder(LI); 1260 LoadInst *TrueLoad = 1261 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1262 LoadInst *FalseLoad = 1263 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".t"); 1264 1265 // Transfer alignment and TBAA info if present. 1266 TrueLoad->setAlignment(LI->getAlignment()); 1267 FalseLoad->setAlignment(LI->getAlignment()); 1268 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1269 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1270 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1271 } 1272 1273 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1274 V->takeName(LI); 1275 LI->replaceAllUsesWith(V); 1276 LI->eraseFromParent(); 1277 } 1278 1279 // Now that all the loads are gone, the select is gone too. 1280 SI->eraseFromParent(); 1281 continue; 1282 } 1283 1284 // Otherwise, we have a PHI node which allows us to push the loads into the 1285 // predecessors. 1286 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1287 if (PN->use_empty()) { 1288 PN->eraseFromParent(); 1289 continue; 1290 } 1291 1292 const Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1293 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1294 PN->getName()+".ld", PN); 1295 1296 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1297 // matter which one we get and if any differ, it doesn't matter. 1298 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1299 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1300 unsigned Align = SomeLoad->getAlignment(); 1301 1302 // Rewrite all loads of the PN to use the new PHI. 1303 while (!PN->use_empty()) { 1304 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1305 LI->replaceAllUsesWith(NewPN); 1306 LI->eraseFromParent(); 1307 } 1308 1309 // Inject loads into all of the pred blocks. Keep track of which blocks we 1310 // insert them into in case we have multiple edges from the same block. 1311 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1312 1313 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1314 BasicBlock *Pred = PN->getIncomingBlock(i); 1315 LoadInst *&Load = InsertedLoads[Pred]; 1316 if (Load == 0) { 1317 Load = new LoadInst(PN->getIncomingValue(i), 1318 PN->getName() + "." + Pred->getName(), 1319 Pred->getTerminator()); 1320 Load->setAlignment(Align); 1321 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1322 } 1323 1324 NewPN->addIncoming(Load, Pred); 1325 } 1326 1327 PN->eraseFromParent(); 1328 } 1329 1330 ++NumAdjusted; 1331 return true; 1332} 1333 1334bool SROA::performPromotion(Function &F) { 1335 std::vector<AllocaInst*> Allocas; 1336 DominatorTree *DT = 0; 1337 if (HasDomTree) 1338 DT = &getAnalysis<DominatorTree>(); 1339 1340 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1341 1342 bool Changed = false; 1343 SmallVector<Instruction*, 64> Insts; 1344 DIBuilder *DIB = 0; 1345 while (1) { 1346 Allocas.clear(); 1347 1348 // Find allocas that are safe to promote, by looking at all instructions in 1349 // the entry node 1350 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1351 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1352 if (tryToMakeAllocaBePromotable(AI, TD)) 1353 Allocas.push_back(AI); 1354 1355 if (Allocas.empty()) break; 1356 1357 if (HasDomTree) 1358 PromoteMemToReg(Allocas, *DT); 1359 else { 1360 SSAUpdater SSA; 1361 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1362 AllocaInst *AI = Allocas[i]; 1363 1364 // Build list of instructions to promote. 1365 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1366 UI != E; ++UI) 1367 Insts.push_back(cast<Instruction>(*UI)); 1368 1369 DbgDeclareInst *DDI = FindAllocaDbgDeclare(AI); 1370 if (DDI && !DIB) 1371 DIB = new DIBuilder(*AI->getParent()->getParent()->getParent()); 1372 AllocaPromoter(Insts, SSA, DDI, DIB).run(AI, Insts); 1373 Insts.clear(); 1374 } 1375 } 1376 NumPromoted += Allocas.size(); 1377 Changed = true; 1378 } 1379 1380 // FIXME: Is there a better way to handle the lazy initialization of DIB 1381 // so that there doesn't need to be an explicit delete? 1382 delete DIB; 1383 1384 return Changed; 1385} 1386 1387 1388/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1389/// SROA. It must be a struct or array type with a small number of elements. 1390static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1391 const Type *T = AI->getAllocatedType(); 1392 // Do not promote any struct into more than 32 separate vars. 1393 if (const StructType *ST = dyn_cast<StructType>(T)) 1394 return ST->getNumElements() <= 32; 1395 // Arrays are much less likely to be safe for SROA; only consider 1396 // them if they are very small. 1397 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 1398 return AT->getNumElements() <= 8; 1399 return false; 1400} 1401 1402 1403// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1404// which runs on all of the malloc/alloca instructions in the function, removing 1405// them if they are only used by getelementptr instructions. 1406// 1407bool SROA::performScalarRepl(Function &F) { 1408 std::vector<AllocaInst*> WorkList; 1409 1410 // Scan the entry basic block, adding allocas to the worklist. 1411 BasicBlock &BB = F.getEntryBlock(); 1412 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1413 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1414 WorkList.push_back(A); 1415 1416 // Process the worklist 1417 bool Changed = false; 1418 while (!WorkList.empty()) { 1419 AllocaInst *AI = WorkList.back(); 1420 WorkList.pop_back(); 1421 1422 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1423 // with unused elements. 1424 if (AI->use_empty()) { 1425 AI->eraseFromParent(); 1426 Changed = true; 1427 continue; 1428 } 1429 1430 // If this alloca is impossible for us to promote, reject it early. 1431 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1432 continue; 1433 1434 // Check to see if this allocation is only modified by a memcpy/memmove from 1435 // a constant global. If this is the case, we can change all users to use 1436 // the constant global instead. This is commonly produced by the CFE by 1437 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1438 // is only subsequently read. 1439 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 1440 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1441 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 1442 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 1443 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1444 TheCopy->eraseFromParent(); // Don't mutate the global. 