ScalarReplAggregates.cpp revision dd68912801861273dc3dca33cfc18357213049a4
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); 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 // If we already decided to turn this into a blob of integer memory, there is 321 // nothing to be done. 322 if (VectorTy && VectorTy->isVoidTy()) 323 return; 324 325 // If this could be contributing to a vector, analyze it. 326 327 // If the In type is a vector that is the same size as the alloca, see if it 328 // matches the existing VecTy. 329 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 330 if (MergeInVectorType(VInTy, Offset)) 331 return; 332 } else if (In->isFloatTy() || In->isDoubleTy() || 333 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 334 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 335 // Full width accesses can be ignored, because they can always be turned 336 // into bitcasts. 337 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 338 if (EltSize == AllocaSize) 339 return; 340 341 // If we're accessing something that could be an element of a vector, see 342 // if the implied vector agrees with what we already have and if Offset is 343 // compatible with it. 344 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 345 (!VectorTy || Offset * 8 < VectorTy->getPrimitiveSizeInBits())) { 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); 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); 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 FromTypeSize = TD.getTypeAllocSize(FromType); 746 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 747 if (FromTypeSize == ToTypeSize) { 748 // If the two types have the same primitive size, use a bit cast. 749 // Otherwise, it is two vectors with the same element type that has 750 // the same allocation size but different number of elements so use 751 // a shuffle vector. 752 if (FromType->getPrimitiveSizeInBits() == 753 ToType->getPrimitiveSizeInBits()) 754 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 755 else 756 return CreateShuffleVectorCast(FromVal, ToType, Builder); 757 } 758 759 if (isPowerOf2_64(FromTypeSize / ToTypeSize)) { 760 assert(!(ToType->isVectorTy() && Offset != 0) && "Can't extract a value " 761 "of a smaller vector type at a nonzero offset."); 762 763 const Type *CastElementTy = getScaledElementType(FromType, ToType, 764 ToTypeSize * 8); 765 unsigned NumCastVectorElements = FromTypeSize / ToTypeSize; 766 767 LLVMContext &Context = FromVal->getContext(); 768 const Type *CastTy = VectorType::get(CastElementTy, 769 NumCastVectorElements); 770 Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp"); 771 772 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); 773 unsigned Elt = Offset/EltSize; 774 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 775 Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get( 776 Type::getInt32Ty(Context), Elt), "tmp"); 777 return Builder.CreateBitCast(Extract, ToType, "tmp"); 778 } 779 780 // Otherwise it must be an element access. 781 unsigned Elt = 0; 782 if (Offset) { 783 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 784 Elt = Offset/EltSize; 785 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 786 } 787 // Return the element extracted out of it. 788 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 789 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 790 if (V->getType() != ToType) 791 V = Builder.CreateBitCast(V, ToType, "tmp"); 792 return V; 793 } 794 795 // If ToType is a first class aggregate, extract out each of the pieces and 796 // use insertvalue's to form the FCA. 797 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 798 const StructLayout &Layout = *TD.getStructLayout(ST); 799 Value *Res = UndefValue::get(ST); 800 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 801 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 802 Offset+Layout.getElementOffsetInBits(i), 803 Builder); 804 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 805 } 806 return Res; 807 } 808 809 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 810 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 811 Value *Res = UndefValue::get(AT); 812 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 813 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 814 Offset+i*EltSize, Builder); 815 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 816 } 817 return Res; 818 } 819 820 // Otherwise, this must be a union that was converted to an integer value. 821 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 822 823 // If this is a big-endian system and the load is narrower than the 824 // full alloca type, we need to do a shift to get the right bits. 825 int ShAmt = 0; 826 if (TD.isBigEndian()) { 827 // On big-endian machines, the lowest bit is stored at the bit offset 828 // from the pointer given by getTypeStoreSizeInBits. This matters for 829 // integers with a bitwidth that is not a multiple of 8. 830 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 831 TD.getTypeStoreSizeInBits(ToType) - Offset; 832 } else { 833 ShAmt = Offset; 834 } 835 836 // Note: we support negative bitwidths (with shl) which are not defined. 837 // We do this to support (f.e.) loads off the end of a structure where 838 // only some bits are used. 839 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 840 FromVal = Builder.CreateLShr(FromVal, 841 ConstantInt::get(FromVal->getType(), 842 ShAmt), "tmp"); 843 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 844 FromVal = Builder.CreateShl(FromVal, 845 ConstantInt::get(FromVal->getType(), 846 -ShAmt), "tmp"); 847 848 // Finally, unconditionally truncate the integer to the right width. 849 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 850 if (LIBitWidth < NTy->getBitWidth()) 851 FromVal = 852 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 853 LIBitWidth), "tmp"); 854 else if (LIBitWidth > NTy->getBitWidth()) 855 FromVal = 856 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 857 LIBitWidth), "tmp"); 858 859 // If the result is an integer, this is a trunc or bitcast. 860 if (ToType->isIntegerTy()) { 861 // Should be done. 862 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 863 // Just do a bitcast, we know the sizes match up. 864 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 865 } else { 866 // Otherwise must be a pointer. 867 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 868 } 869 assert(FromVal->getType() == ToType && "Didn't convert right?"); 870 return FromVal; 871} 872 873/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 874/// or vector value "Old" at the offset specified by Offset. 875/// 876/// This happens when we are converting an "integer union" to a 877/// single integer scalar, or when we are converting a "vector union" to a 878/// vector with insert/extractelement instructions. 879/// 880/// Offset is an offset from the original alloca, in bits that need to be 881/// shifted to the right. 882Value *ConvertToScalarInfo:: 883ConvertScalar_InsertValue(Value *SV, Value *Old, 884 uint64_t Offset, IRBuilder<> &Builder) { 885 // Convert the stored type to the actual type, shift it left to insert 886 // then 'or' into place. 887 const Type *AllocaType = Old->getType(); 888 LLVMContext &Context = Old->getContext(); 889 890 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 891 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 892 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 893 894 // Changing the whole vector with memset or with an access of a different 895 // vector type? 896 if (ValSize == VecSize) { 897 // If the two types have the same primitive size, use a bit cast. 898 // Otherwise, it is two vectors with the same element type that has 899 // the same allocation size but different number of elements so use 900 // a shuffle vector. 901 if (VTy->getPrimitiveSizeInBits() == 902 SV->getType()->getPrimitiveSizeInBits()) 903 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 904 else 905 return CreateShuffleVectorCast(SV, VTy, Builder); 906 } 907 908 if (isPowerOf2_64(VecSize / ValSize)) { 909 assert(!(SV->getType()->isVectorTy() && Offset != 0) && "Can't insert a " 910 "value of a smaller vector type at a nonzero offset."); 911 912 const Type *CastElementTy = getScaledElementType(VTy, SV->getType(), 913 ValSize); 914 unsigned NumCastVectorElements = VecSize / ValSize; 915 916 LLVMContext &Context = SV->getContext(); 917 const Type *OldCastTy = VectorType::get(CastElementTy, 918 NumCastVectorElements); 919 Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp"); 920 921 Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp"); 922 923 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); 924 unsigned Elt = Offset/EltSize; 925 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 926 Value *Insert = 927 Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get( 928 Type::getInt32Ty(Context), Elt), "tmp"); 929 return Builder.CreateBitCast(Insert, AllocaType, "tmp"); 930 } 931 932 // Must be an element insertion. 