ScalarReplAggregates.cpp revision 032c10fee2a4bb731488ce75844878009d3bd409
1//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This transformation implements the well known scalar replacement of 11// aggregates transformation. This xform breaks up alloca instructions of 12// aggregate type (structure or array) into individual alloca instructions for 13// each member (if possible). Then, if possible, it transforms the individual 14// alloca instructions into nice clean scalar SSA form. 15// 16// This combines a simple SRoA algorithm with the Mem2Reg algorithm because 17// often interact, especially for C++ programs. As such, iterating between 18// SRoA, then Mem2Reg until we run out of things to promote works well. 19// 20//===----------------------------------------------------------------------===// 21 22#define DEBUG_TYPE "scalarrepl" 23#include "llvm/Transforms/Scalar.h" 24#include "llvm/Constants.h" 25#include "llvm/DerivedTypes.h" 26#include "llvm/Function.h" 27#include "llvm/GlobalVariable.h" 28#include "llvm/Instructions.h" 29#include "llvm/IntrinsicInst.h" 30#include "llvm/LLVMContext.h" 31#include "llvm/Module.h" 32#include "llvm/Pass.h" 33#include "llvm/Analysis/Dominators.h" 34#include "llvm/Analysis/Loads.h" 35#include "llvm/Analysis/ValueTracking.h" 36#include "llvm/Target/TargetData.h" 37#include "llvm/Transforms/Utils/PromoteMemToReg.h" 38#include "llvm/Transforms/Utils/Local.h" 39#include "llvm/Transforms/Utils/SSAUpdater.h" 40#include "llvm/Support/CallSite.h" 41#include "llvm/Support/Debug.h" 42#include "llvm/Support/ErrorHandling.h" 43#include "llvm/Support/GetElementPtrTypeIterator.h" 44#include "llvm/Support/IRBuilder.h" 45#include "llvm/Support/MathExtras.h" 46#include "llvm/Support/raw_ostream.h" 47#include "llvm/ADT/SetVector.h" 48#include "llvm/ADT/SmallVector.h" 49#include "llvm/ADT/Statistic.h" 50using namespace llvm; 51 52STATISTIC(NumReplaced, "Number of allocas broken up"); 53STATISTIC(NumPromoted, "Number of allocas promoted"); 54STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); 55STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 56STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 57 58namespace { 59 struct SROA : public FunctionPass { 60 SROA(int T, bool hasDT, char &ID) 61 : FunctionPass(ID), HasDomTree(hasDT) { 62 if (T == -1) 63 SRThreshold = 128; 64 else 65 SRThreshold = T; 66 } 67 68 bool runOnFunction(Function &F); 69 70 bool performScalarRepl(Function &F); 71 bool performPromotion(Function &F); 72 73 private: 74 bool HasDomTree; 75 TargetData *TD; 76 77 /// DeadInsts - Keep track of instructions we have made dead, so that 78 /// we can remove them after we are done working. 79 SmallVector<Value*, 32> DeadInsts; 80 81 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 82 /// information about the uses. All these fields are initialized to false 83 /// and set to true when something is learned. 84 struct AllocaInfo { 85 /// The alloca to promote. 86 AllocaInst *AI; 87 88 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite 89 /// looping and avoid redundant work. 90 SmallPtrSet<PHINode*, 8> CheckedPHIs; 91 92 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 93 bool isUnsafe : 1; 94 95 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 96 bool isMemCpySrc : 1; 97 98 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 99 bool isMemCpyDst : 1; 100 101 /// hasSubelementAccess - This is true if a subelement of the alloca is 102 /// ever accessed, or false if the alloca is only accessed with mem 103 /// intrinsics or load/store that only access the entire alloca at once. 104 bool hasSubelementAccess : 1; 105 106 /// hasALoadOrStore - This is true if there are any loads or stores to it. 107 /// The alloca may just be accessed with memcpy, for example, which would 108 /// not set this. 109 bool hasALoadOrStore : 1; 110 111 explicit AllocaInfo(AllocaInst *ai) 112 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 113 hasSubelementAccess(false), hasALoadOrStore(false) {} 114 }; 115 116 unsigned SRThreshold; 117 118 void MarkUnsafe(AllocaInfo &I, Instruction *User) { 119 I.isUnsafe = true; 120 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); 121 } 122 123 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 124 125 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info); 126 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset, 127 AllocaInfo &Info); 128 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info); 129 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 130 const Type *MemOpType, bool isStore, AllocaInfo &Info, 131 Instruction *TheAccess, bool AllowWholeAccess); 132 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 133 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 134 const Type *&IdxTy); 135 136 void DoScalarReplacement(AllocaInst *AI, 137 std::vector<AllocaInst*> &WorkList); 138 void DeleteDeadInstructions(); 139 140 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 141 SmallVector<AllocaInst*, 32> &NewElts); 142 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 143 SmallVector<AllocaInst*, 32> &NewElts); 144 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 145 SmallVector<AllocaInst*, 32> &NewElts); 146 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 147 AllocaInst *AI, 148 SmallVector<AllocaInst*, 32> &NewElts); 149 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 150 SmallVector<AllocaInst*, 32> &NewElts); 151 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 152 SmallVector<AllocaInst*, 32> &NewElts); 153 154 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 155 }; 156 157 // SROA_DT - SROA that uses DominatorTree. 158 struct SROA_DT : public SROA { 159 static char ID; 160 public: 161 SROA_DT(int T = -1) : SROA(T, true, ID) { 162 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 163 } 164 165 // getAnalysisUsage - This pass does not require any passes, but we know it 166 // will not alter the CFG, so say so. 167 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 168 AU.addRequired<DominatorTree>(); 169 AU.setPreservesCFG(); 170 } 171 }; 172 173 // SROA_SSAUp - SROA that uses SSAUpdater. 174 struct SROA_SSAUp : public SROA { 175 static char ID; 176 public: 177 SROA_SSAUp(int T = -1) : SROA(T, false, ID) { 178 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 179 } 180 181 // getAnalysisUsage - This pass does not require any passes, but we know it 182 // will not alter the CFG, so say so. 183 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 184 AU.setPreservesCFG(); 185 } 186 }; 187 188} 189 190char SROA_DT::ID = 0; 191char SROA_SSAUp::ID = 0; 192 193INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 194 "Scalar Replacement of Aggregates (DT)", false, false) 195INITIALIZE_PASS_DEPENDENCY(DominatorTree) 196INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 197 "Scalar Replacement of Aggregates (DT)", false, false) 198 199INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 200 "Scalar Replacement of Aggregates (SSAUp)", false, false) 201INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 202 "Scalar Replacement of Aggregates (SSAUp)", false, false) 203 204// Public interface to the ScalarReplAggregates pass 205FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 206 bool UseDomTree) { 207 if (UseDomTree) 208 return new SROA_DT(Threshold); 209 return new SROA_SSAUp(Threshold); 210} 211 212 213//===----------------------------------------------------------------------===// 214// Convert To Scalar Optimization. 215//===----------------------------------------------------------------------===// 216 217namespace { 218/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 219/// optimization, which scans the uses of an alloca and determines if it can 220/// rewrite it in terms of a single new alloca that can be mem2reg'd. 221class ConvertToScalarInfo { 222 /// AllocaSize - The size of the alloca being considered. 223 unsigned AllocaSize; 224 const TargetData &TD; 225 226 /// IsNotTrivial - This is set to true if there is some access to the object 227 /// which means that mem2reg can't promote it. 228 bool IsNotTrivial; 229 230 /// VectorTy - This tracks the type that we should promote the vector to if 231 /// it is possible to turn it into a vector. This starts out null, and if it 232 /// isn't possible to turn into a vector type, it gets set to VoidTy. 233 const Type *VectorTy; 234 235 /// HadAVector - True if there is at least one vector access to the alloca. 236 /// We don't want to turn random arrays into vectors and use vector element 237 /// insert/extract, but if there are element accesses to something that is 238 /// also declared as a vector, we do want to promote to a vector. 239 bool HadAVector; 240 241public: 242 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 243 : AllocaSize(Size), TD(td) { 244 IsNotTrivial = false; 245 VectorTy = 0; 246 HadAVector = false; 247 } 248 249 AllocaInst *TryConvert(AllocaInst *AI); 250 251private: 252 bool CanConvertToScalar(Value *V, uint64_t Offset); 253 void MergeInType(const Type *In, uint64_t Offset); 254 bool MergeInVectorType(const VectorType *VInTy, uint64_t Offset); 255 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 256 257 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 258 uint64_t Offset, IRBuilder<> &Builder); 259 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 260 uint64_t Offset, IRBuilder<> &Builder); 261}; 262} // end anonymous namespace. 263 264 265/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 266/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 267/// alloca if possible or null if not. 268AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 269 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 270 // out. 271 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 272 return 0; 273 274 // If we were able to find a vector type that can handle this with 275 // insert/extract elements, and if there was at least one use that had 276 // a vector type, promote this to a vector. We don't want to promote 277 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 278 // we just get a lot of insert/extracts. If at least one vector is 279 // involved, then we probably really do have a union of vector/array. 280 const Type *NewTy; 281 if (VectorTy && VectorTy->isVectorTy() && HadAVector) { 282 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 283 << *VectorTy << '\n'); 284 NewTy = VectorTy; // Use the vector type. 285 } else { 286 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 287 // Create and insert the integer alloca. 288 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 289 } 290 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 291 ConvertUsesToScalar(AI, NewAI, 0); 292 return NewAI; 293} 294 295/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy) 296/// so far at the offset specified by Offset (which is specified in bytes). 297/// 298/// There are three cases we handle here: 299/// 1) A union of vector types of the same size and potentially its elements. 300/// Here we turn element accesses into insert/extract element operations. 301/// This promotes a <4 x float> with a store of float to the third element 302/// into a <4 x float> that uses insert element. 303/// 2) A union of vector types with power-of-2 size differences, e.g. a float, 304/// <2 x float> and <4 x float>. Here we turn element accesses into insert 305/// and extract element operations, and <2 x float> accesses into a cast to 306/// <2 x double>, an extract, and a cast back to <2 x float>. 307/// 3) A fully general blob of memory, which we turn into some (potentially 308/// large) integer type with extract and insert operations where the loads 309/// and stores would mutate the memory. We mark this by setting VectorTy 310/// to VoidTy. 