ScalarReplAggregates.cpp revision 85a7c690852d6151acff0d8821762d75bc774ab4
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/ValueTracking.h" 35#include "llvm/Target/TargetData.h" 36#include "llvm/Transforms/Utils/PromoteMemToReg.h" 37#include "llvm/Transforms/Utils/Local.h" 38#include "llvm/Transforms/Utils/SSAUpdater.h" 39#include "llvm/Support/CallSite.h" 40#include "llvm/Support/Debug.h" 41#include "llvm/Support/ErrorHandling.h" 42#include "llvm/Support/GetElementPtrTypeIterator.h" 43#include "llvm/Support/IRBuilder.h" 44#include "llvm/Support/MathExtras.h" 45#include "llvm/Support/raw_ostream.h" 46#include "llvm/ADT/SmallVector.h" 47#include "llvm/ADT/Statistic.h" 48using namespace llvm; 49 50STATISTIC(NumReplaced, "Number of allocas broken up"); 51STATISTIC(NumPromoted, "Number of allocas promoted"); 52STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 53STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 54 55namespace { 56 struct SROA : public FunctionPass { 57 SROA(int T, bool hasDT, char &ID) 58 : FunctionPass(ID), HasDomTree(hasDT) { 59 if (T == -1) 60 SRThreshold = 128; 61 else 62 SRThreshold = T; 63 } 64 65 bool runOnFunction(Function &F); 66 67 bool performScalarRepl(Function &F); 68 bool performPromotion(Function &F); 69 70 private: 71 bool HasDomTree; 72 TargetData *TD; 73 74 /// DeadInsts - Keep track of instructions we have made dead, so that 75 /// we can remove them after we are done working. 76 SmallVector<Value*, 32> DeadInsts; 77 78 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 79 /// information about the uses. All these fields are initialized to false 80 /// and set to true when something is learned. 81 struct AllocaInfo { 82 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 83 bool isUnsafe : 1; 84 85 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 86 bool isMemCpySrc : 1; 87 88 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 89 bool isMemCpyDst : 1; 90 91 /// hasSubelementAccess - This is true if a subelement of the alloca is 92 /// ever accessed, or false if the alloca is only accessed with mem 93 /// intrinsics or load/store that only access the entire alloca at once. 94 bool hasSubelementAccess : 1; 95 96 /// hasALoadOrStore - This is true if there are any loads or stores to it. 97 /// The alloca may just be accessed with memcpy, for example, which would 98 /// not set this. 99 bool hasALoadOrStore : 1; 100 101 AllocaInfo() 102 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 103 hasSubelementAccess(false), hasALoadOrStore(false) {} 104 }; 105 106 unsigned SRThreshold; 107 108 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 109 110 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 111 112 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 113 AllocaInfo &Info); 114 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset, 115 AllocaInfo &Info); 116 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 117 const Type *MemOpType, bool isStore, AllocaInfo &Info); 118 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 119 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 120 const Type *&IdxTy); 121 122 void DoScalarReplacement(AllocaInst *AI, 123 std::vector<AllocaInst*> &WorkList); 124 void DeleteDeadInstructions(); 125 126 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 127 SmallVector<AllocaInst*, 32> &NewElts); 128 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 129 SmallVector<AllocaInst*, 32> &NewElts); 130 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 131 SmallVector<AllocaInst*, 32> &NewElts); 132 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 133 AllocaInst *AI, 134 SmallVector<AllocaInst*, 32> &NewElts); 135 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 136 SmallVector<AllocaInst*, 32> &NewElts); 137 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 138 SmallVector<AllocaInst*, 32> &NewElts); 139 140 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 141 }; 142 143 // SROA_DT - SROA that uses DominatorTree. 144 struct SROA_DT : public SROA { 145 static char ID; 146 public: 147 SROA_DT(int T = -1) : SROA(T, true, ID) { 148 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 149 } 150 151 // getAnalysisUsage - This pass does not require any passes, but we know it 152 // will not alter the CFG, so say so. 153 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 154 AU.addRequired<DominatorTree>(); 155 AU.setPreservesCFG(); 156 } 157 }; 158 159 // SROA_SSAUp - SROA that uses SSAUpdater. 160 struct SROA_SSAUp : public SROA { 161 static char ID; 162 public: 163 SROA_SSAUp(int T = -1) : SROA(T, false, ID) { 164 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 165 } 166 167 // getAnalysisUsage - This pass does not require any passes, but we know it 168 // will not alter the CFG, so say so. 169 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 170 AU.setPreservesCFG(); 171 } 172 }; 173 174} 175 176char SROA_DT::ID = 0; 177char SROA_SSAUp::ID = 0; 178 179INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 180 "Scalar Replacement of Aggregates (DT)", false, false) 181INITIALIZE_PASS_DEPENDENCY(DominatorTree) 182INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 183 "Scalar Replacement of Aggregates (DT)", false, false) 184 185INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 186 "Scalar Replacement of Aggregates (SSAUp)", false, false) 187INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 188 "Scalar Replacement of Aggregates (SSAUp)", false, false) 189 190// Public interface to the ScalarReplAggregates pass 191FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 192 bool UseDomTree) { 193 if (UseDomTree) 194 return new SROA_DT(Threshold); 195 return new SROA_SSAUp(Threshold); 196} 197 198 199//===----------------------------------------------------------------------===// 200// Convert To Scalar Optimization. 