933 assert(SV->getType() == VTy->getElementType()); 934 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 935 unsigned Elt = Offset/EltSize; 936 return Builder.CreateInsertElement(Old, SV, 937 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 938 "tmp"); 939 } 940 941 // If SV is a first-class aggregate value, insert each value recursively. 942 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 943 const StructLayout &Layout = *TD.getStructLayout(ST); 944 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 945 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 946 Old = ConvertScalar_InsertValue(Elt, Old, 947 Offset+Layout.getElementOffsetInBits(i), 948 Builder); 949 } 950 return Old; 951 } 952 953 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 954 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 955 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 956 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 957 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 958 } 959 return Old; 960 } 961 962 // If SV is a float, convert it to the appropriate integer type. 963 // If it is a pointer, do the same. 964 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 965 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 966 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 967 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 968 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 969 SV = Builder.CreateBitCast(SV, 970 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 971 else if (SV->getType()->isPointerTy()) 972 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 973 974 // Zero extend or truncate the value if needed. 975 if (SV->getType() != AllocaType) { 976 if (SV->getType()->getPrimitiveSizeInBits() < 977 AllocaType->getPrimitiveSizeInBits()) 978 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 979 else { 980 // Truncation may be needed if storing more than the alloca can hold 981 // (undefined behavior). 982 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 983 SrcWidth = DestWidth; 984 SrcStoreWidth = DestStoreWidth; 985 } 986 } 987 988 // If this is a big-endian system and the store is narrower than the 989 // full alloca type, we need to do a shift to get the right bits. 990 int ShAmt = 0; 991 if (TD.isBigEndian()) { 992 // On big-endian machines, the lowest bit is stored at the bit offset 993 // from the pointer given by getTypeStoreSizeInBits. This matters for 994 // integers with a bitwidth that is not a multiple of 8. 995 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 996 } else { 997 ShAmt = Offset; 998 } 999 1000 // Note: we support negative bitwidths (with shr) which are not defined. 1001 // We do this to support (f.e.) stores off the end of a structure where 1002 // only some bits in the structure are set. 1003 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1004 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1005 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1006 ShAmt), "tmp"); 1007 Mask <<= ShAmt; 1008 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1009 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1010 -ShAmt), "tmp"); 1011 Mask = Mask.lshr(-ShAmt); 1012 } 1013 1014 // Mask out the bits we are about to insert from the old value, and or 1015 // in the new bits. 1016 if (SrcWidth != DestWidth) { 1017 assert(DestWidth > SrcWidth); 1018 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1019 SV = Builder.CreateOr(Old, SV, "ins"); 1020 } 1021 return SV; 1022} 1023 1024 1025//===----------------------------------------------------------------------===// 1026// SRoA Driver 1027//===----------------------------------------------------------------------===// 1028 1029 1030bool SROA::runOnFunction(Function &F) { 1031 TD = getAnalysisIfAvailable<TargetData>(); 1032 1033 bool Changed = performPromotion(F); 1034 1035 // FIXME: ScalarRepl currently depends on TargetData more than it 1036 // theoretically needs to. It should be refactored in order to support 1037 // target-independent IR. Until this is done, just skip the actual 1038 // scalar-replacement portion of this pass. 1039 if (!TD) return Changed; 1040 1041 while (1) { 1042 bool LocalChange = performScalarRepl(F); 1043 if (!LocalChange) break; // No need to repromote if no scalarrepl 1044 Changed = true; 1045 LocalChange = performPromotion(F); 1046 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 1047 } 1048 1049 return Changed; 1050} 1051 1052namespace { 1053class AllocaPromoter : public LoadAndStorePromoter { 1054 AllocaInst *AI; 1055public: 1056 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 1057 DbgDeclareInst *DD, DIBuilder *&DB) 1058 : LoadAndStorePromoter(Insts, S, DD, DB), AI(0) {} 1059 1060 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 1061 // Remember which alloca we're promoting (for isInstInList). 1062 this->AI = AI; 1063 LoadAndStorePromoter::run(Insts); 1064 AI->eraseFromParent(); 1065 } 1066 1067 virtual bool isInstInList(Instruction *I, 1068 const SmallVectorImpl<Instruction*> &Insts) const { 1069 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1070 return LI->getOperand(0) == AI; 1071 return cast<StoreInst>(I)->getPointerOperand() == AI; 1072 } 1073}; 1074} // end anon namespace 1075 1076/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1077/// subsequently loaded can be rewritten to load both input pointers and then 1078/// select between the result, allowing the load of the alloca to be promoted. 1079/// From this: 1080/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1081/// %V = load i32* %P2 1082/// to: 1083/// %V1 = load i32* %Alloca -> will be mem2reg'd 1084/// %V2 = load i32* %Other 1085/// %V = select i1 %cond, i32 %V1, i32 %V2 1086/// 1087/// We can do this to a select if its only uses are loads and if the operand to 1088/// the select can be loaded unconditionally. 1089static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1090 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1091 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1092 1093 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1094 UI != UE; ++UI) { 1095 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1096 if (LI == 0 || LI->isVolatile()) return false; 1097 1098 // Both operands to the select need to be dereferencable, either absolutely 1099 // (e.g. allocas) or at this point because we can see other accesses to it. 1100 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1101 LI->getAlignment(), TD)) 1102 return false; 1103 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1104 LI->getAlignment(), TD)) 1105 return false; 1106 } 1107 1108 return true; 1109} 1110 1111/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1112/// subsequently loaded can be rewritten to load both input pointers in the pred 1113/// blocks and then PHI the results, allowing the load of the alloca to be 1114/// promoted. 1115/// From this: 1116/// %P2 = phi [i32* %Alloca, i32* %Other] 1117/// %V = load i32* %P2 1118/// to: 1119/// %V1 = load i32* %Alloca -> will be mem2reg'd 1120/// ... 1121/// %V2 = load i32* %Other 1122/// ... 1123/// %V = phi [i32 %V1, i32 %V2] 1124/// 1125/// We can do this to a select if its only uses are loads and if the operand to 1126/// the select can be loaded unconditionally. 1127static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1128 // For now, we can only do this promotion if the load is in the same block as 1129 // the PHI, and if there are no stores between the phi and load. 1130 // TODO: Allow recursive phi users. 1131 // TODO: Allow stores. 1132 BasicBlock *BB = PN->getParent(); 1133 unsigned MaxAlign = 0; 1134 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1135 UI != UE; ++UI) { 1136 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1137 if (LI == 0 || LI->isVolatile()) return false; 1138 1139 // For now we only allow loads in the same block as the PHI. This is a 1140 // common case that happens when instcombine merges two loads through a PHI. 1141 if (LI->getParent() != BB) return false; 1142 1143 // Ensure that there are no instructions between the PHI and the load that 1144 // could store. 1145 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1146 if (BBI->mayWriteToMemory()) 1147 return false; 1148 1149 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1150 } 1151 1152 // Okay, we know that we have one or more loads in the same block as the PHI. 1153 // We can transform this if it is safe to push the loads into the predecessor 1154 // blocks. The only thing to watch out for is that we can't put a possibly 1155 // trapping load in the predecessor if it is a critical edge. 1156 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1157 BasicBlock *Pred = PN->getIncomingBlock(i); 1158 1159 // If the predecessor has a single successor, then the edge isn't critical. 