311void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) { 312 // If we already decided to turn this into a blob of integer memory, there is 313 // nothing to be done. 314 if (VectorTy && VectorTy->isVoidTy()) 315 return; 316 317 // If this could be contributing to a vector, analyze it. 318 319 // If the In type is a vector that is the same size as the alloca, see if it 320 // matches the existing VecTy. 321 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 322 if (MergeInVectorType(VInTy, Offset)) 323 return; 324 } else if (In->isFloatTy() || In->isDoubleTy() || 325 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 326 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 327 // If we're accessing something that could be an element of a vector, see 328 // if the implied vector agrees with what we already have and if Offset is 329 // compatible with it. 330 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 331 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 332 (VectorTy == 0 || 333 cast<VectorType>(VectorTy)->getElementType() 334 ->getPrimitiveSizeInBits()/8 == EltSize)) { 335 if (VectorTy == 0) 336 VectorTy = VectorType::get(In, AllocaSize/EltSize); 337 return; 338 } 339 } 340 341 // Otherwise, we have a case that we can't handle with an optimized vector 342 // form. We can still turn this into a large integer. 343 VectorTy = Type::getVoidTy(In->getContext()); 344} 345 346/// MergeInVectorType - Handles the vector case of MergeInType, returning true 347/// if the type was successfully merged and false otherwise. 348bool ConvertToScalarInfo::MergeInVectorType(const VectorType *VInTy, 349 uint64_t Offset) { 350 // Remember if we saw a vector type. 351 HadAVector = true; 352 353 // TODO: Support nonzero offsets? 354 if (Offset != 0) 355 return false; 356 357 // Only allow vectors that are a power-of-2 away from the size of the alloca. 358 if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8))) 359 return false; 360 361 // If this the first vector we see, remember the type so that we know the 362 // element size. 363 if (!VectorTy) { 364 VectorTy = VInTy; 365 return true; 366 } 367 368 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth(); 369 unsigned InBitWidth = VInTy->getBitWidth(); 370 371 // Vectors of the same size can be converted using a simple bitcast. 372 if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8)) 373 return true; 374 375 const Type *ElementTy = cast<VectorType>(VectorTy)->getElementType(); 376 const Type *InElementTy = cast<VectorType>(VectorTy)->getElementType(); 377 378 // Do not allow mixed integer and floating-point accesses from vectors of 379 // different sizes. 380 if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy()) 381 return false; 382 383 if (ElementTy->isFloatingPointTy()) { 384 // Only allow floating-point vectors of different sizes if they have the 385 // same element type. 386 // TODO: This could be loosened a bit, but would anything benefit? 387 if (ElementTy != InElementTy) 388 return false; 389 390 // There are no arbitrary-precision floating-point types, which limits the 391 // number of legal vector types with larger element types that we can form 392 // to bitcast and extract a subvector. 393 // TODO: We could support some more cases with mixed fp128 and double here. 394 if (!(BitWidth == 64 || BitWidth == 128) || 395 !(InBitWidth == 64 || InBitWidth == 128)) 396 return false; 397 } else { 398 assert(ElementTy->isIntegerTy() && "Vector elements must be either integer " 399 "or floating-point."); 400 unsigned BitWidth = ElementTy->getPrimitiveSizeInBits(); 401 unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits(); 402 403 // Do not allow integer types smaller than a byte or types whose widths are 404 // not a multiple of a byte. 405 if (BitWidth < 8 || InBitWidth < 8 || 406 BitWidth % 8 != 0 || InBitWidth % 8 != 0) 407 return false; 408 } 409 410 // Pick the largest of the two vector types. 411 if (InBitWidth > BitWidth) 412 VectorTy = VInTy; 413 414 return true; 415} 416 417/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 418/// its accesses to a single vector type, return true and set VecTy to 419/// the new type. If we could convert the alloca into a single promotable 420/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 421/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 422/// is the current offset from the base of the alloca being analyzed. 423/// 424/// If we see at least one access to the value that is as a vector type, set the 425/// SawVec flag. 426bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 427 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 428 Instruction *User = cast<Instruction>(*UI); 429 430 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 431 // Don't break volatile loads. 432 if (LI->isVolatile()) 433 return false; 434 // Don't touch MMX operations. 435 if (LI->getType()->isX86_MMXTy()) 436 return false; 437 MergeInType(LI->getType(), Offset); 438 continue; 439 } 440 441 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 442 // Storing the pointer, not into the value? 443 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 444 // Don't touch MMX operations. 445 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 446 return false; 447 MergeInType(SI->getOperand(0)->getType(), Offset); 448 continue; 449 } 450 451 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 452 IsNotTrivial = true; // Can't be mem2reg'd. 453 if (!CanConvertToScalar(BCI, Offset)) 454 return false; 455 continue; 456 } 457 458 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 459 // If this is a GEP with a variable indices, we can't handle it. 460 if (!GEP->hasAllConstantIndices()) 461 return false; 462 463 // Compute the offset that this GEP adds to the pointer. 464 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 465 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 466 &Indices[0], Indices.size()); 467 // See if all uses can be converted. 468 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 469 return false; 470 IsNotTrivial = true; // Can't be mem2reg'd. 471 continue; 472 } 473 474 // If this is a constant sized memset of a constant value (e.g. 0) we can 475 // handle it. 476 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 477 // Store of constant value and constant size. 478 if (!isa<ConstantInt>(MSI->getValue()) || 479 !isa<ConstantInt>(MSI->getLength())) 480 return false; 481 IsNotTrivial = true; // Can't be mem2reg'd. 482 continue; 483 } 484 485 // If this is a memcpy or memmove into or out of the whole allocation, we 486 // can handle it like a load or store of the scalar type. 487 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 488 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 489 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 490 return false; 491 492 IsNotTrivial = true; // Can't be mem2reg'd. 493 continue; 494 } 495 496 // Otherwise, we cannot handle this! 497 return false; 498 } 499 500 return true; 501} 502 503/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 504/// directly. This happens when we are converting an "integer union" to a 505/// single integer scalar, or when we are converting a "vector union" to a 506/// vector with insert/extractelement instructions. 507/// 508/// Offset is an offset from the original alloca, in bits that need to be 509/// shifted to the right. By the end of this, there should be no uses of Ptr. 510void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 511 uint64_t Offset) { 512 while (!Ptr->use_empty()) { 513 Instruction *User = cast<Instruction>(Ptr->use_back()); 514 515 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 516 ConvertUsesToScalar(CI, NewAI, Offset); 517 CI->eraseFromParent(); 518 continue; 519 } 520 521 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 522 // Compute the offset that this GEP adds to the pointer. 523 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 524 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 525 &Indices[0], Indices.size()); 526 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 527 GEP->eraseFromParent(); 528 continue; 529 } 530 531 IRBuilder<> Builder(User); 532 533 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 534 // The load is a bit extract from NewAI shifted right by Offset bits. 535 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 536 Value *NewLoadVal 537 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 538 LI->replaceAllUsesWith(NewLoadVal); 539 LI->eraseFromParent(); 540 continue; 541 } 542 543 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 544 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 545 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 546 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 547 Builder); 548 Builder.CreateStore(New, NewAI); 549 SI->eraseFromParent(); 550 551 // If the load we just inserted is now dead, then the inserted store 552 // overwrote the entire thing. 553 if (Old->use_empty()) 554 Old->eraseFromParent(); 555 continue; 556 } 557 558 // If this is a constant sized memset of a constant value (e.g. 0) we can 559 // transform it into a store of the expanded constant value. 560 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 561 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 562 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 563 if (NumBytes != 0) { 564 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 565 566 // Compute the value replicated the right number of times. 567 APInt APVal(NumBytes*8, Val); 568 569 // Splat the value if non-zero. 570 if (Val) 571 for (unsigned i = 1; i != NumBytes; ++i) 572 APVal |= APVal << 8; 573 574 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 575 Value *New = ConvertScalar_InsertValue( 576 ConstantInt::get(User->getContext(), APVal), 577 Old, Offset, Builder); 578 Builder.CreateStore(New, NewAI); 579 580 // If the load we just inserted is now dead, then the memset overwrote 581 // the entire thing. 582 if (Old->use_empty()) 583 Old->eraseFromParent(); 584 } 585 MSI->eraseFromParent(); 586 continue; 587 } 588 589 // If this is a memcpy or memmove into or out of the whole allocation, we 590 // can handle it like a load or store of the scalar type. 591 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 592 assert(Offset == 0 && "must be store to start of alloca"); 593 594 // If the source and destination are both to the same alloca, then this is 595 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 596 // as appropriate. 597 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 598 599 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 600 // Dest must be OrigAI, change this to be a load from the original 601 // pointer (bitcasted), then a store to our new alloca. 602 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 603 Value *SrcPtr = MTI->getSource(); 604 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 605 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 606 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 607 AIPTy = PointerType::get(AIPTy->getElementType(), 608 SPTy->getAddressSpace()); 609 } 610 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 611 612 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 613 SrcVal->setAlignment(MTI->getAlignment()); 614 Builder.CreateStore(SrcVal, NewAI); 615 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 616 // Src must be OrigAI, change this to be a load from NewAI then a store 617 // through the original dest pointer (bitcasted). 