1445 AI->eraseFromParent(); 1446 ++NumGlobals; 1447 Changed = true; 1448 continue; 1449 } 1450 1451 // Check to see if we can perform the core SROA transformation. We cannot 1452 // transform the allocation instruction if it is an array allocation 1453 // (allocations OF arrays are ok though), and an allocation of a scalar 1454 // value cannot be decomposed at all. 1455 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1456 1457 // Do not promote [0 x %struct]. 1458 if (AllocaSize == 0) continue; 1459 1460 // Do not promote any struct whose size is too big. 1461 if (AllocaSize > SRThreshold) continue; 1462 1463 // If the alloca looks like a good candidate for scalar replacement, and if 1464 // all its users can be transformed, then split up the aggregate into its 1465 // separate elements. 1466 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1467 DoScalarReplacement(AI, WorkList); 1468 Changed = true; 1469 continue; 1470 } 1471 1472 // If we can turn this aggregate value (potentially with casts) into a 1473 // simple scalar value that can be mem2reg'd into a register value. 1474 // IsNotTrivial tracks whether this is something that mem2reg could have 1475 // promoted itself. If so, we don't want to transform it needlessly. Note 1476 // that we can't just check based on the type: the alloca may be of an i32 1477 // but that has pointer arithmetic to set byte 3 of it or something. 1478 if (AllocaInst *NewAI = 1479 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1480 NewAI->takeName(AI); 1481 AI->eraseFromParent(); 1482 ++NumConverted; 1483 Changed = true; 1484 continue; 1485 } 1486 1487 // Otherwise, couldn't process this alloca. 1488 } 1489 1490 return Changed; 1491} 1492 1493/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1494/// predicate, do SROA now. 1495void SROA::DoScalarReplacement(AllocaInst *AI, 1496 std::vector<AllocaInst*> &WorkList) { 1497 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1498 SmallVector<AllocaInst*, 32> ElementAllocas; 1499 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1500 ElementAllocas.reserve(ST->getNumContainedTypes()); 1501 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1502 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1503 AI->getAlignment(), 1504 AI->getName() + "." + Twine(i), AI); 1505 ElementAllocas.push_back(NA); 1506 WorkList.push_back(NA); // Add to worklist for recursive processing 1507 } 1508 } else { 1509 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1510 ElementAllocas.reserve(AT->getNumElements()); 1511 const Type *ElTy = AT->getElementType(); 1512 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1513 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1514 AI->getName() + "." + Twine(i), AI); 1515 ElementAllocas.push_back(NA); 1516 WorkList.push_back(NA); // Add to worklist for recursive processing 1517 } 1518 } 1519 1520 // Now that we have created the new alloca instructions, rewrite all the 1521 // uses of the old alloca. 1522 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1523 1524 // Now erase any instructions that were made dead while rewriting the alloca. 1525 DeleteDeadInstructions(); 1526 AI->eraseFromParent(); 1527 1528 ++NumReplaced; 1529} 1530 1531/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1532/// recursively including all their operands that become trivially dead. 1533void SROA::DeleteDeadInstructions() { 1534 while (!DeadInsts.empty()) { 1535 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1536 1537 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1538 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1539 // Zero out the operand and see if it becomes trivially dead. 1540 // (But, don't add allocas to the dead instruction list -- they are 1541 // already on the worklist and will be deleted separately.) 1542 *OI = 0; 1543 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1544 DeadInsts.push_back(U); 1545 } 1546 1547 I->eraseFromParent(); 1548 } 1549} 1550 1551/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1552/// performing scalar replacement of alloca AI. The results are flagged in 1553/// the Info parameter. Offset indicates the position within AI that is 1554/// referenced by this instruction. 1555void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1556 AllocaInfo &Info) { 1557 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1558 Instruction *User = cast<Instruction>(*UI); 1559 1560 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1561 isSafeForScalarRepl(BC, Offset, Info); 1562 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1563 uint64_t GEPOffset = Offset; 1564 isSafeGEP(GEPI, GEPOffset, Info); 1565 if (!Info.isUnsafe) 1566 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1567 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1568 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1569 if (Length == 0) 1570 return MarkUnsafe(Info, User); 1571 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1572 UI.getOperandNo() == 0, Info, MI, 1573 true /*AllowWholeAccess*/); 1574 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1575 if (LI->isVolatile()) 1576 return MarkUnsafe(Info, User); 1577 const Type *LIType = LI->getType(); 1578 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1579 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1580 Info.hasALoadOrStore = true; 1581 1582 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1583 // Store is ok if storing INTO the pointer, not storing the pointer 1584 if (SI->isVolatile() || SI->getOperand(0) == I) 1585 return MarkUnsafe(Info, User); 1586 1587 const Type *SIType = SI->getOperand(0)->getType(); 1588 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1589 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1590 Info.hasALoadOrStore = true; 1591 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1592 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1593 } else { 1594 return MarkUnsafe(Info, User); 1595 } 1596 if (Info.isUnsafe) return; 1597 } 1598} 1599 1600 1601/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1602/// derived from the alloca, we can often still split the alloca into elements. 