1160 if (Pred->getTerminator()->getNumSuccessors() == 1) 1161 continue; 1162 1163 Value *InVal = PN->getIncomingValue(i); 1164 1165 // If the InVal is an invoke in the pred, we can't put a load on the edge. 1166 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 1167 if (II->getParent() == Pred) 1168 return false; 1169 1170 // If this pointer is always safe to load, or if we can prove that there is 1171 // already a load in the block, then we can move the load to the pred block. 1172 if (InVal->isDereferenceablePointer() || 1173 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1174 continue; 1175 1176 return false; 1177 } 1178 1179 return true; 1180} 1181 1182 1183/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1184/// direct (non-volatile) loads and stores to it. If the alloca is close but 1185/// not quite there, this will transform the code to allow promotion. As such, 1186/// it is a non-pure predicate. 1187static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1188 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1189 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1190 1191 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1192 UI != UE; ++UI) { 1193 User *U = *UI; 1194 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1195 if (LI->isVolatile()) 1196 return false; 1197 continue; 1198 } 1199 1200 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1201 if (SI->getOperand(0) == AI || SI->isVolatile()) 1202 return false; // Don't allow a store OF the AI, only INTO the AI. 1203 continue; 1204 } 1205 1206 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1207 // If the condition being selected on is a constant, fold the select, yes 1208 // this does (rarely) happen early on. 1209 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1210 Value *Result = SI->getOperand(1+CI->isZero()); 1211 SI->replaceAllUsesWith(Result); 1212 SI->eraseFromParent(); 1213 1214 // This is very rare and we just scrambled the use list of AI, start 1215 // over completely. 1216 return tryToMakeAllocaBePromotable(AI, TD); 1217 } 1218 1219 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1220 // loads, then we can transform this by rewriting the select. 1221 if (!isSafeSelectToSpeculate(SI, TD)) 1222 return false; 1223 1224 InstsToRewrite.insert(SI); 1225 continue; 1226 } 1227 1228 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1229 if (PN->use_empty()) { // Dead PHIs can be stripped. 1230 InstsToRewrite.insert(PN); 1231 continue; 1232 } 1233 1234 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1235 // in the pred blocks, then we can transform this by rewriting the PHI. 1236 if (!isSafePHIToSpeculate(PN, TD)) 1237 return false; 1238 1239 InstsToRewrite.insert(PN); 1240 continue; 1241 } 1242 1243 return false; 1244 } 1245 1246 // If there are no instructions to rewrite, then all uses are load/stores and 1247 // we're done! 1248 if (InstsToRewrite.empty()) 1249 return true; 1250 1251 // If we have instructions that need to be rewritten for this to be promotable 1252 // take care of it now. 1253 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1254 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1255 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1256 // loads with a new select. 1257 while (!SI->use_empty()) { 1258 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1259 1260 IRBuilder<> Builder(LI); 1261 LoadInst *TrueLoad = 1262 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1263 LoadInst *FalseLoad = 1264 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".t"); 1265 1266 // Transfer alignment and TBAA info if present. 1267 TrueLoad->setAlignment(LI->getAlignment()); 1268 FalseLoad->setAlignment(LI->getAlignment()); 1269 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1270 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1271 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1272 } 1273 1274 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1275 V->takeName(LI); 1276 LI->replaceAllUsesWith(V); 1277 LI->eraseFromParent(); 1278 } 1279 1280 // Now that all the loads are gone, the select is gone too. 1281 SI->eraseFromParent(); 1282 continue; 1283 } 1284 1285 // Otherwise, we have a PHI node which allows us to push the loads into the 1286 // predecessors. 1287 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1288 if (PN->use_empty()) { 1289 PN->eraseFromParent(); 1290 continue; 1291 } 1292 1293 const Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1294 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1295 PN->getName()+".ld", PN); 1296 1297 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1298 // matter which one we get and if any differ, it doesn't matter. 1299 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1300 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1301 unsigned Align = SomeLoad->getAlignment(); 1302 1303 // Rewrite all loads of the PN to use the new PHI. 1304 while (!PN->use_empty()) { 1305 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1306 LI->replaceAllUsesWith(NewPN); 1307 LI->eraseFromParent(); 1308 } 1309 1310 // Inject loads into all of the pred blocks. Keep track of which blocks we 1311 // insert them into in case we have multiple edges from the same block. 1312 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1313 1314 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1315 BasicBlock *Pred = PN->getIncomingBlock(i); 1316 LoadInst *&Load = InsertedLoads[Pred]; 1317 if (Load == 0) { 1318 Load = new LoadInst(PN->getIncomingValue(i), 1319 PN->getName() + "." + Pred->getName(), 1320 Pred->getTerminator()); 1321 Load->setAlignment(Align); 1322 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1323 } 1324 1325 NewPN->addIncoming(Load, Pred); 1326 } 1327 1328 PN->eraseFromParent(); 1329 } 1330 1331 ++NumAdjusted; 1332 return true; 1333} 1334 1335bool SROA::performPromotion(Function &F) { 1336 std::vector<AllocaInst*> Allocas; 1337 DominatorTree *DT = 0; 1338 if (HasDomTree) 1339 DT = &getAnalysis<DominatorTree>(); 1340 1341 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1342 1343 bool Changed = false; 1344 SmallVector<Instruction*, 64> Insts; 1345 DIBuilder *DIB = 0; 1346 while (1) { 1347 Allocas.clear(); 1348 1349 // Find allocas that are safe to promote, by looking at all instructions in 1350 // the entry node 1351 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1352 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1353 if (tryToMakeAllocaBePromotable(AI, TD)) 1354 Allocas.push_back(AI); 1355 1356 if (Allocas.empty()) break; 1357 1358 if (HasDomTree) 1359 PromoteMemToReg(Allocas, *DT); 1360 else { 1361 SSAUpdater SSA; 1362 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1363 AllocaInst *AI = Allocas[i]; 1364 1365 // Build list of instructions to promote. 1366 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1367 UI != E; ++UI) 1368 Insts.push_back(cast<Instruction>(*UI)); 1369 1370 DbgDeclareInst *DDI = FindAllocaDbgDeclare(AI); 1371 if (DDI && !DIB) 1372 DIB = new DIBuilder(*AI->getParent()->getParent()->getParent()); 1373 AllocaPromoter(Insts, SSA, DDI, DIB).run(AI, Insts); 1374 Insts.clear(); 1375 } 1376 } 1377 NumPromoted += Allocas.size(); 1378 Changed = true; 1379 } 1380 1381 // FIXME: Is there a better way to handle the lazy initialization of DIB 1382 // so that there doesn't need to be an explicit delete? 1383 delete DIB; 1384 1385 return Changed; 1386} 1387 1388 1389/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1390/// SROA. It must be a struct or array type with a small number of elements. 1391static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1392 const Type *T = AI->getAllocatedType(); 1393 // Do not promote any struct into more than 32 separate vars. 1394 if (const StructType *ST = dyn_cast<StructType>(T)) 1395 return ST->getNumElements() <= 32; 1396 // Arrays are much less likely to be safe for SROA; only consider 1397 // them if they are very small. 1398 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 1399 return AT->getNumElements() <= 8; 1400 return false; 1401} 1402 1403 1404// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1405// which runs on all of the malloc/alloca instructions in the function, removing 1406// them if they are only used by getelementptr instructions. 1407// 1408bool SROA::performScalarRepl(Function &F) { 1409 std::vector<AllocaInst*> WorkList; 1410 1411 // Scan the entry basic block, adding allocas to the worklist. 