618 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 619 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 620 621 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 622 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 623 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 624 AIPTy = PointerType::get(AIPTy->getElementType(), 625 DPTy->getAddressSpace()); 626 } 627 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 628 629 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 630 NewStore->setAlignment(MTI->getAlignment()); 631 } else { 632 // Noop transfer. Src == Dst 633 } 634 635 MTI->eraseFromParent(); 636 continue; 637 } 638 639 llvm_unreachable("Unsupported operation!"); 640 } 641} 642 643/// getScaledElementType - Gets a scaled element type for a partial vector 644/// access of an alloca. The input type must be an integer or float, and 645/// the resulting type must be an integer, float or double. 646static const Type *getScaledElementType(const Type *OldTy, unsigned Scale) { 647 assert((OldTy->isIntegerTy() || OldTy->isFloatTy()) && "Partial vector " 648 "accesses must be scaled from integer or float elements."); 649 650 LLVMContext &Context = OldTy->getContext(); 651 unsigned Size = OldTy->getPrimitiveSizeInBits() * Scale; 652 653 if (OldTy->isIntegerTy()) 654 return Type::getIntNTy(Context, Size); 655 if (Size == 32) 656 return Type::getFloatTy(Context); 657 if (Size == 64) 658 return Type::getDoubleTy(Context); 659 660 llvm_unreachable("Invalid type for a partial vector access of an alloca!"); 661} 662 663/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 664/// or vector value FromVal, extracting the bits from the offset specified by 665/// Offset. This returns the value, which is of type ToType. 666/// 667/// This happens when we are converting an "integer union" to a single 668/// integer scalar, or when we are converting a "vector union" to a vector with 669/// insert/extractelement instructions. 670/// 671/// Offset is an offset from the original alloca, in bits that need to be 672/// shifted to the right. 673Value *ConvertToScalarInfo:: 674ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 675 uint64_t Offset, IRBuilder<> &Builder) { 676 // If the load is of the whole new alloca, no conversion is needed. 677 if (FromVal->getType() == ToType && Offset == 0) 678 return FromVal; 679 680 // If the result alloca is a vector type, this is either an element 681 // access or a bitcast to another vector type of the same size. 682 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 683 if (ToType->isVectorTy()) { 684 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 685 if (ToTypeSize == AllocaSize) 686 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 687 688 assert(isPowerOf2_64(AllocaSize / ToTypeSize) && 689 "Partial vector access of an alloca must have a power-of-2 size " 690 "ratio."); 691 assert(Offset == 0 && "Can't extract a value of a smaller vector type " 692 "from a nonzero offset."); 693 694 const Type *ToElementTy = cast<VectorType>(ToType)->getElementType(); 695 unsigned Scale = AllocaSize / ToTypeSize; 696 const Type *CastElementTy = getScaledElementType(ToElementTy, Scale); 697 unsigned NumCastVectorElements = VTy->getNumElements() / Scale; 698 699 LLVMContext &Context = FromVal->getContext(); 700 const Type *CastTy = VectorType::get(CastElementTy, 701 NumCastVectorElements); 702 Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp"); 703 Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get( 704 Type::getInt32Ty(Context), 0), "tmp"); 705 return Builder.CreateBitCast(Extract, ToType, "tmp"); 706 } 707 708 // Otherwise it must be an element access. 709 unsigned Elt = 0; 710 if (Offset) { 711 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 712 Elt = Offset/EltSize; 713 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 714 } 715 // Return the element extracted out of it. 716 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 717 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 718 if (V->getType() != ToType) 719 V = Builder.CreateBitCast(V, ToType, "tmp"); 720 return V; 721 } 722 723 // If ToType is a first class aggregate, extract out each of the pieces and 724 // use insertvalue's to form the FCA. 725 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 726 const StructLayout &Layout = *TD.getStructLayout(ST); 727 Value *Res = UndefValue::get(ST); 728 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 729 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 730 Offset+Layout.getElementOffsetInBits(i), 731 Builder); 732 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 733 } 734 return Res; 735 } 736 737 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 738 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 739 Value *Res = UndefValue::get(AT); 740 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 741 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 742 Offset+i*EltSize, Builder); 743 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 744 } 745 return Res; 746 } 747 748 // Otherwise, this must be a union that was converted to an integer value. 749 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 750 751 // If this is a big-endian system and the load is narrower than the 752 // full alloca type, we need to do a shift to get the right bits. 753 int ShAmt = 0; 754 if (TD.isBigEndian()) { 755 // On big-endian machines, the lowest bit is stored at the bit offset 756 // from the pointer given by getTypeStoreSizeInBits. This matters for 757 // integers with a bitwidth that is not a multiple of 8. 758 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 759 TD.getTypeStoreSizeInBits(ToType) - Offset; 760 } else { 761 ShAmt = Offset; 762 } 763 764 // Note: we support negative bitwidths (with shl) which are not defined. 765 // We do this to support (f.e.) loads off the end of a structure where 766 // only some bits are used. 767 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 768 FromVal = Builder.CreateLShr(FromVal, 769 ConstantInt::get(FromVal->getType(), 770 ShAmt), "tmp"); 771 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 772 FromVal = Builder.CreateShl(FromVal, 773 ConstantInt::get(FromVal->getType(), 774 -ShAmt), "tmp"); 775 776 // Finally, unconditionally truncate the integer to the right width. 777 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 778 if (LIBitWidth < NTy->getBitWidth()) 779 FromVal = 780 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 781 LIBitWidth), "tmp"); 782 else if (LIBitWidth > NTy->getBitWidth()) 783 FromVal = 784 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 785 LIBitWidth), "tmp"); 786 787 // If the result is an integer, this is a trunc or bitcast. 788 if (ToType->isIntegerTy()) { 789 // Should be done. 790 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 791 // Just do a bitcast, we know the sizes match up. 792 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 793 } else { 794 // Otherwise must be a pointer. 795 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 796 } 797 assert(FromVal->getType() == ToType && "Didn't convert right?"); 798 return FromVal; 799} 800 801/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 802/// or vector value "Old" at the offset specified by Offset. 803/// 804/// This happens when we are converting an "integer union" to a 805/// single integer scalar, or when we are converting a "vector union" to a 806/// vector with insert/extractelement instructions. 807/// 808/// Offset is an offset from the original alloca, in bits that need to be 809/// shifted to the right. 810Value *ConvertToScalarInfo:: 811ConvertScalar_InsertValue(Value *SV, Value *Old, 812 uint64_t Offset, IRBuilder<> &Builder) { 813 // Convert the stored type to the actual type, shift it left to insert 814 // then 'or' into place. 815 const Type *AllocaType = Old->getType(); 816 LLVMContext &Context = Old->getContext(); 817 818 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 819 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 820 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 821 822 // Changing the whole vector with memset or with an access of a different 823 // vector type? 824 if (ValSize == VecSize) 825 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 826 827 if (SV->getType()->isVectorTy() && isPowerOf2_64(VecSize / ValSize)) { 828 assert(Offset == 0 && "Can't insert a value of a smaller vector type at " 829 "a nonzero offset."); 830 831 const Type *ToElementTy = 832 cast<VectorType>(SV->getType())->getElementType(); 833 unsigned Scale = VecSize / ValSize; 834 const Type *CastElementTy = getScaledElementType(ToElementTy, Scale); 835 unsigned NumCastVectorElements = VTy->getNumElements() / Scale; 836 837 LLVMContext &Context = SV->getContext(); 838 const Type *OldCastTy = VectorType::get(CastElementTy, 839 NumCastVectorElements); 840 Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp"); 841 842 Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp"); 843 Value *Insert = 844 Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get( 845 Type::getInt32Ty(Context), 0), "tmp"); 846 return Builder.CreateBitCast(Insert, AllocaType, "tmp"); 847 } 848 849 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 850 851 // Must be an element insertion. 852 unsigned Elt = Offset/EltSize; 853 854 if (SV->getType() != VTy->getElementType()) 855 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 856 857 SV = Builder.CreateInsertElement(Old, SV, 858 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 859 "tmp"); 860 return SV; 861 } 862 863 // If SV is a first-class aggregate value, insert each value recursively. 864 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 865 const StructLayout &Layout = *TD.getStructLayout(ST); 866 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 867 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 868 Old = ConvertScalar_InsertValue(Elt, Old, 869 Offset+Layout.getElementOffsetInBits(i), 870 Builder); 871 } 872 return Old; 873 } 874 875 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 876 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 877 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 878 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 879 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 880 } 881 return Old; 882 } 883 884 // If SV is a float, convert it to the appropriate integer type. 885 // If it is a pointer, do the same. 886 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 887 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 888 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 889 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 890 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 891 SV = Builder.CreateBitCast(SV, 892 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 893 else if (SV->getType()->isPointerTy()) 894 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 895 896 // Zero extend or truncate the value if needed. 897 if (SV->getType() != AllocaType) { 898 if (SV->getType()->getPrimitiveSizeInBits() < 899 AllocaType->getPrimitiveSizeInBits()) 900 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 901 else { 902 // Truncation may be needed if storing more than the alloca can hold 903 // (undefined behavior). 904 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 905 SrcWidth = DestWidth; 906 SrcStoreWidth = DestStoreWidth; 907 } 908 } 909 910 // If this is a big-endian system and the store is narrower than the 911 // full alloca type, we need to do a shift to get the right bits. 912 int ShAmt = 0; 913 if (TD.isBigEndian()) { 914 // On big-endian machines, the lowest bit is stored at the bit offset 915 // from the pointer given by getTypeStoreSizeInBits. This matters for 916 // integers with a bitwidth that is not a multiple of 8. 