201//===----------------------------------------------------------------------===// 202 203namespace { 204/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 205/// optimization, which scans the uses of an alloca and determines if it can 206/// rewrite it in terms of a single new alloca that can be mem2reg'd. 207class ConvertToScalarInfo { 208 /// AllocaSize - The size of the alloca being considered. 209 unsigned AllocaSize; 210 const TargetData &TD; 211 212 /// IsNotTrivial - This is set to true if there is some access to the object 213 /// which means that mem2reg can't promote it. 214 bool IsNotTrivial; 215 216 /// VectorTy - This tracks the type that we should promote the vector to if 217 /// it is possible to turn it into a vector. This starts out null, and if it 218 /// isn't possible to turn into a vector type, it gets set to VoidTy. 219 const Type *VectorTy; 220 221 /// HadAVector - True if there is at least one vector access to the alloca. 222 /// We don't want to turn random arrays into vectors and use vector element 223 /// insert/extract, but if there are element accesses to something that is 224 /// also declared as a vector, we do want to promote to a vector. 225 bool HadAVector; 226 227public: 228 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 229 : AllocaSize(Size), TD(td) { 230 IsNotTrivial = false; 231 VectorTy = 0; 232 HadAVector = false; 233 } 234 235 AllocaInst *TryConvert(AllocaInst *AI); 236 237private: 238 bool CanConvertToScalar(Value *V, uint64_t Offset); 239 void MergeInType(const Type *In, uint64_t Offset); 240 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 241 242 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 243 uint64_t Offset, IRBuilder<> &Builder); 244 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 245 uint64_t Offset, IRBuilder<> &Builder); 246}; 247} // end anonymous namespace. 248 249 250/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 251/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 252/// alloca if possible or null if not. 253AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 254 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 255 // out. 256 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 257 return 0; 258 259 // If we were able to find a vector type that can handle this with 260 // insert/extract elements, and if there was at least one use that had 261 // a vector type, promote this to a vector. We don't want to promote 262 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 263 // we just get a lot of insert/extracts. If at least one vector is 264 // involved, then we probably really do have a union of vector/array. 265 const Type *NewTy; 266 if (VectorTy && VectorTy->isVectorTy() && HadAVector) { 267 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 268 << *VectorTy << '\n'); 269 NewTy = VectorTy; // Use the vector type. 270 } else { 271 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 272 // Create and insert the integer alloca. 273 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 274 } 275 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 276 ConvertUsesToScalar(AI, NewAI, 0); 277 return NewAI; 278} 279 280/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy) 281/// so far at the offset specified by Offset (which is specified in bytes). 282/// 283/// There are two cases we handle here: 284/// 1) A union of vector types of the same size and potentially its elements. 285/// Here we turn element accesses into insert/extract element operations. 286/// This promotes a <4 x float> with a store of float to the third element 287/// into a <4 x float> that uses insert element. 288/// 2) A fully general blob of memory, which we turn into some (potentially 289/// large) integer type with extract and insert operations where the loads 290/// and stores would mutate the memory. We mark this by setting VectorTy 291/// to VoidTy. 292void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) { 293 // If we already decided to turn this into a blob of integer memory, there is 294 // nothing to be done. 295 if (VectorTy && VectorTy->isVoidTy()) 296 return; 297 298 // If this could be contributing to a vector, analyze it. 299 300 // If the In type is a vector that is the same size as the alloca, see if it 301 // matches the existing VecTy. 302 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 303 // Remember if we saw a vector type. 304 HadAVector = true; 305 306 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 307 // If we're storing/loading a vector of the right size, allow it as a 308 // vector. If this the first vector we see, remember the type so that 309 // we know the element size. If this is a subsequent access, ignore it 310 // even if it is a differing type but the same size. Worst case we can 311 // bitcast the resultant vectors. 312 if (VectorTy == 0) 313 VectorTy = VInTy; 314 return; 315 } 316 } else if (In->isFloatTy() || In->isDoubleTy() || 317 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 318 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 319 // If we're accessing something that could be an element of a vector, see 320 // if the implied vector agrees with what we already have and if Offset is 321 // compatible with it. 322 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 323 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 324 (VectorTy == 0 || 325 cast<VectorType>(VectorTy)->getElementType() 326 ->getPrimitiveSizeInBits()/8 == EltSize)) { 327 if (VectorTy == 0) 328 VectorTy = VectorType::get(In, AllocaSize/EltSize); 329 return; 330 } 331 } 332 333 // Otherwise, we have a case that we can't handle with an optimized vector 334 // form. We can still turn this into a large integer. 