1603/// This is useful if we have a large alloca where one element is phi'd 1604/// together somewhere: we can SRoA and promote all the other elements even if 1605/// we end up not being able to promote this one. 1606/// 1607/// All we require is that the uses of the PHI do not index into other parts of 1608/// the alloca. The most important use case for this is single load and stores 1609/// that are PHI'd together, which can happen due to code sinking. 1610void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1611 AllocaInfo &Info) { 1612 // If we've already checked this PHI, don't do it again. 1613 if (PHINode *PN = dyn_cast<PHINode>(I)) 1614 if (!Info.CheckedPHIs.insert(PN)) 1615 return; 1616 1617 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1618 Instruction *User = cast<Instruction>(*UI); 1619 1620 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1621 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1622 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1623 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1624 // but would have to prove that we're staying inside of an element being 1625 // promoted. 1626 if (!GEPI->hasAllZeroIndices()) 1627 return MarkUnsafe(Info, User); 1628 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1629 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1630 if (LI->isVolatile()) 1631 return MarkUnsafe(Info, User); 1632 const Type *LIType = LI->getType(); 1633 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1634 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1635 Info.hasALoadOrStore = true; 1636 1637 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1638 // Store is ok if storing INTO the pointer, not storing the pointer 1639 if (SI->isVolatile() || SI->getOperand(0) == I) 1640 return MarkUnsafe(Info, User); 1641 1642 const Type *SIType = SI->getOperand(0)->getType(); 1643 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1644 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1645 Info.hasALoadOrStore = true; 1646 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1647 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1648 } else { 1649 return MarkUnsafe(Info, User); 1650 } 1651 if (Info.isUnsafe) return; 1652 } 1653} 1654 1655/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1656/// replacement. It is safe when all the indices are constant, in-bounds 1657/// references, and when the resulting offset corresponds to an element within 1658/// the alloca type. The results are flagged in the Info parameter. Upon 1659/// return, Offset is adjusted as specified by the GEP indices. 1660void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1661 uint64_t &Offset, AllocaInfo &Info) { 1662 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1663 if (GEPIt == E) 1664 return; 1665 1666 // Walk through the GEP type indices, checking the types that this indexes 1667 // into. 1668 for (; GEPIt != E; ++GEPIt) { 1669 // Ignore struct elements, no extra checking needed for these. 1670 if ((*GEPIt)->isStructTy()) 1671 continue; 1672 1673 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1674 if (!IdxVal) 1675 return MarkUnsafe(Info, GEPI); 1676 } 1677 1678 // Compute the offset due to this GEP and check if the alloca has a 1679 // component element at that offset. 1680 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1681 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1682 &Indices[0], Indices.size()); 1683 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1684 MarkUnsafe(Info, GEPI); 1685} 1686 1687/// isHomogeneousAggregate - Check if type T is a struct or array containing 1688/// elements of the same type (which is always true for arrays). If so, 1689/// return true with NumElts and EltTy set to the number of elements and the 1690/// element type, respectively. 1691static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1692 const Type *&EltTy) { 1693 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1694 NumElts = AT->getNumElements(); 1695 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1696 return true; 1697 } 1698 if (const StructType *ST = dyn_cast<StructType>(T)) { 1699 NumElts = ST->getNumContainedTypes(); 1700 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1701 for (unsigned n = 1; n < NumElts; ++n) { 1702 if (ST->getContainedType(n) != EltTy) 1703 return false; 1704 } 1705 return true; 1706 } 1707 return false; 1708} 1709 1710/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1711/// "homogeneous" aggregates with the same element type and number of elements. 1712static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1713 if (T1 == T2) 1714 return true; 1715 1716 unsigned NumElts1, NumElts2; 1717 const Type *EltTy1, *EltTy2; 1718 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1719 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1720 NumElts1 == NumElts2 && 1721 EltTy1 == EltTy2) 1722 return true; 1723 1724 return false; 1725} 1726 1727/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1728/// alloca or has an offset and size that corresponds to a component element 1729/// within it. The offset checked here may have been formed from a GEP with a 1730/// pointer bitcasted to a different type. 1731/// 1732/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1733/// unit. If false, it only allows accesses known to be in a single element. 1734void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1735 const Type *MemOpType, bool isStore, 1736 AllocaInfo &Info, Instruction *TheAccess, 1737 bool AllowWholeAccess) { 1738 // Check if this is a load/store of the entire alloca. 1739 if (Offset == 0 && AllowWholeAccess && 1740 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1741 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1742 // loads/stores (which are essentially the same as the MemIntrinsics with 1743 // regard to copying padding between elements). But, if an alloca is 1744 // flagged as both a source and destination of such operations, we'll need 1745 // to check later for padding between elements. 