917 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 918 } else { 919 ShAmt = Offset; 920 } 921 922 // Note: we support negative bitwidths (with shr) which are not defined. 923 // We do this to support (f.e.) stores off the end of a structure where 924 // only some bits in the structure are set. 925 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 926 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 927 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 928 ShAmt), "tmp"); 929 Mask <<= ShAmt; 930 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 931 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 932 -ShAmt), "tmp"); 933 Mask = Mask.lshr(-ShAmt); 934 } 935 936 // Mask out the bits we are about to insert from the old value, and or 937 // in the new bits. 938 if (SrcWidth != DestWidth) { 939 assert(DestWidth > SrcWidth); 940 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 941 SV = Builder.CreateOr(Old, SV, "ins"); 942 } 943 return SV; 944} 945 946 947//===----------------------------------------------------------------------===// 948// SRoA Driver 949//===----------------------------------------------------------------------===// 950 951 952bool SROA::runOnFunction(Function &F) { 953 TD = getAnalysisIfAvailable<TargetData>(); 954 955 bool Changed = performPromotion(F); 956 957 // FIXME: ScalarRepl currently depends on TargetData more than it 958 // theoretically needs to. It should be refactored in order to support 959 // target-independent IR. Until this is done, just skip the actual 960 // scalar-replacement portion of this pass. 961 if (!TD) return Changed; 962 963 while (1) { 964 bool LocalChange = performScalarRepl(F); 965 if (!LocalChange) break; // No need to repromote if no scalarrepl 966 Changed = true; 967 LocalChange = performPromotion(F); 968 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 969 } 970 971 return Changed; 972} 973 974namespace { 975class AllocaPromoter : public LoadAndStorePromoter { 976 AllocaInst *AI; 977public: 978 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S) 979 : LoadAndStorePromoter(Insts, S), AI(0) {} 980 981 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 982 // Remember which alloca we're promoting (for isInstInList). 983 this->AI = AI; 984 LoadAndStorePromoter::run(Insts); 985 AI->eraseFromParent(); 986 } 987 988 virtual bool isInstInList(Instruction *I, 989 const SmallVectorImpl<Instruction*> &Insts) const { 990 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 991 return LI->getOperand(0) == AI; 992 return cast<StoreInst>(I)->getPointerOperand() == AI; 993 } 994}; 995} // end anon namespace 996 997/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 998/// subsequently loaded can be rewritten to load both input pointers and then 999/// select between the result, allowing the load of the alloca to be promoted. 1000/// From this: 1001/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1002/// %V = load i32* %P2 1003/// to: 1004/// %V1 = load i32* %Alloca -> will be mem2reg'd 1005/// %V2 = load i32* %Other 1006/// %V = select i1 %cond, i32 %V1, i32 %V2 1007/// 1008/// We can do this to a select if its only uses are loads and if the operand to 1009/// the select can be loaded unconditionally. 1010static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1011 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1012 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1013 1014 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1015 UI != UE; ++UI) { 1016 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1017 if (LI == 0 || LI->isVolatile()) return false; 1018 1019 // Both operands to the select need to be dereferencable, either absolutely 1020 // (e.g. allocas) or at this point because we can see other accesses to it. 1021 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1022 LI->getAlignment(), TD)) 1023 return false; 1024 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1025 LI->getAlignment(), TD)) 1026 return false; 1027 } 1028 1029 return true; 1030} 1031 1032/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1033/// subsequently loaded can be rewritten to load both input pointers in the pred 1034/// blocks and then PHI the results, allowing the load of the alloca to be 1035/// promoted. 1036/// From this: 1037/// %P2 = phi [i32* %Alloca, i32* %Other] 1038/// %V = load i32* %P2 1039/// to: 1040/// %V1 = load i32* %Alloca -> will be mem2reg'd 1041/// ... 1042/// %V2 = load i32* %Other 1043/// ... 1044/// %V = phi [i32 %V1, i32 %V2] 1045/// 1046/// We can do this to a select if its only uses are loads and if the operand to 1047/// the select can be loaded unconditionally. 1048static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1049 // For now, we can only do this promotion if the load is in the same block as 1050 // the PHI, and if there are no stores between the phi and load. 1051 // TODO: Allow recursive phi users. 1052 // TODO: Allow stores. 1053 BasicBlock *BB = PN->getParent(); 1054 unsigned MaxAlign = 0; 1055 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1056 UI != UE; ++UI) { 1057 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1058 if (LI == 0 || LI->isVolatile()) return false; 1059 1060 // For now we only allow loads in the same block as the PHI. This is a 1061 // common case that happens when instcombine merges two loads through a PHI. 1062 if (LI->getParent() != BB) return false; 1063 1064 // Ensure that there are no instructions between the PHI and the load that 1065 // could store. 1066 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1067 if (BBI->mayWriteToMemory()) 1068 return false; 1069 1070 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1071 } 1072 1073 // Okay, we know that we have one or more loads in the same block as the PHI. 1074 // We can transform this if it is safe to push the loads into the predecessor 1075 // blocks. The only thing to watch out for is that we can't put a possibly 1076 // trapping load in the predecessor if it is a critical edge. 1077 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1078 BasicBlock *Pred = PN->getIncomingBlock(i); 1079 1080 // If the predecessor has a single successor, then the edge isn't critical. 1081 if (Pred->getTerminator()->getNumSuccessors() == 1) 1082 continue; 1083 1084 Value *InVal = PN->getIncomingValue(i); 1085 1086 // If the InVal is an invoke in the pred, we can't put a load on the edge. 1087 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 1088 if (II->getParent() == Pred) 1089 return false; 1090 1091 // If this pointer is always safe to load, or if we can prove that there is 1092 // already a load in the block, then we can move the load to the pred block. 1093 if (InVal->isDereferenceablePointer() || 1094 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1095 continue; 1096 1097 return false; 1098 } 1099 1100 return true; 1101} 1102 1103 1104/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1105/// direct (non-volatile) loads and stores to it. If the alloca is close but 1106/// not quite there, this will transform the code to allow promotion. As such, 1107/// it is a non-pure predicate. 1108static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1109 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1110 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1111 1112 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1113 UI != UE; ++UI) { 1114 User *U = *UI; 1115 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1116 if (LI->isVolatile()) 1117 return false; 1118 continue; 1119 } 1120 1121 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1122 if (SI->getOperand(0) == AI || SI->isVolatile()) 1123 return false; // Don't allow a store OF the AI, only INTO the AI. 1124 continue; 1125 } 1126 1127 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1128 // If the condition being selected on is a constant, fold the select, yes 1129 // this does (rarely) happen early on. 1130 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1131 Value *Result = SI->getOperand(1+CI->isZero()); 1132 SI->replaceAllUsesWith(Result); 1133 SI->eraseFromParent(); 1134 1135 // This is very rare and we just scrambled the use list of AI, start 1136 // over completely. 1137 return tryToMakeAllocaBePromotable(AI, TD); 1138 } 1139 1140 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1141 // loads, then we can transform this by rewriting the select. 1142 if (!isSafeSelectToSpeculate(SI, TD)) 1143 return false; 1144 1145 InstsToRewrite.insert(SI); 1146 continue; 1147 } 1148 1149 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1150 if (PN->use_empty()) { // Dead PHIs can be stripped. 1151 InstsToRewrite.insert(PN); 1152 continue; 1153 } 1154 1155 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1156 // in the pred blocks, then we can transform this by rewriting the PHI. 1157 if (!isSafePHIToSpeculate(PN, TD)) 1158 return false; 1159 1160 InstsToRewrite.insert(PN); 1161 continue; 1162 } 1163 1164 return false; 1165 } 1166 1167 // If there are no instructions to rewrite, then all uses are load/stores and 1168 // we're done! 1169 if (InstsToRewrite.empty()) 1170 return true; 1171 1172 // If we have instructions that need to be rewritten for this to be promotable 1173 // take care of it now. 1174 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1175 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1176 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1177 // loads with a new select. 1178 while (!SI->use_empty()) { 1179 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1180 1181 IRBuilder<> Builder(LI); 1182 LoadInst *TrueLoad = 1183 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1184 LoadInst *FalseLoad = 1185 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".t"); 1186 1187 // Transfer alignment and TBAA info if present. 1188 TrueLoad->setAlignment(LI->getAlignment()); 1189 FalseLoad->setAlignment(LI->getAlignment()); 1190 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1191 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1192 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1193 } 1194 1195 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1196 V->takeName(LI); 1197 LI->replaceAllUsesWith(V); 1198 LI->eraseFromParent(); 1199 } 1200 1201 // Now that all the loads are gone, the select is gone too. 1202 SI->eraseFromParent(); 1203 continue; 1204 } 1205 1206 // Otherwise, we have a PHI node which allows us to push the loads into the 1207 // predecessors. 1208 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1209 if (PN->use_empty()) { 1210 PN->eraseFromParent(); 1211 continue; 1212 } 1213 1214 const Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1215 PHINode *NewPN = PHINode::Create(LoadTy, PN->getName()+".ld", PN); 1216 1217 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1218 // matter which one we get and if any differ, it doesn't matter. 1219 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1220 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1221 unsigned Align = SomeLoad->getAlignment(); 1222 1223 // Rewrite all loads of the PN to use the new PHI. 1224 while (!PN->use_empty()) { 1225 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1226 LI->replaceAllUsesWith(NewPN); 1227 LI->eraseFromParent(); 1228 } 1229 1230 // Inject loads into all of the pred blocks. Keep track of which blocks we 1231 // insert them into in case we have multiple edges from the same block. 