1412 BasicBlock &BB = F.getEntryBlock(); 1413 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1414 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1415 WorkList.push_back(A); 1416 1417 // Process the worklist 1418 bool Changed = false; 1419 while (!WorkList.empty()) { 1420 AllocaInst *AI = WorkList.back(); 1421 WorkList.pop_back(); 1422 1423 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1424 // with unused elements. 1425 if (AI->use_empty()) { 1426 AI->eraseFromParent(); 1427 Changed = true; 1428 continue; 1429 } 1430 1431 // If this alloca is impossible for us to promote, reject it early. 1432 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1433 continue; 1434 1435 // Check to see if this allocation is only modified by a memcpy/memmove from 1436 // a constant global. If this is the case, we can change all users to use 1437 // the constant global instead. This is commonly produced by the CFE by 1438 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1439 // is only subsequently read. 1440 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 1441 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1442 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 1443 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 1444 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1445 TheCopy->eraseFromParent(); // Don't mutate the global. 1446 AI->eraseFromParent(); 1447 ++NumGlobals; 1448 Changed = true; 1449 continue; 1450 } 1451 1452 // Check to see if we can perform the core SROA transformation. We cannot 1453 // transform the allocation instruction if it is an array allocation 1454 // (allocations OF arrays are ok though), and an allocation of a scalar 1455 // value cannot be decomposed at all. 1456 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1457 1458 // Do not promote [0 x %struct]. 1459 if (AllocaSize == 0) continue; 1460 1461 // Do not promote any struct whose size is too big. 1462 if (AllocaSize > SRThreshold) continue; 1463 1464 // If the alloca looks like a good candidate for scalar replacement, and if 1465 // all its users can be transformed, then split up the aggregate into its 1466 // separate elements. 1467 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1468 DoScalarReplacement(AI, WorkList); 1469 Changed = true; 1470 continue; 1471 } 1472 1473 // If we can turn this aggregate value (potentially with casts) into a 1474 // simple scalar value that can be mem2reg'd into a register value. 1475 // IsNotTrivial tracks whether this is something that mem2reg could have 1476 // promoted itself. If so, we don't want to transform it needlessly. Note 1477 // that we can't just check based on the type: the alloca may be of an i32 1478 // but that has pointer arithmetic to set byte 3 of it or something. 1479 if (AllocaInst *NewAI = 1480 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1481 NewAI->takeName(AI); 1482 AI->eraseFromParent(); 1483 ++NumConverted; 1484 Changed = true; 1485 continue; 1486 } 1487 1488 // Otherwise, couldn't process this alloca. 1489 } 1490 1491 return Changed; 1492} 1493 1494/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1495/// predicate, do SROA now. 1496void SROA::DoScalarReplacement(AllocaInst *AI, 1497 std::vector<AllocaInst*> &WorkList) { 1498 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1499 SmallVector<AllocaInst*, 32> ElementAllocas; 1500 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1501 ElementAllocas.reserve(ST->getNumContainedTypes()); 1502 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1503 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1504 AI->getAlignment(), 1505 AI->getName() + "." + Twine(i), AI); 1506 ElementAllocas.push_back(NA); 1507 WorkList.push_back(NA); // Add to worklist for recursive processing 1508 } 1509 } else { 1510 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1511 ElementAllocas.reserve(AT->getNumElements()); 1512 const Type *ElTy = AT->getElementType(); 1513 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1514 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1515 AI->getName() + "." + Twine(i), AI); 1516 ElementAllocas.push_back(NA); 1517 WorkList.push_back(NA); // Add to worklist for recursive processing 1518 } 1519 } 1520 1521 // Now that we have created the new alloca instructions, rewrite all the 1522 // uses of the old alloca. 1523 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1524 1525 // Now erase any instructions that were made dead while rewriting the alloca. 1526 DeleteDeadInstructions(); 1527 AI->eraseFromParent(); 1528 1529 ++NumReplaced; 1530} 1531 1532/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1533/// recursively including all their operands that become trivially dead. 1534void SROA::DeleteDeadInstructions() { 1535 while (!DeadInsts.empty()) { 1536 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1537 1538 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1539 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1540 // Zero out the operand and see if it becomes trivially dead. 1541 // (But, don't add allocas to the dead instruction list -- they are 1542 // already on the worklist and will be deleted separately.) 1543 *OI = 0; 1544 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1545 DeadInsts.push_back(U); 1546 } 1547 1548 I->eraseFromParent(); 1549 } 1550} 1551 1552/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1553/// performing scalar replacement of alloca AI. The results are flagged in 1554/// the Info parameter. Offset indicates the position within AI that is 1555/// referenced by this instruction. 1556void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1557 AllocaInfo &Info) { 1558 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1559 Instruction *User = cast<Instruction>(*UI); 1560 1561 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1562 isSafeForScalarRepl(BC, Offset, Info); 1563 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1564 uint64_t GEPOffset = Offset; 1565 isSafeGEP(GEPI, GEPOffset, Info); 1566 if (!Info.isUnsafe) 1567 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1568 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1569 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1570 if (Length == 0) 1571 return MarkUnsafe(Info, User); 1572 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1573 UI.getOperandNo() == 0, Info, MI, 1574 true /*AllowWholeAccess*/); 1575 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1576 if (LI->isVolatile()) 1577 return MarkUnsafe(Info, User); 1578 const Type *LIType = LI->getType(); 1579 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1580 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1581 Info.hasALoadOrStore = true; 1582 1583 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1584 // Store is ok if storing INTO the pointer, not storing the pointer 1585 if (SI->isVolatile() || SI->getOperand(0) == I) 1586 return MarkUnsafe(Info, User); 1587 1588 const Type *SIType = SI->getOperand(0)->getType(); 1589 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1590 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1591 Info.hasALoadOrStore = true; 1592 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1593 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1594 } else { 1595 return MarkUnsafe(Info, User); 1596 } 1597 if (Info.isUnsafe) return; 1598 } 1599} 1600 1601 1602/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1603/// derived from the alloca, we can often still split the alloca into elements. 1604/// This is useful if we have a large alloca where one element is phi'd 1605/// together somewhere: we can SRoA and promote all the other elements even if 1606/// we end up not being able to promote this one. 1607/// 1608/// All we require is that the uses of the PHI do not index into other parts of 1609/// the alloca. The most important use case for this is single load and stores 1610/// that are PHI'd together, which can happen due to code sinking. 1611void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1612 AllocaInfo &Info) { 1613 // If we've already checked this PHI, don't do it again. 1614 if (PHINode *PN = dyn_cast<PHINode>(I)) 1615 if (!Info.CheckedPHIs.insert(PN)) 1616 return; 1617 1618 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1619 Instruction *User = cast<Instruction>(*UI); 1620 1621 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1622 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1623 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1624 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1625 // but would have to prove that we're staying inside of an element being 1626 // promoted. 