335 VectorTy = Type::getVoidTy(In->getContext()); 336} 337 338/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 339/// its accesses to a single vector type, return true and set VecTy to 340/// the new type. If we could convert the alloca into a single promotable 341/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 342/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 343/// is the current offset from the base of the alloca being analyzed. 344/// 345/// If we see at least one access to the value that is as a vector type, set the 346/// SawVec flag. 347bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 348 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 349 Instruction *User = cast<Instruction>(*UI); 350 351 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 352 // Don't break volatile loads. 353 if (LI->isVolatile()) 354 return false; 355 // Don't touch MMX operations. 356 if (LI->getType()->isX86_MMXTy()) 357 return false; 358 MergeInType(LI->getType(), Offset); 359 continue; 360 } 361 362 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 363 // Storing the pointer, not into the value? 364 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 365 // Don't touch MMX operations. 366 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 367 return false; 368 MergeInType(SI->getOperand(0)->getType(), Offset); 369 continue; 370 } 371 372 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 373 IsNotTrivial = true; // Can't be mem2reg'd. 374 if (!CanConvertToScalar(BCI, Offset)) 375 return false; 376 continue; 377 } 378 379 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 380 // If this is a GEP with a variable indices, we can't handle it. 381 if (!GEP->hasAllConstantIndices()) 382 return false; 383 384 // Compute the offset that this GEP adds to the pointer. 385 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 386 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 387 &Indices[0], Indices.size()); 388 // See if all uses can be converted. 389 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 390 return false; 391 IsNotTrivial = true; // Can't be mem2reg'd. 392 continue; 393 } 394 395 // If this is a constant sized memset of a constant value (e.g. 0) we can 396 // handle it. 397 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 398 // Store of constant value and constant size. 399 if (!isa<ConstantInt>(MSI->getValue()) || 400 !isa<ConstantInt>(MSI->getLength())) 401 return false; 402 IsNotTrivial = true; // Can't be mem2reg'd. 403 continue; 404 } 405 406 // If this is a memcpy or memmove into or out of the whole allocation, we 407 // can handle it like a load or store of the scalar type. 408 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 409 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 410 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 411 return false; 412 413 IsNotTrivial = true; // Can't be mem2reg'd. 414 continue; 415 } 416 417 // Otherwise, we cannot handle this! 418 return false; 419 } 420 421 return true; 422} 423 424/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 425/// directly. This happens when we are converting an "integer union" to a 426/// single integer scalar, or when we are converting a "vector union" to a 427/// vector with insert/extractelement instructions. 428/// 429/// Offset is an offset from the original alloca, in bits that need to be 430/// shifted to the right. By the end of this, there should be no uses of Ptr. 431void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 432 uint64_t Offset) { 433 while (!Ptr->use_empty()) { 434 Instruction *User = cast<Instruction>(Ptr->use_back()); 435 436 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 437 ConvertUsesToScalar(CI, NewAI, Offset); 438 CI->eraseFromParent(); 439 continue; 440 } 441 442 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 443 // Compute the offset that this GEP adds to the pointer. 444 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 445 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 446 &Indices[0], Indices.size()); 447 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 448 GEP->eraseFromParent(); 449 continue; 450 } 451 452 IRBuilder<> Builder(User); 453 454 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 455 // The load is a bit extract from NewAI shifted right by Offset bits. 456 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 457 Value *NewLoadVal 458 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 459 LI->replaceAllUsesWith(NewLoadVal); 460 LI->eraseFromParent(); 461 continue; 462 } 463 464 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 465 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 466 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 467 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 468 Builder); 469 Builder.CreateStore(New, NewAI); 470 SI->eraseFromParent(); 471 472 // If the load we just inserted is now dead, then the inserted store 473 // overwrote the entire thing. 474 if (Old->use_empty()) 475 Old->eraseFromParent(); 476 continue; 477 } 478 479 // If this is a constant sized memset of a constant value (e.g. 0) we can 480 // transform it into a store of the expanded constant value. 481 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 482 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 483 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 484 if (NumBytes != 0) { 485 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 486 487 // Compute the value replicated the right number of times. 