1746 if (!MemOpType || MemOpType->isIntegerTy()) { 1747 if (isStore) 1748 Info.isMemCpyDst = true; 1749 else 1750 Info.isMemCpySrc = true; 1751 return; 1752 } 1753 // This is also safe for references using a type that is compatible with 1754 // the type of the alloca, so that loads/stores can be rewritten using 1755 // insertvalue/extractvalue. 1756 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1757 Info.hasSubelementAccess = true; 1758 return; 1759 } 1760 } 1761 // Check if the offset/size correspond to a component within the alloca type. 1762 const Type *T = Info.AI->getAllocatedType(); 1763 if (TypeHasComponent(T, Offset, MemSize)) { 1764 Info.hasSubelementAccess = true; 1765 return; 1766 } 1767 1768 return MarkUnsafe(Info, TheAccess); 1769} 1770 1771/// TypeHasComponent - Return true if T has a component type with the 1772/// specified offset and size. If Size is zero, do not check the size. 1773bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1774 const Type *EltTy; 1775 uint64_t EltSize; 1776 if (const StructType *ST = dyn_cast<StructType>(T)) { 1777 const StructLayout *Layout = TD->getStructLayout(ST); 1778 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1779 EltTy = ST->getContainedType(EltIdx); 1780 EltSize = TD->getTypeAllocSize(EltTy); 1781 Offset -= Layout->getElementOffset(EltIdx); 1782 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1783 EltTy = AT->getElementType(); 1784 EltSize = TD->getTypeAllocSize(EltTy); 1785 if (Offset >= AT->getNumElements() * EltSize) 1786 return false; 1787 Offset %= EltSize; 1788 } else { 1789 return false; 1790 } 1791 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1792 return true; 1793 // Check if the component spans multiple elements. 1794 if (Offset + Size > EltSize) 1795 return false; 1796 return TypeHasComponent(EltTy, Offset, Size); 1797} 1798 1799/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1800/// the instruction I, which references it, to use the separate elements. 1801/// Offset indicates the position within AI that is referenced by this 1802/// instruction. 1803void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1804 SmallVector<AllocaInst*, 32> &NewElts) { 1805 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1806 Use &TheUse = UI.getUse(); 1807 Instruction *User = cast<Instruction>(*UI++); 1808 1809 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1810 RewriteBitCast(BC, AI, Offset, NewElts); 1811 continue; 1812 } 1813 1814 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1815 RewriteGEP(GEPI, AI, Offset, NewElts); 1816 continue; 1817 } 1818 1819 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1820 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1821 uint64_t MemSize = Length->getZExtValue(); 1822 if (Offset == 0 && 1823 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1824 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1825 // Otherwise the intrinsic can only touch a single element and the 1826 // address operand will be updated, so nothing else needs to be done. 1827 continue; 1828 } 1829 1830 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1831 const Type *LIType = LI->getType(); 1832 1833 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1834 // Replace: 1835 // %res = load { i32, i32 }* %alloc 1836 // with: 1837 // %load.0 = load i32* %alloc.0 1838 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1839 // %load.1 = load i32* %alloc.1 1840 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1841 // (Also works for arrays instead of structs) 1842 Value *Insert = UndefValue::get(LIType); 1843 IRBuilder<> Builder(LI); 1844 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1845 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1846 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1847 } 1848 LI->replaceAllUsesWith(Insert); 1849 DeadInsts.push_back(LI); 1850 } else if (LIType->isIntegerTy() && 1851 TD->getTypeAllocSize(LIType) == 1852 TD->getTypeAllocSize(AI->getAllocatedType())) { 1853 // If this is a load of the entire alloca to an integer, rewrite it. 1854 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1855 } 1856 continue; 1857 } 1858 1859 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1860 Value *Val = SI->getOperand(0); 1861 const Type *SIType = Val->getType(); 1862 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1863 // Replace: 1864 // store { i32, i32 } %val, { i32, i32 }* %alloc 1865 // with: 1866 // %val.0 = extractvalue { i32, i32 } %val, 0 1867 // store i32 %val.0, i32* %alloc.0 1868 // %val.1 = extractvalue { i32, i32 } %val, 1 1869 // store i32 %val.1, i32* %alloc.1 1870 // (Also works for arrays instead of structs) 1871 IRBuilder<> Builder(SI); 1872 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1873 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1874 Builder.CreateStore(Extract, NewElts[i]); 1875 } 1876 DeadInsts.push_back(SI); 1877 } else if (SIType->isIntegerTy() && 1878 TD->getTypeAllocSize(SIType) == 1879 TD->getTypeAllocSize(AI->getAllocatedType())) { 1880 // If this is a store of the entire alloca from an integer, rewrite it. 1881 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1882 } 1883 continue; 1884 } 1885 1886 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1887 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1888 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1889 // the new pointer. 1890 if (!isa<AllocaInst>(I)) continue; 1891 1892 assert(Offset == 0 && NewElts[0] && 1893 "Direct alloca use should have a zero offset"); 1894 1895 // If we have a use of the alloca, we know the derived uses will be 1896 // utilizing just the first element of the scalarized result. Insert a 1897 // bitcast of the first alloca before the user as required. 