1627 if (!GEPI->hasAllZeroIndices()) 1628 return MarkUnsafe(Info, User); 1629 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1630 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1631 if (LI->isVolatile()) 1632 return MarkUnsafe(Info, User); 1633 const Type *LIType = LI->getType(); 1634 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1635 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1636 Info.hasALoadOrStore = true; 1637 1638 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1639 // Store is ok if storing INTO the pointer, not storing the pointer 1640 if (SI->isVolatile() || SI->getOperand(0) == I) 1641 return MarkUnsafe(Info, User); 1642 1643 const Type *SIType = SI->getOperand(0)->getType(); 1644 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1645 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1646 Info.hasALoadOrStore = true; 1647 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1648 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1649 } else { 1650 return MarkUnsafe(Info, User); 1651 } 1652 if (Info.isUnsafe) return; 1653 } 1654} 1655 1656/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1657/// replacement. It is safe when all the indices are constant, in-bounds 1658/// references, and when the resulting offset corresponds to an element within 1659/// the alloca type. The results are flagged in the Info parameter. Upon 1660/// return, Offset is adjusted as specified by the GEP indices. 1661void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1662 uint64_t &Offset, AllocaInfo &Info) { 1663 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1664 if (GEPIt == E) 1665 return; 1666 1667 // Walk through the GEP type indices, checking the types that this indexes 1668 // into. 1669 for (; GEPIt != E; ++GEPIt) { 1670 // Ignore struct elements, no extra checking needed for these. 1671 if ((*GEPIt)->isStructTy()) 1672 continue; 1673 1674 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1675 if (!IdxVal) 1676 return MarkUnsafe(Info, GEPI); 1677 } 1678 1679 // Compute the offset due to this GEP and check if the alloca has a 1680 // component element at that offset. 1681 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1682 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1683 &Indices[0], Indices.size()); 1684 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1685 MarkUnsafe(Info, GEPI); 1686} 1687 1688/// isHomogeneousAggregate - Check if type T is a struct or array containing 1689/// elements of the same type (which is always true for arrays). If so, 1690/// return true with NumElts and EltTy set to the number of elements and the 1691/// element type, respectively. 1692static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1693 const Type *&EltTy) { 1694 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1695 NumElts = AT->getNumElements(); 1696 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1697 return true; 1698 } 1699 if (const StructType *ST = dyn_cast<StructType>(T)) { 1700 NumElts = ST->getNumContainedTypes(); 1701 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1702 for (unsigned n = 1; n < NumElts; ++n) { 1703 if (ST->getContainedType(n) != EltTy) 1704 return false; 1705 } 1706 return true; 1707 } 1708 return false; 1709} 1710 1711/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1712/// "homogeneous" aggregates with the same element type and number of elements. 1713static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1714 if (T1 == T2) 1715 return true; 1716 1717 unsigned NumElts1, NumElts2; 1718 const Type *EltTy1, *EltTy2; 1719 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1720 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1721 NumElts1 == NumElts2 && 1722 EltTy1 == EltTy2) 1723 return true; 1724 1725 return false; 1726} 1727 1728/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1729/// alloca or has an offset and size that corresponds to a component element 1730/// within it. The offset checked here may have been formed from a GEP with a 1731/// pointer bitcasted to a different type. 1732/// 1733/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1734/// unit. If false, it only allows accesses known to be in a single element. 1735void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1736 const Type *MemOpType, bool isStore, 1737 AllocaInfo &Info, Instruction *TheAccess, 1738 bool AllowWholeAccess) { 1739 // Check if this is a load/store of the entire alloca. 1740 if (Offset == 0 && AllowWholeAccess && 1741 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1742 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1743 // loads/stores (which are essentially the same as the MemIntrinsics with 1744 // regard to copying padding between elements). But, if an alloca is 1745 // flagged as both a source and destination of such operations, we'll need 1746 // to check later for padding between elements. 1747 if (!MemOpType || MemOpType->isIntegerTy()) { 1748 if (isStore) 1749 Info.isMemCpyDst = true; 1750 else 1751 Info.isMemCpySrc = true; 1752 return; 1753 } 1754 // This is also safe for references using a type that is compatible with 1755 // the type of the alloca, so that loads/stores can be rewritten using 1756 // insertvalue/extractvalue. 1757 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1758 Info.hasSubelementAccess = true; 1759 return; 1760 } 1761 } 1762 // Check if the offset/size correspond to a component within the alloca type. 1763 const Type *T = Info.AI->getAllocatedType(); 1764 if (TypeHasComponent(T, Offset, MemSize)) { 1765 Info.hasSubelementAccess = true; 1766 return; 1767 } 1768 1769 return MarkUnsafe(Info, TheAccess); 1770} 1771 1772/// TypeHasComponent - Return true if T has a component type with the 1773/// specified offset and size. If Size is zero, do not check the size. 1774bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1775 const Type *EltTy; 1776 uint64_t EltSize; 1777 if (const StructType *ST = dyn_cast<StructType>(T)) { 1778 const StructLayout *Layout = TD->getStructLayout(ST); 1779 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1780 EltTy = ST->getContainedType(EltIdx); 1781 EltSize = TD->getTypeAllocSize(EltTy); 1782 Offset -= Layout->getElementOffset(EltIdx); 1783 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1784 EltTy = AT->getElementType(); 1785 EltSize = TD->getTypeAllocSize(EltTy); 1786 if (Offset >= AT->getNumElements() * EltSize) 1787 return false; 1788 Offset %= EltSize; 1789 } else { 1790 return false; 1791 } 1792 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1793 return true; 1794 // Check if the component spans multiple elements. 1795 if (Offset + Size > EltSize) 1796 return false; 1797 return TypeHasComponent(EltTy, Offset, Size); 1798} 1799 1800/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1801/// the instruction I, which references it, to use the separate elements. 1802/// Offset indicates the position within AI that is referenced by this 1803/// instruction. 1804void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1805 SmallVector<AllocaInst*, 32> &NewElts) { 1806 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1807 Use &TheUse = UI.getUse(); 1808 Instruction *User = cast<Instruction>(*UI++); 1809 1810 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1811 RewriteBitCast(BC, AI, Offset, NewElts); 1812 continue; 1813 } 1814 1815 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1816 RewriteGEP(GEPI, AI, Offset, NewElts); 1817 continue; 1818 } 1819 1820 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1821 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1822 uint64_t MemSize = Length->getZExtValue(); 1823 if (Offset == 0 && 1824 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1825 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1826 // Otherwise the intrinsic can only touch a single element and the 1827 // address operand will be updated, so nothing else needs to be done. 1828 continue; 1829 } 1830 1831 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1832 const Type *LIType = LI->getType(); 1833 1834 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1835 // Replace: 1836 // %res = load { i32, i32 }* %alloc 1837 // with: 1838 // %load.0 = load i32* %alloc.0 1839 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1840 // %load.1 = load i32* %alloc.1 1841 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1842 // (Also works for arrays instead of structs) 1843 Value *Insert = UndefValue::get(LIType); 1844 IRBuilder<> Builder(LI); 1845 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1846 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1847 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1848 } 1849 LI->replaceAllUsesWith(Insert); 1850 DeadInsts.push_back(LI); 1851 } else if (LIType->isIntegerTy() && 1852 TD->getTypeAllocSize(LIType) == 1853 TD->getTypeAllocSize(AI->getAllocatedType())) { 1854 // If this is a load of the entire alloca to an integer, rewrite it. 1855 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1856 } 1857 continue; 1858 } 1859 1860 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1861 Value *Val = SI->getOperand(0); 1862 const Type *SIType = Val->getType(); 1863 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1864 // Replace: 1865 // store { i32, i32 } %val, { i32, i32 }* %alloc 1866 // with: 1867 // %val.0 = extractvalue { i32, i32 } %val, 0 1868 // store i32 %val.0, i32* %alloc.0 1869 // %val.1 = extractvalue { i32, i32 } %val, 1 1870 // store i32 %val.1, i32* %alloc.1 1871 // (Also works for arrays instead of structs) 1872 IRBuilder<> Builder(SI); 1873 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1874 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1875 Builder.CreateStore(Extract, NewElts[i]); 1876 } 1877 DeadInsts.push_back(SI); 1878 } else if (SIType->isIntegerTy() && 1879 TD->getTypeAllocSize(SIType) == 1880 TD->getTypeAllocSize(AI->getAllocatedType())) { 1881 // If this is a store of the entire alloca from an integer, rewrite it. 1882 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1883 } 1884 continue; 1885 } 1886 1887 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1888 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1889 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1890 // the new pointer. 