1232 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1233 1234 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1235 BasicBlock *Pred = PN->getIncomingBlock(i); 1236 LoadInst *&Load = InsertedLoads[Pred]; 1237 if (Load == 0) { 1238 Load = new LoadInst(PN->getIncomingValue(i), 1239 PN->getName() + "." + Pred->getName(), 1240 Pred->getTerminator()); 1241 Load->setAlignment(Align); 1242 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1243 } 1244 1245 NewPN->addIncoming(Load, Pred); 1246 } 1247 1248 PN->eraseFromParent(); 1249 } 1250 1251 ++NumAdjusted; 1252 return true; 1253} 1254 1255 1256bool SROA::performPromotion(Function &F) { 1257 std::vector<AllocaInst*> Allocas; 1258 DominatorTree *DT = 0; 1259 if (HasDomTree) 1260 DT = &getAnalysis<DominatorTree>(); 1261 1262 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1263 1264 bool Changed = false; 1265 SmallVector<Instruction*, 64> Insts; 1266 while (1) { 1267 Allocas.clear(); 1268 1269 // Find allocas that are safe to promote, by looking at all instructions in 1270 // the entry node 1271 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1272 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1273 if (tryToMakeAllocaBePromotable(AI, TD)) 1274 Allocas.push_back(AI); 1275 1276 if (Allocas.empty()) break; 1277 1278 if (HasDomTree) 1279 PromoteMemToReg(Allocas, *DT); 1280 else { 1281 SSAUpdater SSA; 1282 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1283 AllocaInst *AI = Allocas[i]; 1284 1285 // Build list of instructions to promote. 1286 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1287 UI != E; ++UI) 1288 Insts.push_back(cast<Instruction>(*UI)); 1289 1290 AllocaPromoter(Insts, SSA).run(AI, Insts); 1291 Insts.clear(); 1292 } 1293 } 1294 NumPromoted += Allocas.size(); 1295 Changed = true; 1296 } 1297 1298 return Changed; 1299} 1300 1301 1302/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1303/// SROA. It must be a struct or array type with a small number of elements. 1304static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1305 const Type *T = AI->getAllocatedType(); 1306 // Do not promote any struct into more than 32 separate vars. 1307 if (const StructType *ST = dyn_cast<StructType>(T)) 1308 return ST->getNumElements() <= 32; 1309 // Arrays are much less likely to be safe for SROA; only consider 1310 // them if they are very small. 1311 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 1312 return AT->getNumElements() <= 8; 1313 return false; 1314} 1315 1316 1317// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1318// which runs on all of the malloc/alloca instructions in the function, removing 1319// them if they are only used by getelementptr instructions. 1320// 1321bool SROA::performScalarRepl(Function &F) { 1322 std::vector<AllocaInst*> WorkList; 1323 1324 // Scan the entry basic block, adding allocas to the worklist. 1325 BasicBlock &BB = F.getEntryBlock(); 1326 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1327 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1328 WorkList.push_back(A); 1329 1330 // Process the worklist 1331 bool Changed = false; 1332 while (!WorkList.empty()) { 1333 AllocaInst *AI = WorkList.back(); 1334 WorkList.pop_back(); 1335 1336 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1337 // with unused elements. 1338 if (AI->use_empty()) { 1339 AI->eraseFromParent(); 1340 Changed = true; 1341 continue; 1342 } 1343 1344 // If this alloca is impossible for us to promote, reject it early. 1345 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1346 continue; 1347 1348 // Check to see if this allocation is only modified by a memcpy/memmove from 1349 // a constant global. If this is the case, we can change all users to use 1350 // the constant global instead. This is commonly produced by the CFE by 1351 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1352 // is only subsequently read. 1353 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 1354 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1355 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 1356 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 1357 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1358 TheCopy->eraseFromParent(); // Don't mutate the global. 1359 AI->eraseFromParent(); 1360 ++NumGlobals; 1361 Changed = true; 1362 continue; 1363 } 1364 1365 // Check to see if we can perform the core SROA transformation. We cannot 1366 // transform the allocation instruction if it is an array allocation 1367 // (allocations OF arrays are ok though), and an allocation of a scalar 1368 // value cannot be decomposed at all. 1369 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1370 1371 // Do not promote [0 x %struct]. 1372 if (AllocaSize == 0) continue; 1373 1374 // Do not promote any struct whose size is too big. 1375 if (AllocaSize > SRThreshold) continue; 1376 1377 // If the alloca looks like a good candidate for scalar replacement, and if 1378 // all its users can be transformed, then split up the aggregate into its 1379 // separate elements. 1380 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1381 DoScalarReplacement(AI, WorkList); 1382 Changed = true; 1383 continue; 1384 } 1385 1386 // If we can turn this aggregate value (potentially with casts) into a 1387 // simple scalar value that can be mem2reg'd into a register value. 1388 // IsNotTrivial tracks whether this is something that mem2reg could have 1389 // promoted itself. If so, we don't want to transform it needlessly. Note 1390 // that we can't just check based on the type: the alloca may be of an i32 1391 // but that has pointer arithmetic to set byte 3 of it or something. 1392 if (AllocaInst *NewAI = 1393 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1394 NewAI->takeName(AI); 1395 AI->eraseFromParent(); 1396 ++NumConverted; 1397 Changed = true; 1398 continue; 1399 } 1400 1401 // Otherwise, couldn't process this alloca. 1402 } 1403 1404 return Changed; 1405} 1406 1407/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1408/// predicate, do SROA now. 1409void SROA::DoScalarReplacement(AllocaInst *AI, 1410 std::vector<AllocaInst*> &WorkList) { 1411 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1412 SmallVector<AllocaInst*, 32> ElementAllocas; 1413 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1414 ElementAllocas.reserve(ST->getNumContainedTypes()); 1415 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1416 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1417 AI->getAlignment(), 1418 AI->getName() + "." + Twine(i), AI); 1419 ElementAllocas.push_back(NA); 1420 WorkList.push_back(NA); // Add to worklist for recursive processing 1421 } 1422 } else { 1423 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1424 ElementAllocas.reserve(AT->getNumElements()); 1425 const Type *ElTy = AT->getElementType(); 1426 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1427 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1428 AI->getName() + "." + Twine(i), AI); 1429 ElementAllocas.push_back(NA); 1430 WorkList.push_back(NA); // Add to worklist for recursive processing 1431 } 1432 } 1433 1434 // Now that we have created the new alloca instructions, rewrite all the 1435 // uses of the old alloca. 1436 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1437 1438 // Now erase any instructions that were made dead while rewriting the alloca. 1439 DeleteDeadInstructions(); 1440 AI->eraseFromParent(); 1441 1442 ++NumReplaced; 1443} 1444 1445/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1446/// recursively including all their operands that become trivially dead. 1447void SROA::DeleteDeadInstructions() { 1448 while (!DeadInsts.empty()) { 1449 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1450 1451 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1452 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1453 // Zero out the operand and see if it becomes trivially dead. 1454 // (But, don't add allocas to the dead instruction list -- they are 1455 // already on the worklist and will be deleted separately.) 1456 *OI = 0; 1457 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1458 DeadInsts.push_back(U); 1459 } 1460 1461 I->eraseFromParent(); 1462 } 1463} 1464 1465/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1466/// performing scalar replacement of alloca AI. The results are flagged in 1467/// the Info parameter. Offset indicates the position within AI that is 1468/// referenced by this instruction. 1469void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1470 AllocaInfo &Info) { 1471 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1472 Instruction *User = cast<Instruction>(*UI); 1473 1474 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1475 isSafeForScalarRepl(BC, Offset, Info); 1476 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1477 uint64_t GEPOffset = Offset; 1478 isSafeGEP(GEPI, GEPOffset, Info); 1479 if (!Info.isUnsafe) 1480 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1481 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1482 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1483 if (Length == 0) 1484 return MarkUnsafe(Info, User); 1485 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1486 UI.getOperandNo() == 0, Info, MI, 1487 true /*AllowWholeAccess*/); 1488 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1489 if (LI->isVolatile()) 1490 return MarkUnsafe(Info, User); 1491 const Type *LIType = LI->getType(); 1492 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1493 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1494 Info.hasALoadOrStore = true; 1495 1496 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1497 // Store is ok if storing INTO the pointer, not storing the pointer 1498 if (SI->isVolatile() || SI->getOperand(0) == I) 1499 return MarkUnsafe(Info, User); 1500 1501 const Type *SIType = SI->getOperand(0)->getType(); 1502 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1503 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1504 Info.hasALoadOrStore = true; 1505 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1506 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1507 } else { 1508 return MarkUnsafe(Info, User); 1509 } 1510 if (Info.isUnsafe) return; 1511 } 1512} 1513 1514 1515/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1516/// derived from the alloca, we can often still split the alloca into elements. 1517/// This is useful if we have a large alloca where one element is phi'd 1518/// together somewhere: we can SRoA and promote all the other elements even if 1519/// we end up not being able to promote this one. 1520/// 1521/// All we require is that the uses of the PHI do not index into other parts of 1522/// the alloca. The most important use case for this is single load and stores 1523/// that are PHI'd together, which can happen due to code sinking. 1524void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1525 AllocaInfo &Info) { 1526 // If we've already checked this PHI, don't do it again. 1527 if (PHINode *PN = dyn_cast<PHINode>(I)) 1528 if (!Info.CheckedPHIs.insert(PN)) 1529 return; 1530 1531 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1532 Instruction *User = cast<Instruction>(*UI); 1533 1534 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1535 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1536 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1537 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1538 // but would have to prove that we're staying inside of an element being 1539 // promoted. 