1898 AllocaInst *NewAI = NewElts[0]; 1899 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1900 NewAI->moveBefore(BCI); 1901 TheUse = BCI; 1902 continue; 1903 } 1904 } 1905} 1906 1907/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1908/// and recursively continue updating all of its uses. 1909void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1910 SmallVector<AllocaInst*, 32> &NewElts) { 1911 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1912 if (BC->getOperand(0) != AI) 1913 return; 1914 1915 // The bitcast references the original alloca. Replace its uses with 1916 // references to the first new element alloca. 1917 Instruction *Val = NewElts[0]; 1918 if (Val->getType() != BC->getDestTy()) { 1919 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1920 Val->takeName(BC); 1921 } 1922 BC->replaceAllUsesWith(Val); 1923 DeadInsts.push_back(BC); 1924} 1925 1926/// FindElementAndOffset - Return the index of the element containing Offset 1927/// within the specified type, which must be either a struct or an array. 1928/// Sets T to the type of the element and Offset to the offset within that 1929/// element. IdxTy is set to the type of the index result to be used in a 1930/// GEP instruction. 1931uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1932 const Type *&IdxTy) { 1933 uint64_t Idx = 0; 1934 if (const StructType *ST = dyn_cast<StructType>(T)) { 1935 const StructLayout *Layout = TD->getStructLayout(ST); 1936 Idx = Layout->getElementContainingOffset(Offset); 1937 T = ST->getContainedType(Idx); 1938 Offset -= Layout->getElementOffset(Idx); 1939 IdxTy = Type::getInt32Ty(T->getContext()); 1940 return Idx; 1941 } 1942 const ArrayType *AT = cast<ArrayType>(T); 1943 T = AT->getElementType(); 1944 uint64_t EltSize = TD->getTypeAllocSize(T); 1945 Idx = Offset / EltSize; 1946 Offset -= Idx * EltSize; 1947 IdxTy = Type::getInt64Ty(T->getContext()); 1948 return Idx; 1949} 1950 1951/// RewriteGEP - Check if this GEP instruction moves the pointer across 1952/// elements of the alloca that are being split apart, and if so, rewrite 1953/// the GEP to be relative to the new element. 1954void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1955 SmallVector<AllocaInst*, 32> &NewElts) { 1956 uint64_t OldOffset = Offset; 1957 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1958 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1959 &Indices[0], Indices.size()); 1960 1961 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1962 1963 const Type *T = AI->getAllocatedType(); 1964 const Type *IdxTy; 1965 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1966 if (GEPI->getOperand(0) == AI) 1967 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1968 1969 T = AI->getAllocatedType(); 1970 uint64_t EltOffset = Offset; 1971 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1972 1973 // If this GEP does not move the pointer across elements of the alloca 1974 // being split, then it does not needs to be rewritten. 1975 if (Idx == OldIdx) 1976 return; 1977 1978 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1979 SmallVector<Value*, 8> NewArgs; 1980 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1981 while (EltOffset != 0) { 1982 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1983 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1984 } 1985 Instruction *Val = NewElts[Idx]; 1986 if (NewArgs.size() > 1) { 1987 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1988 NewArgs.end(), "", GEPI); 1989 Val->takeName(GEPI); 1990 } 1991 if (Val->getType() != GEPI->getType()) 1992 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1993 GEPI->replaceAllUsesWith(Val); 1994 DeadInsts.push_back(GEPI); 1995} 1996 1997/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1998/// Rewrite it to copy or set the elements of the scalarized memory. 1999void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2000 AllocaInst *AI, 2001 SmallVector<AllocaInst*, 32> &NewElts) { 2002 // If this is a memcpy/memmove, construct the other pointer as the 2003 // appropriate type. The "Other" pointer is the pointer that goes to memory 2004 // that doesn't have anything to do with the alloca that we are promoting. For 2005 // memset, this Value* stays null. 2006 Value *OtherPtr = 0; 2007 unsigned MemAlignment = MI->getAlignment(); 2008 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2009 if (Inst == MTI->getRawDest()) 2010 OtherPtr = MTI->getRawSource(); 2011 else { 2012 assert(Inst == MTI->getRawSource()); 2013 OtherPtr = MTI->getRawDest(); 2014 } 2015 } 2016 2017 // If there is an other pointer, we want to convert it to the same pointer 2018 // type as AI has, so we can GEP through it safely. 2019 if (OtherPtr) { 2020 unsigned AddrSpace = 2021 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2022 2023 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2024 // optimization, but it's also required to detect the corner case where 2025 // both pointer operands are referencing the same memory, and where 2026 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2027 // function is only called for mem intrinsics that access the whole 2028 // aggregate, so non-zero GEPs are not an issue here.) 2029 OtherPtr = OtherPtr->stripPointerCasts(); 2030 2031 // Copying the alloca to itself is a no-op: just delete it. 2032 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2033 // This code will run twice for a no-op memcpy -- once for each operand. 2034 // Put only one reference to MI on the DeadInsts list. 2035 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2036 E = DeadInsts.end(); I != E; ++I) 2037 if (*I == MI) return; 2038 DeadInsts.push_back(MI); 2039 return; 2040 } 2041 2042 // If the pointer is not the right type, insert a bitcast to the right 2043 // type. 2044 const Type *NewTy = 2045 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2046 2047 if (OtherPtr->getType() != NewTy) 2048 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2049 } 2050 2051 // Process each element of the aggregate. 2052 bool SROADest = MI->getRawDest() == Inst; 2053 2054 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2055 2056 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2057 // If this is a memcpy/memmove, emit a GEP of the other element address. 2058 Value *OtherElt = 0; 2059 unsigned OtherEltAlign = MemAlignment; 2060 2061 if (OtherPtr) { 2062 Value *Idx[2] = { Zero, 2063 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2064 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 2065 OtherPtr->getName()+"."