1891 if (!isa<AllocaInst>(I)) continue; 1892 1893 assert(Offset == 0 && NewElts[0] && 1894 "Direct alloca use should have a zero offset"); 1895 1896 // If we have a use of the alloca, we know the derived uses will be 1897 // utilizing just the first element of the scalarized result. Insert a 1898 // bitcast of the first alloca before the user as required. 1899 AllocaInst *NewAI = NewElts[0]; 1900 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1901 NewAI->moveBefore(BCI); 1902 TheUse = BCI; 1903 continue; 1904 } 1905 } 1906} 1907 1908/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1909/// and recursively continue updating all of its uses. 1910void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1911 SmallVector<AllocaInst*, 32> &NewElts) { 1912 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1913 if (BC->getOperand(0) != AI) 1914 return; 1915 1916 // The bitcast references the original alloca. Replace its uses with 1917 // references to the first new element alloca. 1918 Instruction *Val = NewElts[0]; 1919 if (Val->getType() != BC->getDestTy()) { 1920 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1921 Val->takeName(BC); 1922 } 1923 BC->replaceAllUsesWith(Val); 1924 DeadInsts.push_back(BC); 1925} 1926 1927/// FindElementAndOffset - Return the index of the element containing Offset 1928/// within the specified type, which must be either a struct or an array. 1929/// Sets T to the type of the element and Offset to the offset within that 1930/// element. IdxTy is set to the type of the index result to be used in a 1931/// GEP instruction. 1932uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1933 const Type *&IdxTy) { 1934 uint64_t Idx = 0; 1935 if (const StructType *ST = dyn_cast<StructType>(T)) { 1936 const StructLayout *Layout = TD->getStructLayout(ST); 1937 Idx = Layout->getElementContainingOffset(Offset); 1938 T = ST->getContainedType(Idx); 1939 Offset -= Layout->getElementOffset(Idx); 1940 IdxTy = Type::getInt32Ty(T->getContext()); 1941 return Idx; 1942 } 1943 const ArrayType *AT = cast<ArrayType>(T); 1944 T = AT->getElementType(); 1945 uint64_t EltSize = TD->getTypeAllocSize(T); 1946 Idx = Offset / EltSize; 1947 Offset -= Idx * EltSize; 1948 IdxTy = Type::getInt64Ty(T->getContext()); 1949 return Idx; 1950} 1951 1952/// RewriteGEP - Check if this GEP instruction moves the pointer across 1953/// elements of the alloca that are being split apart, and if so, rewrite 1954/// the GEP to be relative to the new element. 1955void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1956 SmallVector<AllocaInst*, 32> &NewElts) { 1957 uint64_t OldOffset = Offset; 1958 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1959 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1960 &Indices[0], Indices.size()); 1961 1962 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1963 1964 const Type *T = AI->getAllocatedType(); 1965 const Type *IdxTy; 1966 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1967 if (GEPI->getOperand(0) == AI) 1968 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1969 1970 T = AI->getAllocatedType(); 1971 uint64_t EltOffset = Offset; 1972 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1973 1974 // If this GEP does not move the pointer across elements of the alloca 1975 // being split, then it does not needs to be rewritten. 1976 if (Idx == OldIdx) 1977 return; 1978 1979 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1980 SmallVector<Value*, 8> NewArgs; 1981 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1982 while (EltOffset != 0) { 1983 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1984 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1985 } 1986 Instruction *Val = NewElts[Idx]; 1987 if (NewArgs.size() > 1) { 1988 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1989 NewArgs.end(), "", GEPI); 1990 Val->takeName(GEPI); 1991 } 1992 if (Val->getType() != GEPI->getType()) 1993 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1994 GEPI->replaceAllUsesWith(Val); 1995 DeadInsts.push_back(GEPI); 1996} 1997 1998/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1999/// Rewrite it to copy or set the elements of the scalarized memory. 2000void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2001 AllocaInst *AI, 2002 SmallVector<AllocaInst*, 32> &NewElts) { 2003 // If this is a memcpy/memmove, construct the other pointer as the 2004 // appropriate type. The "Other" pointer is the pointer that goes to memory 2005 // that doesn't have anything to do with the alloca that we are promoting. For 2006 // memset, this Value* stays null. 2007 Value *OtherPtr = 0; 2008 unsigned MemAlignment = MI->getAlignment(); 2009 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2010 if (Inst == MTI->getRawDest()) 2011 OtherPtr = MTI->getRawSource(); 2012 else { 2013 assert(Inst == MTI->getRawSource()); 2014 OtherPtr = MTI->getRawDest(); 2015 } 2016 } 2017 2018 // If there is an other pointer, we want to convert it to the same pointer 2019 // type as AI has, so we can GEP through it safely. 2020 if (OtherPtr) { 2021 unsigned AddrSpace = 2022 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2023 2024 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2025 // optimization, but it's also required to detect the corner case where 2026 // both pointer operands are referencing the same memory, and where 2027 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2028 // function is only called for mem intrinsics that access the whole 2029 // aggregate, so non-zero GEPs are not an issue here.) 2030 OtherPtr = OtherPtr->stripPointerCasts(); 2031 2032 // Copying the alloca to itself is a no-op: just delete it. 2033 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2034 // This code will run twice for a no-op memcpy -- once for each operand. 2035 // Put only one reference to MI on the DeadInsts list. 2036 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2037 E = DeadInsts.end(); I != E; ++I) 2038 if (*I == MI) return; 2039 DeadInsts.push_back(MI); 2040 return; 2041 } 2042 2043 // If the pointer is not the right type, insert a bitcast to the right 2044 // type. 2045 const Type *NewTy = 2046 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2047 2048 if (OtherPtr->getType() != NewTy) 2049 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2050 } 2051 2052 // Process each element of the aggregate. 2053 bool SROADest = MI->getRawDest() == Inst; 2054 2055 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2056 2057 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2058 // If this is a memcpy/memmove, emit a GEP of the other element address. 2059 Value *OtherElt = 0; 2060 unsigned OtherEltAlign = MemAlignment; 2061 2062 if (OtherPtr) { 2063 Value *Idx[2] = { Zero, 2064 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2065 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 2066 OtherPtr->getName()+"."+Twine(i), 2067 MI); 2068 uint64_t EltOffset; 2069 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2070 const Type *OtherTy = OtherPtrTy->getElementType(); 2071 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 2072 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2073 } else { 2074 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2075 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2076 } 2077 2078 // The alignment of the other pointer is the guaranteed alignment of the 2079 // element, which is affected by both the known alignment of the whole 2080 // mem intrinsic and the alignment of the element. If the alignment of 2081 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2082 // known alignment is just 4 bytes. 