1540 if (!GEPI->hasAllZeroIndices()) 1541 return MarkUnsafe(Info, User); 1542 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1543 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1544 if (LI->isVolatile()) 1545 return MarkUnsafe(Info, User); 1546 const Type *LIType = LI->getType(); 1547 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1548 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1549 Info.hasALoadOrStore = true; 1550 1551 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1552 // Store is ok if storing INTO the pointer, not storing the pointer 1553 if (SI->isVolatile() || SI->getOperand(0) == I) 1554 return MarkUnsafe(Info, User); 1555 1556 const Type *SIType = SI->getOperand(0)->getType(); 1557 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1558 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1559 Info.hasALoadOrStore = true; 1560 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1561 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1562 } else { 1563 return MarkUnsafe(Info, User); 1564 } 1565 if (Info.isUnsafe) return; 1566 } 1567} 1568 1569/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1570/// replacement. It is safe when all the indices are constant, in-bounds 1571/// references, and when the resulting offset corresponds to an element within 1572/// the alloca type. The results are flagged in the Info parameter. Upon 1573/// return, Offset is adjusted as specified by the GEP indices. 1574void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1575 uint64_t &Offset, AllocaInfo &Info) { 1576 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1577 if (GEPIt == E) 1578 return; 1579 1580 // Walk through the GEP type indices, checking the types that this indexes 1581 // into. 1582 for (; GEPIt != E; ++GEPIt) { 1583 // Ignore struct elements, no extra checking needed for these. 1584 if ((*GEPIt)->isStructTy()) 1585 continue; 1586 1587 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1588 if (!IdxVal) 1589 return MarkUnsafe(Info, GEPI); 1590 } 1591 1592 // Compute the offset due to this GEP and check if the alloca has a 1593 // component element at that offset. 1594 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1595 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1596 &Indices[0], Indices.size()); 1597 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1598 MarkUnsafe(Info, GEPI); 1599} 1600 1601/// isHomogeneousAggregate - Check if type T is a struct or array containing 1602/// elements of the same type (which is always true for arrays). If so, 1603/// return true with NumElts and EltTy set to the number of elements and the 1604/// element type, respectively. 1605static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1606 const Type *&EltTy) { 1607 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1608 NumElts = AT->getNumElements(); 1609 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1610 return true; 1611 } 1612 if (const StructType *ST = dyn_cast<StructType>(T)) { 1613 NumElts = ST->getNumContainedTypes(); 1614 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1615 for (unsigned n = 1; n < NumElts; ++n) { 1616 if (ST->getContainedType(n) != EltTy) 1617 return false; 1618 } 1619 return true; 1620 } 1621 return false; 1622} 1623 1624/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1625/// "homogeneous" aggregates with the same element type and number of elements. 1626static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1627 if (T1 == T2) 1628 return true; 1629 1630 unsigned NumElts1, NumElts2; 1631 const Type *EltTy1, *EltTy2; 1632 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1633 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1634 NumElts1 == NumElts2 && 1635 EltTy1 == EltTy2) 1636 return true; 1637 1638 return false; 1639} 1640 1641/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1642/// alloca or has an offset and size that corresponds to a component element 1643/// within it. The offset checked here may have been formed from a GEP with a 1644/// pointer bitcasted to a different type. 1645/// 1646/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1647/// unit. If false, it only allows accesses known to be in a single element. 1648void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1649 const Type *MemOpType, bool isStore, 1650 AllocaInfo &Info, Instruction *TheAccess, 1651 bool AllowWholeAccess) { 1652 // Check if this is a load/store of the entire alloca. 1653 if (Offset == 0 && AllowWholeAccess && 1654 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1655 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1656 // loads/stores (which are essentially the same as the MemIntrinsics with 1657 // regard to copying padding between elements). But, if an alloca is 1658 // flagged as both a source and destination of such operations, we'll need 1659 // to check later for padding between elements. 1660 if (!MemOpType || MemOpType->isIntegerTy()) { 1661 if (isStore) 1662 Info.isMemCpyDst = true; 1663 else 1664 Info.isMemCpySrc = true; 1665 return; 1666 } 1667 // This is also safe for references using a type that is compatible with 1668 // the type of the alloca, so that loads/stores can be rewritten using 1669 // insertvalue/extractvalue. 1670 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1671 Info.hasSubelementAccess = true; 1672 return; 1673 } 1674 } 1675 // Check if the offset/size correspond to a component within the alloca type. 1676 const Type *T = Info.AI->getAllocatedType(); 1677 if (TypeHasComponent(T, Offset, MemSize)) { 1678 Info.hasSubelementAccess = true; 1679 return; 1680 } 1681 1682 return MarkUnsafe(Info, TheAccess); 1683} 1684 1685/// TypeHasComponent - Return true if T has a component type with the 1686/// specified offset and size. If Size is zero, do not check the size. 1687bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1688 const Type *EltTy; 1689 uint64_t EltSize; 1690 if (const StructType *ST = dyn_cast<StructType>(T)) { 1691 const StructLayout *Layout = TD->getStructLayout(ST); 1692 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1693 EltTy = ST->getContainedType(EltIdx); 1694 EltSize = TD->getTypeAllocSize(EltTy); 1695 Offset -= Layout->getElementOffset(EltIdx); 1696 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1697 EltTy = AT->getElementType(); 1698 EltSize = TD->getTypeAllocSize(EltTy); 1699 if (Offset >= AT->getNumElements() * EltSize) 1700 return false; 1701 Offset %= EltSize; 1702 } else { 1703 return false; 1704 } 1705 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1706 return true; 1707 // Check if the component spans multiple elements. 1708 if (Offset + Size > EltSize) 1709 return false; 1710 return TypeHasComponent(EltTy, Offset, Size); 1711} 1712 1713/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1714/// the instruction I, which references it, to use the separate elements. 1715/// Offset indicates the position within AI that is referenced by this 1716/// instruction. 1717void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1718 SmallVector<AllocaInst*, 32> &NewElts) { 1719 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1720 Use &TheUse = UI.getUse(); 1721 Instruction *User = cast<Instruction>(*UI++); 1722 1723 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1724 RewriteBitCast(BC, AI, Offset, NewElts); 1725 continue; 1726 } 1727 1728 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1729 RewriteGEP(GEPI, AI, Offset, NewElts); 1730 continue; 1731 } 1732 1733 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1734 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1735 uint64_t MemSize = Length->getZExtValue(); 1736 if (Offset == 0 && 1737 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1738 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1739 // Otherwise the intrinsic can only touch a single element and the 1740 // address operand will be updated, so nothing else needs to be done. 1741 continue; 1742 } 1743 1744 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1745 const Type *LIType = LI->getType(); 1746 1747 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1748 // Replace: 1749 // %res = load { i32, i32 }* %alloc 1750 // with: 1751 // %load.0 = load i32* %alloc.0 1752 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1753 // %load.1 = load i32* %alloc.1 1754 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1755 // (Also works for arrays instead of structs) 1756 Value *Insert = UndefValue::get(LIType); 1757 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1758 Value *Load = new LoadInst(NewElts[i], "load", LI); 1759 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1760 } 1761 LI->replaceAllUsesWith(Insert); 1762 DeadInsts.push_back(LI); 1763 } else if (LIType->isIntegerTy() && 1764 TD->getTypeAllocSize(LIType) == 1765 TD->getTypeAllocSize(AI->getAllocatedType())) { 1766 // If this is a load of the entire alloca to an integer, rewrite it. 1767 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1768 } 1769 continue; 1770 } 1771 1772 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1773 Value *Val = SI->getOperand(0); 1774 const Type *SIType = Val->getType(); 1775 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1776 // Replace: 1777 // store { i32, i32 } %val, { i32, i32 }* %alloc 1778 // with: 1779 // %val.0 = extractvalue { i32, i32 } %val, 0 1780 // store i32 %val.0, i32* %alloc.0 1781 // %val.1 = extractvalue { i32, i32 } %val, 1 1782 // store i32 %val.1, i32* %alloc.1 1783 // (Also works for arrays instead of structs) 1784 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1785 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1786 new StoreInst(Extract, NewElts[i], SI); 1787 } 1788 DeadInsts.push_back(SI); 1789 } else if (SIType->isIntegerTy() && 1790 TD->getTypeAllocSize(SIType) == 1791 TD->getTypeAllocSize(AI->getAllocatedType())) { 1792 // If this is a store of the entire alloca from an integer, rewrite it. 1793 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1794 } 1795 continue; 1796 } 1797 1798 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1799 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1800 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1801 // the new pointer. 1802 if (!isa<AllocaInst>(I)) continue; 1803 1804 assert(Offset == 0 && NewElts[0] && 1805 "Direct alloca use should have a zero offset"); 1806 1807 // If we have a use of the alloca, we know the derived uses will be 1808 // utilizing just the first element of the scalarized result. Insert a 1809 // bitcast of the first alloca before the user as required. 1810 AllocaInst *NewAI = NewElts[0]; 1811 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1812 NewAI->moveBefore(BCI); 1813 TheUse = BCI; 1814 continue; 1815 } 1816 } 1817} 1818 1819/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1820/// and recursively continue updating all of its uses. 1821void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1822 SmallVector<AllocaInst*, 32> &NewElts) { 1823 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1824 if (BC->getOperand(0) != AI) 1825 return; 1826 1827 // The bitcast references the original alloca. Replace its uses with 1828 // references to the first new element alloca. 1829 Instruction *Val = NewElts[0]; 1830 if (Val->getType() != BC->getDestTy()) { 1831 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1832 Val->takeName(BC); 1833 } 1834 BC->replaceAllUsesWith(Val); 1835 DeadInsts.push_back(BC); 1836} 1837 1838/// FindElementAndOffset - Return the index of the element containing Offset 1839/// within the specified type, which must be either a struct or an array. 1840/// Sets T to the type of the element and Offset to the offset within that 1841/// element. IdxTy is set to the type of the index result to be used in a 1842/// GEP instruction. 