+Twine(i), 2066 MI); 2067 uint64_t EltOffset; 2068 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2069 const Type *OtherTy = OtherPtrTy->getElementType(); 2070 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 2071 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2072 } else { 2073 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2074 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2075 } 2076 2077 // The alignment of the other pointer is the guaranteed alignment of the 2078 // element, which is affected by both the known alignment of the whole 2079 // mem intrinsic and the alignment of the element. If the alignment of 2080 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2081 // known alignment is just 4 bytes. 2082 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2083 } 2084 2085 Value *EltPtr = NewElts[i]; 2086 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2087 2088 // If we got down to a scalar, insert a load or store as appropriate. 2089 if (EltTy->isSingleValueType()) { 2090 if (isa<MemTransferInst>(MI)) { 2091 if (SROADest) { 2092 // From Other to Alloca. 2093 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2094 new StoreInst(Elt, EltPtr, MI); 2095 } else { 2096 // From Alloca to Other. 2097 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2098 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2099 } 2100 continue; 2101 } 2102 assert(isa<MemSetInst>(MI)); 2103 2104 // If the stored element is zero (common case), just store a null 2105 // constant. 2106 Constant *StoreVal; 2107 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2108 if (CI->isZero()) { 2109 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2110 } else { 2111 // If EltTy is a vector type, get the element type. 2112 const Type *ValTy = EltTy->getScalarType(); 2113 2114 // Construct an integer with the right value. 2115 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2116 APInt OneVal(EltSize, CI->getZExtValue()); 2117 APInt TotalVal(OneVal); 2118 // Set each byte. 2119 for (unsigned i = 0; 8*i < EltSize; ++i) { 2120 TotalVal = TotalVal.shl(8); 2121 TotalVal |= OneVal; 2122 } 2123 2124 // Convert the integer value to the appropriate type. 2125 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2126 if (ValTy->isPointerTy()) 2127 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2128 else if (ValTy->isFloatingPointTy()) 2129 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2130 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2131 2132 // If the requested value was a vector constant, create it. 2133 if (EltTy != ValTy) { 2134 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 2135 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2136 StoreVal = ConstantVector::get(Elts); 2137 } 2138 } 2139 new StoreInst(StoreVal, EltPtr, MI); 2140 continue; 2141 } 2142 // Otherwise, if we're storing a byte variable, use a memset call for 2143 // this element. 2144 } 2145 2146 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2147 2148 IRBuilder<> Builder(MI); 2149 2150 // Finally, insert the meminst for this element. 2151 if (isa<MemSetInst>(MI)) { 2152 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2153 MI->isVolatile()); 2154 } else { 2155 assert(isa<MemTransferInst>(MI)); 2156 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2157 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2158 2159 if (isa<MemCpyInst>(MI)) 2160 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2161 else 2162 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2163 } 2164 } 2165 DeadInsts.push_back(MI); 2166} 2167 2168/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2169/// overwrites the entire allocation. Extract out the pieces of the stored 2170/// integer and store them individually. 2171void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2172 SmallVector<AllocaInst*, 32> &NewElts){ 2173 // Extract each element out of the integer according to its structure offset 2174 // and store the element value to the individual alloca. 2175 Value *SrcVal = SI->getOperand(0); 2176 const Type *AllocaEltTy = AI->getAllocatedType(); 2177 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2178 2179 IRBuilder<> Builder(SI); 2180 2181 // Handle tail padding by extending the operand 2182 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2183 SrcVal = Builder.CreateZExt(SrcVal, 2184 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2185 2186 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2187 << '\n'); 2188 2189 // There are two forms here: AI could be an array or struct. Both cases 2190 // have different ways to compute the element offset. 2191 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2192 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2193 2194 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2195 // Get the number of bits to shift SrcVal to get the value. 2196 const Type *FieldTy = EltSTy->getElementType(i); 2197 uint64_t Shift = Layout->getElementOffsetInBits(i); 2198 2199 if (TD->isBigEndian()) 2200 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2201 2202 Value *EltVal = SrcVal; 2203 if (Shift) { 2204 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2205 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2206 } 2207 2208 // Truncate down to an integer of the right size. 2209 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2210 2211 // Ignore zero sized fields like {}, they obviously contain no data. 2212 if (FieldSizeBits == 0) continue; 2213 2214 if (FieldSizeBits != AllocaSizeBits) 2215 EltVal = Builder.CreateTrunc(EltVal, 2216 IntegerType::get(SI->getContext(), FieldSizeBits)); 2217 Value *DestField = NewElts[i]; 2218 if (EltVal->getType() == FieldTy) { 2219 // Storing to an integer field of this size, just do it. 2220 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2221 // Bitcast to the right element type (for fp/vector values). 2222 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2223 } else { 2224 // Otherwise, bitcast the dest pointer (for aggregates). 