2083 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2084 } 2085 2086 Value *EltPtr = NewElts[i]; 2087 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2088 2089 // If we got down to a scalar, insert a load or store as appropriate. 2090 if (EltTy->isSingleValueType()) { 2091 if (isa<MemTransferInst>(MI)) { 2092 if (SROADest) { 2093 // From Other to Alloca. 2094 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2095 new StoreInst(Elt, EltPtr, MI); 2096 } else { 2097 // From Alloca to Other. 2098 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2099 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2100 } 2101 continue; 2102 } 2103 assert(isa<MemSetInst>(MI)); 2104 2105 // If the stored element is zero (common case), just store a null 2106 // constant. 2107 Constant *StoreVal; 2108 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2109 if (CI->isZero()) { 2110 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2111 } else { 2112 // If EltTy is a vector type, get the element type. 2113 const Type *ValTy = EltTy->getScalarType(); 2114 2115 // Construct an integer with the right value. 2116 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2117 APInt OneVal(EltSize, CI->getZExtValue()); 2118 APInt TotalVal(OneVal); 2119 // Set each byte. 2120 for (unsigned i = 0; 8*i < EltSize; ++i) { 2121 TotalVal = TotalVal.shl(8); 2122 TotalVal |= OneVal; 2123 } 2124 2125 // Convert the integer value to the appropriate type. 2126 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2127 if (ValTy->isPointerTy()) 2128 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2129 else if (ValTy->isFloatingPointTy()) 2130 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2131 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2132 2133 // If the requested value was a vector constant, create it. 2134 if (EltTy != ValTy) { 2135 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 2136 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2137 StoreVal = ConstantVector::get(Elts); 2138 } 2139 } 2140 new StoreInst(StoreVal, EltPtr, MI); 2141 continue; 2142 } 2143 // Otherwise, if we're storing a byte variable, use a memset call for 2144 // this element. 2145 } 2146 2147 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2148 2149 IRBuilder<> Builder(MI); 2150 2151 // Finally, insert the meminst for this element. 2152 if (isa<MemSetInst>(MI)) { 2153 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2154 MI->isVolatile()); 2155 } else { 2156 assert(isa<MemTransferInst>(MI)); 2157 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2158 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2159 2160 if (isa<MemCpyInst>(MI)) 2161 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2162 else 2163 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2164 } 2165 } 2166 DeadInsts.push_back(MI); 2167} 2168 2169/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2170/// overwrites the entire allocation. Extract out the pieces of the stored 2171/// integer and store them individually. 2172void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2173 SmallVector<AllocaInst*, 32> &NewElts){ 2174 // Extract each element out of the integer according to its structure offset 2175 // and store the element value to the individual alloca. 2176 Value *SrcVal = SI->getOperand(0); 2177 const Type *AllocaEltTy = AI->getAllocatedType(); 2178 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2179 2180 IRBuilder<> Builder(SI); 2181 2182 // Handle tail padding by extending the operand 2183 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2184 SrcVal = Builder.CreateZExt(SrcVal, 2185 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2186 2187 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2188 << '\n'); 2189 2190 // There are two forms here: AI could be an array or struct. Both cases 2191 // have different ways to compute the element offset. 2192 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2193 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2194 2195 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2196 // Get the number of bits to shift SrcVal to get the value. 2197 const Type *FieldTy = EltSTy->getElementType(i); 2198 uint64_t Shift = Layout->getElementOffsetInBits(i); 2199 2200 if (TD->isBigEndian()) 2201 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2202 2203 Value *EltVal = SrcVal; 2204 if (Shift) { 2205 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2206 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2207 } 2208 2209 // Truncate down to an integer of the right size. 2210 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2211 2212 // Ignore zero sized fields like {}, they obviously contain no data. 2213 if (FieldSizeBits == 0) continue; 2214 2215 if (FieldSizeBits != AllocaSizeBits) 2216 EltVal = Builder.CreateTrunc(EltVal, 2217 IntegerType::get(SI->getContext(), FieldSizeBits)); 2218 Value *DestField = NewElts[i]; 2219 if (EltVal->getType() == FieldTy) { 2220 // Storing to an integer field of this size, just do it. 2221 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2222 // Bitcast to the right element type (for fp/vector values). 2223 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2224 } else { 2225 // Otherwise, bitcast the dest pointer (for aggregates). 2226 DestField = Builder.CreateBitCast(DestField, 2227 PointerType::getUnqual(EltVal->getType())); 2228 } 2229 new StoreInst(EltVal, DestField, SI); 2230 } 2231 2232 } else { 2233 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2234 const Type *ArrayEltTy = ATy->getElementType(); 2235 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2236 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2237 2238 uint64_t Shift; 2239 2240 if (TD->isBigEndian()) 2241 Shift = AllocaSizeBits-ElementOffset; 2242 else 2243 Shift = 0; 2244 2245 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2246 // Ignore zero sized fields like {}, they obviously contain no data. 2247 if (ElementSizeBits == 0) continue; 2248 2249 Value *EltVal = SrcVal; 2250 if (Shift) { 2251 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2252 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2253 } 2254 2255 // Truncate down to an integer of the right size. 2256 if (ElementSizeBits != AllocaSizeBits) 2257 EltVal = Builder.CreateTrunc(EltVal, 2258 IntegerType::get(SI->getContext(), 2259 ElementSizeBits)); 2260 Value *DestField = NewElts[i]; 2261 if (EltVal->getType() == ArrayEltTy) { 2262 // Storing to an integer field of this size, just do it. 2263 } else if (ArrayEltTy->isFloatingPointTy() || 2264 ArrayEltTy->isVectorTy()) { 2265 // Bitcast to the right element type (for fp/vector values). 2266 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2267 } else { 2268 // Otherwise, bitcast the dest pointer (for aggregates). 2269 DestField = Builder.CreateBitCast(DestField, 2270 PointerType::getUnqual(EltVal->getType())); 2271 } 2272 new StoreInst(EltVal, DestField, SI); 2273 2274 if (TD->isBigEndian()) 2275 Shift -= ElementOffset; 2276 else 2277 Shift += ElementOffset; 2278 } 2279 } 2280 2281 DeadInsts.push_back(SI); 2282} 2283 2284/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2285/// an integer. Load the individual pieces to form the aggregate value. 2286void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2287 SmallVector<AllocaInst*, 32> &NewElts) { 2288 // Extract each element out of the NewElts according to its structure offset 2289 // and form the result value. 2290 const Type *AllocaEltTy = AI->getAllocatedType(); 2291 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2292 2293 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2294 << '\n'); 2295 2296 // There are two forms here: AI could be an array or struct. Both cases 2297 // have different ways to compute the element offset. 2298 const StructLayout *Layout = 0; 2299 uint64_t ArrayEltBitOffset = 0; 2300 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2301 Layout = TD->getStructLayout(EltSTy); 2302 } else { 2303 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2304 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2305 } 2306 2307 Value *ResultVal = 2308 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2309 2310 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2311 // Load the value from the alloca. If the NewElt is an aggregate, cast 2312 // the pointer to an integer of the same size before doing the load. 2313 Value *SrcField = NewElts[i]; 2314 const Type *FieldTy = 2315 cast<PointerType>(SrcField->getType())->getElementType(); 2316 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2317 2318 // Ignore zero sized fields like {}, they obviously contain no data. 2319 if (FieldSizeBits == 0) continue; 2320 2321 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2322 FieldSizeBits); 2323 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2324 !FieldTy->isVectorTy()) 2325 SrcField = new BitCastInst(SrcField, 2326 PointerType::getUnqual(FieldIntTy), 2327 "", LI); 2328 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2329 2330 // If SrcField is a fp or vector of the right size but that isn't an 2331 // integer type, bitcast to an integer so we can shift it. 2332 if (SrcField->getType() != FieldIntTy) 2333 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2334 2335 // Zero extend the field to be the same size as the final alloca so that 2336 // we can shift and insert it. 2337 if (SrcField->getType() != ResultVal->getType()) 2338 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2339 2340 // Determine the number of bits to shift SrcField. 