1843uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1844 const Type *&IdxTy) { 1845 uint64_t Idx = 0; 1846 if (const StructType *ST = dyn_cast<StructType>(T)) { 1847 const StructLayout *Layout = TD->getStructLayout(ST); 1848 Idx = Layout->getElementContainingOffset(Offset); 1849 T = ST->getContainedType(Idx); 1850 Offset -= Layout->getElementOffset(Idx); 1851 IdxTy = Type::getInt32Ty(T->getContext()); 1852 return Idx; 1853 } 1854 const ArrayType *AT = cast<ArrayType>(T); 1855 T = AT->getElementType(); 1856 uint64_t EltSize = TD->getTypeAllocSize(T); 1857 Idx = Offset / EltSize; 1858 Offset -= Idx * EltSize; 1859 IdxTy = Type::getInt64Ty(T->getContext()); 1860 return Idx; 1861} 1862 1863/// RewriteGEP - Check if this GEP instruction moves the pointer across 1864/// elements of the alloca that are being split apart, and if so, rewrite 1865/// the GEP to be relative to the new element. 1866void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1867 SmallVector<AllocaInst*, 32> &NewElts) { 1868 uint64_t OldOffset = Offset; 1869 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1870 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1871 &Indices[0], Indices.size()); 1872 1873 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1874 1875 const Type *T = AI->getAllocatedType(); 1876 const Type *IdxTy; 1877 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1878 if (GEPI->getOperand(0) == AI) 1879 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1880 1881 T = AI->getAllocatedType(); 1882 uint64_t EltOffset = Offset; 1883 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1884 1885 // If this GEP does not move the pointer across elements of the alloca 1886 // being split, then it does not needs to be rewritten. 1887 if (Idx == OldIdx) 1888 return; 1889 1890 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1891 SmallVector<Value*, 8> NewArgs; 1892 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1893 while (EltOffset != 0) { 1894 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1895 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1896 } 1897 Instruction *Val = NewElts[Idx]; 1898 if (NewArgs.size() > 1) { 1899 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1900 NewArgs.end(), "", GEPI); 1901 Val->takeName(GEPI); 1902 } 1903 if (Val->getType() != GEPI->getType()) 1904 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1905 GEPI->replaceAllUsesWith(Val); 1906 DeadInsts.push_back(GEPI); 1907} 1908 1909/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1910/// Rewrite it to copy or set the elements of the scalarized memory. 1911void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1912 AllocaInst *AI, 1913 SmallVector<AllocaInst*, 32> &NewElts) { 1914 // If this is a memcpy/memmove, construct the other pointer as the 1915 // appropriate type. The "Other" pointer is the pointer that goes to memory 1916 // that doesn't have anything to do with the alloca that we are promoting. For 1917 // memset, this Value* stays null. 1918 Value *OtherPtr = 0; 1919 unsigned MemAlignment = MI->getAlignment(); 1920 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1921 if (Inst == MTI->getRawDest()) 1922 OtherPtr = MTI->getRawSource(); 1923 else { 1924 assert(Inst == MTI->getRawSource()); 1925 OtherPtr = MTI->getRawDest(); 1926 } 1927 } 1928 1929 // If there is an other pointer, we want to convert it to the same pointer 1930 // type as AI has, so we can GEP through it safely. 1931 if (OtherPtr) { 1932 unsigned AddrSpace = 1933 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 1934 1935 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1936 // optimization, but it's also required to detect the corner case where 1937 // both pointer operands are referencing the same memory, and where 1938 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1939 // function is only called for mem intrinsics that access the whole 1940 // aggregate, so non-zero GEPs are not an issue here.) 1941 OtherPtr = OtherPtr->stripPointerCasts(); 1942 1943 // Copying the alloca to itself is a no-op: just delete it. 1944 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1945 // This code will run twice for a no-op memcpy -- once for each operand. 1946 // Put only one reference to MI on the DeadInsts list. 1947 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1948 E = DeadInsts.end(); I != E; ++I) 1949 if (*I == MI) return; 1950 DeadInsts.push_back(MI); 1951 return; 1952 } 1953 1954 // If the pointer is not the right type, insert a bitcast to the right 1955 // type. 1956 const Type *NewTy = 1957 PointerType::get(AI->getType()->getElementType(), AddrSpace); 1958 1959 if (OtherPtr->getType() != NewTy) 1960 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 1961 } 1962 1963 // Process each element of the aggregate. 1964 bool SROADest = MI->getRawDest() == Inst; 1965 1966 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1967 1968 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1969 // If this is a memcpy/memmove, emit a GEP of the other element address. 1970 Value *OtherElt = 0; 1971 unsigned OtherEltAlign = MemAlignment; 1972 1973 if (OtherPtr) { 1974 Value *Idx[2] = { Zero, 1975 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1976 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1977 OtherPtr->getName()+"."+Twine(i), 1978 MI); 1979 uint64_t EltOffset; 1980 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1981 const Type *OtherTy = OtherPtrTy->getElementType(); 1982 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1983 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1984 } else { 1985 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 1986 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1987 } 1988 1989 // The alignment of the other pointer is the guaranteed alignment of the 1990 // element, which is affected by both the known alignment of the whole 1991 // mem intrinsic and the alignment of the element. If the alignment of 1992 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 1993 // known alignment is just 4 bytes. 1994 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 1995 } 1996 1997 Value *EltPtr = NewElts[i]; 1998 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 1999 2000 // If we got down to a scalar, insert a load or store as appropriate. 2001 if (EltTy->isSingleValueType()) { 2002 if (isa<MemTransferInst>(MI)) { 2003 if (SROADest) { 2004 // From Other to Alloca. 2005 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2006 new StoreInst(Elt, EltPtr, MI); 2007 } else { 2008 // From Alloca to Other. 2009 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2010 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2011 } 2012 continue; 2013 } 2014 assert(isa<MemSetInst>(MI)); 2015 2016 // If the stored element is zero (common case), just store a null 2017 // constant. 2018 Constant *StoreVal; 2019 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2020 if (CI->isZero()) { 2021 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2022 } else { 2023 // If EltTy is a vector type, get the element type. 2024 const Type *ValTy = EltTy->getScalarType(); 2025 2026 // Construct an integer with the right value. 2027 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2028 APInt OneVal(EltSize, CI->getZExtValue()); 2029 APInt TotalVal(OneVal); 2030 // Set each byte. 2031 for (unsigned i = 0; 8*i < EltSize; ++i) { 2032 TotalVal = TotalVal.shl(8); 2033 TotalVal |= OneVal; 2034 } 2035 2036 // Convert the integer value to the appropriate type. 2037 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2038 if (ValTy->isPointerTy()) 2039 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2040 else if (ValTy->isFloatingPointTy()) 2041 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2042 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2043 2044 // If the requested value was a vector constant, create it. 2045 if (EltTy != ValTy) { 2046 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 2047 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2048 StoreVal = ConstantVector::get(Elts); 2049 } 2050 } 2051 new StoreInst(StoreVal, EltPtr, MI); 2052 continue; 2053 } 2054 // Otherwise, if we're storing a byte variable, use a memset call for 2055 // this element. 2056 } 2057 2058 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2059 2060 IRBuilder<> Builder(MI); 2061 2062 // Finally, insert the meminst for this element. 2063 if (isa<MemSetInst>(MI)) { 2064 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2065 MI->isVolatile()); 2066 } else { 2067 assert(isa<MemTransferInst>(MI)); 2068 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2069 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2070 2071 if (isa<MemCpyInst>(MI)) 2072 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2073 else 2074 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2075 } 2076 } 2077 DeadInsts.push_back(MI); 2078} 2079 2080/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2081/// overwrites the entire allocation. Extract out the pieces of the stored 2082/// integer and store them individually. 2083void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2084 SmallVector<AllocaInst*, 32> &NewElts){ 2085 // Extract each element out of the integer according to its structure offset 2086 // and store the element value to the individual alloca. 2087 Value *SrcVal = SI->getOperand(0); 2088 const Type *AllocaEltTy = AI->getAllocatedType(); 2089 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2090 2091 IRBuilder<> Builder(SI); 2092 2093 // Handle tail padding by extending the operand 2094 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2095 SrcVal = Builder.CreateZExt(SrcVal, 2096 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2097 2098 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2099 << '\n'); 2100 2101 // There are two forms here: AI could be an array or struct. Both cases 2102 // have different ways to compute the element offset. 2103 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2104 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2105 2106 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2107 // Get the number of bits to shift SrcVal to get the value. 2108 const Type *FieldTy = EltSTy->getElementType(i); 2109 uint64_t Shift = Layout->getElementOffsetInBits(i); 2110 2111 if (TD->isBigEndian()) 2112 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2113 2114 Value *EltVal = SrcVal; 2115 if (Shift) { 2116 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2117 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2118 } 2119 2120 // Truncate down to an integer of the right size. 2121 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2122 2123 // Ignore zero sized fields like {}, they obviously contain no data. 2124 if (FieldSizeBits == 0) continue; 2125 2126 if (FieldSizeBits != AllocaSizeBits) 2127 EltVal = Builder.CreateTrunc(EltVal, 2128 IntegerType::get(SI->getContext(), FieldSizeBits)); 2129 Value *DestField = NewElts[i]; 2130 if (EltVal->getType() == FieldTy) { 2131 // Storing to an integer field of this size, just do it. 2132 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2133 // Bitcast to the right element type (for fp/vector values). 2134 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2135 } else { 2136 // Otherwise, bitcast the dest pointer (for aggregates). 