2225 DestField = Builder.CreateBitCast(DestField, 2226 PointerType::getUnqual(EltVal->getType())); 2227 } 2228 new StoreInst(EltVal, DestField, SI); 2229 } 2230 2231 } else { 2232 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2233 const Type *ArrayEltTy = ATy->getElementType(); 2234 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2235 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2236 2237 uint64_t Shift; 2238 2239 if (TD->isBigEndian()) 2240 Shift = AllocaSizeBits-ElementOffset; 2241 else 2242 Shift = 0; 2243 2244 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2245 // Ignore zero sized fields like {}, they obviously contain no data. 2246 if (ElementSizeBits == 0) continue; 2247 2248 Value *EltVal = SrcVal; 2249 if (Shift) { 2250 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2251 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2252 } 2253 2254 // Truncate down to an integer of the right size. 2255 if (ElementSizeBits != AllocaSizeBits) 2256 EltVal = Builder.CreateTrunc(EltVal, 2257 IntegerType::get(SI->getContext(), 2258 ElementSizeBits)); 2259 Value *DestField = NewElts[i]; 2260 if (EltVal->getType() == ArrayEltTy) { 2261 // Storing to an integer field of this size, just do it. 2262 } else if (ArrayEltTy->isFloatingPointTy() || 2263 ArrayEltTy->isVectorTy()) { 2264 // Bitcast to the right element type (for fp/vector values). 2265 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2266 } else { 2267 // Otherwise, bitcast the dest pointer (for aggregates). 2268 DestField = Builder.CreateBitCast(DestField, 2269 PointerType::getUnqual(EltVal->getType())); 2270 } 2271 new StoreInst(EltVal, DestField, SI); 2272 2273 if (TD->isBigEndian()) 2274 Shift -= ElementOffset; 2275 else 2276 Shift += ElementOffset; 2277 } 2278 } 2279 2280 DeadInsts.push_back(SI); 2281} 2282 2283/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2284/// an integer. Load the individual pieces to form the aggregate value. 2285void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2286 SmallVector<AllocaInst*, 32> &NewElts) { 2287 // Extract each element out of the NewElts according to its structure offset 2288 // and form the result value. 2289 const Type *AllocaEltTy = AI->getAllocatedType(); 2290 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2291 2292 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2293 << '\n'); 2294 2295 // There are two forms here: AI could be an array or struct. Both cases 2296 // have different ways to compute the element offset. 2297 const StructLayout *Layout = 0; 2298 uint64_t ArrayEltBitOffset = 0; 2299 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2300 Layout = TD->getStructLayout(EltSTy); 2301 } else { 2302 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2303 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2304 } 2305 2306 Value *ResultVal = 2307 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2308 2309 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2310 // Load the value from the alloca. If the NewElt is an aggregate, cast 2311 // the pointer to an integer of the same size before doing the load. 2312 Value *SrcField = NewElts[i]; 2313 const Type *FieldTy = 2314 cast<PointerType>(SrcField->getType())->getElementType(); 2315 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2316 2317 // Ignore zero sized fields like {}, they obviously contain no data. 2318 if (FieldSizeBits == 0) continue; 2319 2320 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2321 FieldSizeBits); 2322 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2323 !FieldTy->isVectorTy()) 2324 SrcField = new BitCastInst(SrcField, 2325 PointerType::getUnqual(FieldIntTy), 2326 "", LI); 2327 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2328 2329 // If SrcField is a fp or vector of the right size but that isn't an 2330 // integer type, bitcast to an integer so we can shift it. 2331 if (SrcField->getType() != FieldIntTy) 2332 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2333 2334 // Zero extend the field to be the same size as the final alloca so that 2335 // we can shift and insert it. 2336 if (SrcField->getType() != ResultVal->getType()) 2337 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2338 2339 // Determine the number of bits to shift SrcField. 2340 uint64_t Shift; 2341 if (Layout) // Struct case. 2342 Shift = Layout->getElementOffsetInBits(i); 2343 else // Array case. 2344 Shift = i*ArrayEltBitOffset; 2345 2346 if (TD->isBigEndian()) 2347 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2348 2349 if (Shift) { 2350 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2351 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2352 } 2353 2354 // Don't create an 'or x, 0' on the first iteration. 2355 if (!isa<Constant>(ResultVal) || 2356 !cast<Constant>(ResultVal)->isNullValue()) 2357 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2358 else 2359 ResultVal = SrcField; 2360 } 2361 2362 // Handle tail padding by truncating the result 2363 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2364 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2365 2366 LI->replaceAllUsesWith(ResultVal); 2367 DeadInsts.push_back(LI); 2368} 2369 2370/// HasPadding - Return true if the specified type has any structure or 2371/// alignment padding in between the elements that would be split apart 2372/// by SROA; return false otherwise. 2373static bool HasPadding(const Type *Ty, const TargetData &TD) { 2374 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2375 Ty = ATy->getElementType(); 2376 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2377 } 2378 2379 // SROA currently handles only Arrays and Structs. 2380 const StructType *STy = cast<StructType>(Ty); 2381 const StructLayout *SL = TD.getStructLayout(STy); 2382 unsigned PrevFieldBitOffset = 0; 2383 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2384 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2385 2386 // Check to see if there is any padding between this element and the 2387 // previous one. 2388 if (i) { 2389 unsigned PrevFieldEnd = 2390 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2391 if (PrevFieldEnd < FieldBitOffset) 2392 return true; 2393 } 2394 PrevFieldBitOffset = FieldBitOffset; 2395 } 2396 // Check for tail padding. 