488 APInt APVal(NumBytes*8, Val); 489 490 // Splat the value if non-zero. 491 if (Val) 492 for (unsigned i = 1; i != NumBytes; ++i) 493 APVal |= APVal << 8; 494 495 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 496 Value *New = ConvertScalar_InsertValue( 497 ConstantInt::get(User->getContext(), APVal), 498 Old, Offset, Builder); 499 Builder.CreateStore(New, NewAI); 500 501 // If the load we just inserted is now dead, then the memset overwrote 502 // the entire thing. 503 if (Old->use_empty()) 504 Old->eraseFromParent(); 505 } 506 MSI->eraseFromParent(); 507 continue; 508 } 509 510 // If this is a memcpy or memmove into or out of the whole allocation, we 511 // can handle it like a load or store of the scalar type. 512 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 513 assert(Offset == 0 && "must be store to start of alloca"); 514 515 // If the source and destination are both to the same alloca, then this is 516 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 517 // as appropriate. 518 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, 0)); 519 520 if (GetUnderlyingObject(MTI->getSource(), 0) != OrigAI) { 521 // Dest must be OrigAI, change this to be a load from the original 522 // pointer (bitcasted), then a store to our new alloca. 523 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 524 Value *SrcPtr = MTI->getSource(); 525 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 526 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 527 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 528 AIPTy = PointerType::get(AIPTy->getElementType(), 529 SPTy->getAddressSpace()); 530 } 531 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 532 533 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 534 SrcVal->setAlignment(MTI->getAlignment()); 535 Builder.CreateStore(SrcVal, NewAI); 536 } else if (GetUnderlyingObject(MTI->getDest(), 0) != OrigAI) { 537 // Src must be OrigAI, change this to be a load from NewAI then a store 538 // through the original dest pointer (bitcasted). 539 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 540 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 541 542 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 543 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 544 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 545 AIPTy = PointerType::get(AIPTy->getElementType(), 546 DPTy->getAddressSpace()); 547 } 548 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 549 550 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 551 NewStore->setAlignment(MTI->getAlignment()); 552 } else { 553 // Noop transfer. Src == Dst 554 } 555 556 MTI->eraseFromParent(); 557 continue; 558 } 559 560 llvm_unreachable("Unsupported operation!"); 561 } 562} 563 564/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 565/// or vector value FromVal, extracting the bits from the offset specified by 566/// Offset. This returns the value, which is of type ToType. 567/// 568/// This happens when we are converting an "integer union" to a single 569/// integer scalar, or when we are converting a "vector union" to a vector with 570/// insert/extractelement instructions. 571/// 572/// Offset is an offset from the original alloca, in bits that need to be 573/// shifted to the right. 574Value *ConvertToScalarInfo:: 575ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 576 uint64_t Offset, IRBuilder<> &Builder) { 577 // If the load is of the whole new alloca, no conversion is needed. 578 if (FromVal->getType() == ToType && Offset == 0) 579 return FromVal; 580 581 // If the result alloca is a vector type, this is either an element 582 // access or a bitcast to another vector type of the same size. 583 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 584 if (ToType->isVectorTy()) 585 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 586 587 // Otherwise it must be an element access. 588 unsigned Elt = 0; 589 if (Offset) { 590 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 591 Elt = Offset/EltSize; 592 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 593 } 594 // Return the element extracted out of it. 595 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 596 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 597 if (V->getType() != ToType) 598 V = Builder.CreateBitCast(V, ToType, "tmp"); 599 return V; 600 } 601 602 // If ToType is a first class aggregate, extract out each of the pieces and 603 // use insertvalue's to form the FCA. 604 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 605 const StructLayout &Layout = *TD.getStructLayout(ST); 606 Value *Res = UndefValue::get(ST); 607 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 608 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 609 Offset+Layout.getElementOffsetInBits(i), 610 Builder); 611 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 612 } 613 return Res; 614 } 615 616 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 617 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 618 Value *Res = UndefValue::get(AT); 619 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 620 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 621 Offset+i*EltSize, Builder); 622 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 623 } 624 return Res; 625 } 626 627 // Otherwise, this must be a union that was converted to an integer value. 