2137 DestField = Builder.CreateBitCast(DestField, 2138 PointerType::getUnqual(EltVal->getType())); 2139 } 2140 new StoreInst(EltVal, DestField, SI); 2141 } 2142 2143 } else { 2144 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2145 const Type *ArrayEltTy = ATy->getElementType(); 2146 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2147 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2148 2149 uint64_t Shift; 2150 2151 if (TD->isBigEndian()) 2152 Shift = AllocaSizeBits-ElementOffset; 2153 else 2154 Shift = 0; 2155 2156 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2157 // Ignore zero sized fields like {}, they obviously contain no data. 2158 if (ElementSizeBits == 0) continue; 2159 2160 Value *EltVal = SrcVal; 2161 if (Shift) { 2162 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2163 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2164 } 2165 2166 // Truncate down to an integer of the right size. 2167 if (ElementSizeBits != AllocaSizeBits) 2168 EltVal = Builder.CreateTrunc(EltVal, 2169 IntegerType::get(SI->getContext(), 2170 ElementSizeBits)); 2171 Value *DestField = NewElts[i]; 2172 if (EltVal->getType() == ArrayEltTy) { 2173 // Storing to an integer field of this size, just do it. 2174 } else if (ArrayEltTy->isFloatingPointTy() || 2175 ArrayEltTy->isVectorTy()) { 2176 // Bitcast to the right element type (for fp/vector values). 2177 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2178 } else { 2179 // Otherwise, bitcast the dest pointer (for aggregates). 2180 DestField = Builder.CreateBitCast(DestField, 2181 PointerType::getUnqual(EltVal->getType())); 2182 } 2183 new StoreInst(EltVal, DestField, SI); 2184 2185 if (TD->isBigEndian()) 2186 Shift -= ElementOffset; 2187 else 2188 Shift += ElementOffset; 2189 } 2190 } 2191 2192 DeadInsts.push_back(SI); 2193} 2194 2195/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2196/// an integer. Load the individual pieces to form the aggregate value. 2197void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2198 SmallVector<AllocaInst*, 32> &NewElts) { 2199 // Extract each element out of the NewElts according to its structure offset 2200 // and form the result value. 2201 const Type *AllocaEltTy = AI->getAllocatedType(); 2202 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2203 2204 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2205 << '\n'); 2206 2207 // There are two forms here: AI could be an array or struct. Both cases 2208 // have different ways to compute the element offset. 2209 const StructLayout *Layout = 0; 2210 uint64_t ArrayEltBitOffset = 0; 2211 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2212 Layout = TD->getStructLayout(EltSTy); 2213 } else { 2214 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2215 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2216 } 2217 2218 Value *ResultVal = 2219 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2220 2221 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2222 // Load the value from the alloca. If the NewElt is an aggregate, cast 2223 // the pointer to an integer of the same size before doing the load. 2224 Value *SrcField = NewElts[i]; 2225 const Type *FieldTy = 2226 cast<PointerType>(SrcField->getType())->getElementType(); 2227 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2228 2229 // Ignore zero sized fields like {}, they obviously contain no data. 2230 if (FieldSizeBits == 0) continue; 2231 2232 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2233 FieldSizeBits); 2234 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2235 !FieldTy->isVectorTy()) 2236 SrcField = new BitCastInst(SrcField, 2237 PointerType::getUnqual(FieldIntTy), 2238 "", LI); 2239 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2240 2241 // If SrcField is a fp or vector of the right size but that isn't an 2242 // integer type, bitcast to an integer so we can shift it. 2243 if (SrcField->getType() != FieldIntTy) 2244 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2245 2246 // Zero extend the field to be the same size as the final alloca so that 2247 // we can shift and insert it. 2248 if (SrcField->getType() != ResultVal->getType()) 2249 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2250 2251 // Determine the number of bits to shift SrcField. 2252 uint64_t Shift; 2253 if (Layout) // Struct case. 2254 Shift = Layout->getElementOffsetInBits(i); 2255 else // Array case. 2256 Shift = i*ArrayEltBitOffset; 2257 2258 if (TD->isBigEndian()) 2259 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2260 2261 if (Shift) { 2262 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2263 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2264 } 2265 2266 // Don't create an 'or x, 0' on the first iteration. 2267 if (!isa<Constant>(ResultVal) || 2268 !cast<Constant>(ResultVal)->isNullValue()) 2269 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2270 else 2271 ResultVal = SrcField; 2272 } 2273 2274 // Handle tail padding by truncating the result 2275 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2276 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2277 2278 LI->replaceAllUsesWith(ResultVal); 2279 DeadInsts.push_back(LI); 2280} 2281 2282/// HasPadding - Return true if the specified type has any structure or 2283/// alignment padding in between the elements that would be split apart 2284/// by SROA; return false otherwise. 2285static bool HasPadding(const Type *Ty, const TargetData &TD) { 2286 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2287 Ty = ATy->getElementType(); 2288 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2289 } 2290 2291 // SROA currently handles only Arrays and Structs. 2292 const StructType *STy = cast<StructType>(Ty); 2293 const StructLayout *SL = TD.getStructLayout(STy); 2294 unsigned PrevFieldBitOffset = 0; 2295 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2296 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2297 2298 // Check to see if there is any padding between this element and the 2299 // previous one. 2300 if (i) { 2301 unsigned PrevFieldEnd = 2302 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2303 if (PrevFieldEnd < FieldBitOffset) 2304 return true; 2305 } 2306 PrevFieldBitOffset = FieldBitOffset; 2307 } 2308 // Check for tail padding. 2309 if (unsigned EltCount = STy->getNumElements()) { 2310 unsigned PrevFieldEnd = PrevFieldBitOffset + 2311 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2312 if (PrevFieldEnd < SL->getSizeInBits()) 2313 return true; 2314 } 2315 return false; 2316} 2317 2318/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2319/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2320/// or 1 if safe after canonicalization has been performed. 2321bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2322 // Loop over the use list of the alloca. We can only transform it if all of 2323 // the users are safe to transform. 2324 AllocaInfo Info(AI); 2325 2326 isSafeForScalarRepl(AI, 0, Info); 2327 if (Info.isUnsafe) { 2328 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2329 return false; 2330 } 2331 2332 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2333 // source and destination, we have to be careful. In particular, the memcpy 2334 // could be moving around elements that live in structure padding of the LLVM 2335 // types, but may actually be used. In these cases, we refuse to promote the 2336 // struct. 2337 if (Info.isMemCpySrc && Info.isMemCpyDst && 2338 HasPadding(AI->getAllocatedType(), *TD)) 2339 return false; 2340 2341 // If the alloca never has an access to just *part* of it, but is accessed 2342 // via loads and stores, then we should use ConvertToScalarInfo to promote 2343 // the alloca instead of promoting each piece at a time and inserting fission 2344 // and fusion code. 2345 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2346 // If the struct/array just has one element, use basic SRoA. 2347 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2348 if (ST->getNumElements() > 1) return false; 2349 } else { 2350 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2351 return false; 2352 } 2353 } 2354 2355 return true; 2356} 2357 2358 2359 2360/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2361/// some part of a constant global variable. This intentionally only accepts 2362/// constant expressions because we don't can't rewrite arbitrary instructions. 2363static bool PointsToConstantGlobal(Value *V) { 2364 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2365 return GV->isConstant(); 2366 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2367 if (CE->getOpcode() == Instruction::BitCast || 2368 CE->getOpcode() == Instruction::GetElementPtr) 2369 return PointsToConstantGlobal(CE->getOperand(0)); 2370 return false; 2371} 2372 2373/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2374/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2375/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2376/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2377/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2378/// the alloca, and if the source pointer is a pointer to a constant global, we 2379/// can optimize this. 2380static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2381 bool isOffset) { 2382 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2383 User *U = cast<Instruction>(*UI); 2384 2385 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2386 // Ignore non-volatile loads, they are always ok. 2387 if (LI->isVolatile()) return false; 2388 continue; 2389 } 2390 2391 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2392 // If uses of the bitcast are ok, we are ok. 2393 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 2394 return false; 2395 continue; 2396 } 2397 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2398 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2399 // doesn't, it does. 2400 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2401 isOffset || !GEP->hasAllZeroIndices())) 2402 return false; 2403 continue; 2404 } 2405 2406 if (CallSite CS = U) { 2407 // If this is a readonly/readnone call site, then we know it is just a 2408 // load and we can ignore it. 2409 if (CS.onlyReadsMemory()) 2410 continue; 2411 2412 // If this is the function being called then we treat it like a load and 2413 // ignore it. 2414 if (CS.isCallee(UI)) 2415 continue; 2416 2417 // If this is being passed as a byval argument, the caller is making a 2418 // copy, so it is only a read of the alloca. 2419 unsigned ArgNo = CS.getArgumentNo(UI); 2420 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2421 continue; 2422 } 2423 2424 // If this is isn't our memcpy/memmove, reject it as something we can't 2425 // handle. 2426 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2427 if (MI == 0) 2428 return false; 2429 2430 // If the transfer is using the alloca as a source of the transfer, then 2431 // ignore it since it is a load (unless the transfer is volatile). 2432 if (UI.getOperandNo() == 1) { 2433 if (MI->isVolatile()) return false; 2434 continue; 2435 } 2436 2437 // If we already have seen a copy, reject the second one. 2438 if (TheCopy) return false; 2439 2440 // If the pointer has been offset from the start of the alloca, we can't 2441 // safely handle this. 2442 if (isOffset) return false; 2443 2444 // If the memintrinsic isn't using the alloca as the dest, reject it. 2445 if (UI.getOperandNo() != 0) return false; 2446 2447 // If the source of the memcpy/move is not a constant global, reject it. 2448 if (!PointsToConstantGlobal(MI->getSource())) 2449 return false; 2450 2451 // Otherwise, the transform is safe. Remember the copy instruction. 2452 TheCopy = MI; 2453 } 2454 return true; 2455} 2456 2457/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2458/// modified by a copy from a constant global. If we can prove this, we can 2459/// replace any uses of the alloca with uses of the global directly. 2460MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 2461 MemTransferInst *TheCopy = 0; 2462 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 2463 return TheCopy; 2464 return 0; 2465} 2466