2397 if (unsigned EltCount = STy->getNumElements()) { 2398 unsigned PrevFieldEnd = PrevFieldBitOffset + 2399 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2400 if (PrevFieldEnd < SL->getSizeInBits()) 2401 return true; 2402 } 2403 return false; 2404} 2405 2406/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2407/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2408/// or 1 if safe after canonicalization has been performed. 2409bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2410 // Loop over the use list of the alloca. We can only transform it if all of 2411 // the users are safe to transform. 2412 AllocaInfo Info(AI); 2413 2414 isSafeForScalarRepl(AI, 0, Info); 2415 if (Info.isUnsafe) { 2416 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2417 return false; 2418 } 2419 2420 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2421 // source and destination, we have to be careful. In particular, the memcpy 2422 // could be moving around elements that live in structure padding of the LLVM 2423 // types, but may actually be used. In these cases, we refuse to promote the 2424 // struct. 2425 if (Info.isMemCpySrc && Info.isMemCpyDst && 2426 HasPadding(AI->getAllocatedType(), *TD)) 2427 return false; 2428 2429 // If the alloca never has an access to just *part* of it, but is accessed 2430 // via loads and stores, then we should use ConvertToScalarInfo to promote 2431 // the alloca instead of promoting each piece at a time and inserting fission 2432 // and fusion code. 2433 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2434 // If the struct/array just has one element, use basic SRoA. 2435 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2436 if (ST->getNumElements() > 1) return false; 2437 } else { 2438 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2439 return false; 2440 } 2441 } 2442 2443 return true; 2444} 2445 2446 2447 2448/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2449/// some part of a constant global variable. This intentionally only accepts 2450/// constant expressions because we don't can't rewrite arbitrary instructions. 2451static bool PointsToConstantGlobal(Value *V) { 2452 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2453 return GV->isConstant(); 2454 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2455 if (CE->getOpcode() == Instruction::BitCast || 2456 CE->getOpcode() == Instruction::GetElementPtr) 2457 return PointsToConstantGlobal(CE->getOperand(0)); 2458 return false; 2459} 2460 2461/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2462/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2463/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2464/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2465/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2466/// the alloca, and if the source pointer is a pointer to a constant global, we 2467/// can optimize this. 2468static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2469 bool isOffset) { 2470 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2471 User *U = cast<Instruction>(*UI); 2472 2473 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2474 // Ignore non-volatile loads, they are always ok. 2475 if (LI->isVolatile()) return false; 2476 continue; 2477 } 2478 2479 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2480 // If uses of the bitcast are ok, we are ok. 2481 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 2482 return false; 2483 continue; 2484 } 2485 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2486 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2487 // doesn't, it does. 2488 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2489 isOffset || !GEP->hasAllZeroIndices())) 2490 return false; 2491 continue; 2492 } 2493 2494 if (CallSite CS = U) { 2495 // If this is the function being called then we treat it like a load and 2496 // ignore it. 2497 if (CS.isCallee(UI)) 2498 continue; 2499 2500 // If this is a readonly/readnone call site, then we know it is just a 2501 // load (but one that potentially returns the value itself), so we can 2502 // ignore it if we know that the value isn't captured. 2503 unsigned ArgNo = CS.getArgumentNo(UI); 2504 if (CS.onlyReadsMemory() && 2505 (CS.getInstruction()->use_empty() || 2506 CS.paramHasAttr(ArgNo+1, Attribute::NoCapture))) 2507 continue; 2508 2509 // If this is being passed as a byval argument, the caller is making a 2510 // copy, so it is only a read of the alloca. 2511 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2512 continue; 2513 } 2514 2515 // If this is isn't our memcpy/memmove, reject it as something we can't 2516 // handle. 2517 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2518 if (MI == 0) 2519 return false; 2520 2521 // If the transfer is using the alloca as a source of the transfer, then 2522 // ignore it since it is a load (unless the transfer is volatile). 2523 if (UI.getOperandNo() == 1) { 2524 if (MI->isVolatile()) return false; 2525 continue; 2526 } 2527 2528 // If we already have seen a copy, reject the second one. 2529 if (TheCopy) return false; 2530 2531 // If the pointer has been offset from the start of the alloca, we can't 2532 // safely handle this. 2533 if (isOffset) return false; 2534 2535 // If the memintrinsic isn't using the alloca as the dest, reject it. 2536 if (UI.getOperandNo() != 0) return false; 2537 2538 // If the source of the memcpy/move is not a constant global, reject it. 2539 if (!PointsToConstantGlobal(MI->getSource())) 2540 return false; 2541 2542 // Otherwise, the transform is safe. Remember the copy instruction. 2543 TheCopy = MI; 2544 } 2545 return true; 2546} 2547 2548/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2549/// modified by a copy from a constant global. If we can prove this, we can 2550/// replace any uses of the alloca with uses of the global directly. 2551MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 2552 MemTransferInst *TheCopy = 0; 2553 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 2554 return TheCopy; 2555 return 0; 2556} 2557