2341 uint64_t Shift; 2342 if (Layout) // Struct case. 2343 Shift = Layout->getElementOffsetInBits(i); 2344 else // Array case. 2345 Shift = i*ArrayEltBitOffset; 2346 2347 if (TD->isBigEndian()) 2348 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2349 2350 if (Shift) { 2351 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2352 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2353 } 2354 2355 // Don't create an 'or x, 0' on the first iteration. 2356 if (!isa<Constant>(ResultVal) || 2357 !cast<Constant>(ResultVal)->isNullValue()) 2358 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2359 else 2360 ResultVal = SrcField; 2361 } 2362 2363 // Handle tail padding by truncating the result 2364 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2365 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2366 2367 LI->replaceAllUsesWith(ResultVal); 2368 DeadInsts.push_back(LI); 2369} 2370 2371/// HasPadding - Return true if the specified type has any structure or 2372/// alignment padding in between the elements that would be split apart 2373/// by SROA; return false otherwise. 2374static bool HasPadding(const Type *Ty, const TargetData &TD) { 2375 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2376 Ty = ATy->getElementType(); 2377 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2378 } 2379 2380 // SROA currently handles only Arrays and Structs. 2381 const StructType *STy = cast<StructType>(Ty); 2382 const StructLayout *SL = TD.getStructLayout(STy); 2383 unsigned PrevFieldBitOffset = 0; 2384 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2385 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2386 2387 // Check to see if there is any padding between this element and the 2388 // previous one. 2389 if (i) { 2390 unsigned PrevFieldEnd = 2391 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2392 if (PrevFieldEnd < FieldBitOffset) 2393 return true; 2394 } 2395 PrevFieldBitOffset = FieldBitOffset; 2396 } 2397 // Check for tail padding. 2398 if (unsigned EltCount = STy->getNumElements()) { 2399 unsigned PrevFieldEnd = PrevFieldBitOffset + 2400 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2401 if (PrevFieldEnd < SL->getSizeInBits()) 2402 return true; 2403 } 2404 return false; 2405} 2406 2407/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2408/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2409/// or 1 if safe after canonicalization has been performed. 2410bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2411 // Loop over the use list of the alloca. We can only transform it if all of 2412 // the users are safe to transform. 2413 AllocaInfo Info(AI); 2414 2415 isSafeForScalarRepl(AI, 0, Info); 2416 if (Info.isUnsafe) { 2417 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2418 return false; 2419 } 2420 2421 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2422 // source and destination, we have to be careful. In particular, the memcpy 2423 // could be moving around elements that live in structure padding of the LLVM 2424 // types, but may actually be used. In these cases, we refuse to promote the 2425 // struct. 2426 if (Info.isMemCpySrc && Info.isMemCpyDst && 2427 HasPadding(AI->getAllocatedType(), *TD)) 2428 return false; 2429 2430 // If the alloca never has an access to just *part* of it, but is accessed 2431 // via loads and stores, then we should use ConvertToScalarInfo to promote 2432 // the alloca instead of promoting each piece at a time and inserting fission 2433 // and fusion code. 2434 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2435 // If the struct/array just has one element, use basic SRoA. 2436 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2437 if (ST->getNumElements() > 1) return false; 2438 } else { 2439 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2440 return false; 2441 } 2442 } 2443 2444 return true; 2445} 2446 2447 2448 2449/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2450/// some part of a constant global variable. This intentionally only accepts 2451/// constant expressions because we don't can't rewrite arbitrary instructions. 2452static bool PointsToConstantGlobal(Value *V) { 2453 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2454 return GV->isConstant(); 2455 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2456 if (CE->getOpcode() == Instruction::BitCast || 2457 CE->getOpcode() == Instruction::GetElementPtr) 2458 return PointsToConstantGlobal(CE->getOperand(0)); 2459 return false; 2460} 2461 2462/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2463/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2464/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2465/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2466/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2467/// the alloca, and if the source pointer is a pointer to a constant global, we 2468/// can optimize this. 2469static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2470 bool isOffset) { 2471 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2472 User *U = cast<Instruction>(*UI); 2473 2474 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2475 // Ignore non-volatile loads, they are always ok. 2476 if (LI->isVolatile()) return false; 2477 continue; 2478 } 2479 2480 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2481 // If uses of the bitcast are ok, we are ok. 2482 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 2483 return false; 2484 continue; 2485 } 2486 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2487 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2488 // doesn't, it does. 2489 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2490 isOffset || !GEP->hasAllZeroIndices())) 2491 return false; 2492 continue; 2493 } 2494 2495 if (CallSite CS = U) { 2496 // If this is the function being called then we treat it like a load and 2497 // ignore it. 2498 if (CS.isCallee(UI)) 2499 continue; 2500 2501 // If this is a readonly/readnone call site, then we know it is just a 2502 // load (but one that potentially returns the value itself), so we can 2503 // ignore it if we know that the value isn't captured. 2504 unsigned ArgNo = CS.getArgumentNo(UI); 2505 if (CS.onlyReadsMemory() && 2506 (CS.getInstruction()->use_empty() || 2507 CS.paramHasAttr(ArgNo+1, Attribute::NoCapture))) 2508 continue; 2509 2510 // If this is being passed as a byval argument, the caller is making a 2511 // copy, so it is only a read of the alloca. 2512 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2513 continue; 2514 } 2515 2516 // If this is isn't our memcpy/memmove, reject it as something we can't 2517 // handle. 2518 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2519 if (MI == 0) 2520 return false; 2521 2522 // If the transfer is using the alloca as a source of the transfer, then 2523 // ignore it since it is a load (unless the transfer is volatile). 2524 if (UI.getOperandNo() == 1) { 2525 if (MI->isVolatile()) return false; 2526 continue; 2527 } 2528 2529 // If we already have seen a copy, reject the second one. 2530 if (TheCopy) return false; 2531 2532 // If the pointer has been offset from the start of the alloca, we can't 2533 // safely handle this. 2534 if (isOffset) return false; 2535 2536 // If the memintrinsic isn't using the alloca as the dest, reject it. 2537 if (UI.getOperandNo() != 0) return false; 2538 2539 // If the source of the memcpy/move is not a constant global, reject it. 2540 if (!PointsToConstantGlobal(MI->getSource())) 2541 return false; 2542 2543 // Otherwise, the transform is safe. Remember the copy instruction. 2544 TheCopy = MI; 2545 } 2546 return true; 2547} 2548 2549/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2550/// modified by a copy from a constant global. If we can prove this, we can 2551/// replace any uses of the alloca with uses of the global directly. 2552MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 2553 MemTransferInst *TheCopy = 0; 2554 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 2555 return TheCopy; 2556 return 0; 2557} 2558