628 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 629 630 // If this is a big-endian system and the load is narrower than the 631 // full alloca type, we need to do a shift to get the right bits. 632 int ShAmt = 0; 633 if (TD.isBigEndian()) { 634 // On big-endian machines, the lowest bit is stored at the bit offset 635 // from the pointer given by getTypeStoreSizeInBits. This matters for 636 // integers with a bitwidth that is not a multiple of 8. 637 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 638 TD.getTypeStoreSizeInBits(ToType) - Offset; 639 } else { 640 ShAmt = Offset; 641 } 642 643 // Note: we support negative bitwidths (with shl) which are not defined. 644 // We do this to support (f.e.) loads off the end of a structure where 645 // only some bits are used. 646 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 647 FromVal = Builder.CreateLShr(FromVal, 648 ConstantInt::get(FromVal->getType(), 649 ShAmt), "tmp"); 650 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 651 FromVal = Builder.CreateShl(FromVal, 652 ConstantInt::get(FromVal->getType(), 653 -ShAmt), "tmp"); 654 655 // Finally, unconditionally truncate the integer to the right width. 656 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 657 if (LIBitWidth < NTy->getBitWidth()) 658 FromVal = 659 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 660 LIBitWidth), "tmp"); 661 else if (LIBitWidth > NTy->getBitWidth()) 662 FromVal = 663 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 664 LIBitWidth), "tmp"); 665 666 // If the result is an integer, this is a trunc or bitcast. 667 if (ToType->isIntegerTy()) { 668 // Should be done. 669 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 670 // Just do a bitcast, we know the sizes match up. 671 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 672 } else { 673 // Otherwise must be a pointer. 674 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 675 } 676 assert(FromVal->getType() == ToType && "Didn't convert right?"); 677 return FromVal; 678} 679 680/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 681/// or vector value "Old" at the offset specified by Offset. 682/// 683/// This happens when we are converting an "integer union" to a 684/// single integer scalar, or when we are converting a "vector union" to a 685/// vector with insert/extractelement instructions. 686/// 687/// Offset is an offset from the original alloca, in bits that need to be 688/// shifted to the right. 689Value *ConvertToScalarInfo:: 690ConvertScalar_InsertValue(Value *SV, Value *Old, 691 uint64_t Offset, IRBuilder<> &Builder) { 692 // Convert the stored type to the actual type, shift it left to insert 693 // then 'or' into place. 694 const Type *AllocaType = Old->getType(); 695 LLVMContext &Context = Old->getContext(); 696 697 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 698 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 699 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 700 701 // Changing the whole vector with memset or with an access of a different 702 // vector type? 703 if (ValSize == VecSize) 704 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 705 706 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 707 708 // Must be an element insertion. 709 unsigned Elt = Offset/EltSize; 710 711 if (SV->getType() != VTy->getElementType()) 712 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 713 714 SV = Builder.CreateInsertElement(Old, SV, 715 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 716 "tmp"); 717 return SV; 718 } 719 720 // If SV is a first-class aggregate value, insert each value recursively. 721 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 722 const StructLayout &Layout = *TD.getStructLayout(ST); 723 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 724 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 725 Old = ConvertScalar_InsertValue(Elt, Old, 726 Offset+Layout.getElementOffsetInBits(i), 727 Builder); 728 } 729 return Old; 730 } 731 732 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 733 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 734 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 735 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 736 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 737 } 738 return Old; 739 } 740 741 // If SV is a float, convert it to the appropriate integer type. 742 // If it is a pointer, do the same. 743 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 744 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 745 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 746 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 747 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 748 SV = Builder.CreateBitCast(SV, 749 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 750 else if (SV->getType()->isPointerTy()) 751 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 752 753 // Zero extend or truncate the value if needed. 754 if (SV->getType() != AllocaType) { 755 if (SV->getType()->getPrimitiveSizeInBits() < 756 AllocaType->getPrimitiveSizeInBits()) 757 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 758 else { 759 // Truncation may be needed if storing more than the alloca can hold 760 // (undefined behavior). 761 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 762 SrcWidth = DestWidth; 763 SrcStoreWidth = DestStoreWidth; 764 } 765 } 766 767 // If this is a big-endian system and the store is narrower than the 768 // full alloca type, we need to do a shift to get the right bits. 769 int ShAmt = 0; 770 if (TD.isBigEndian()) { 771 // On big-endian machines, the lowest bit is stored at the bit offset 772 // from the pointer given by getTypeStoreSizeInBits. This matters for 773 // integers with a bitwidth that is not a multiple of 8. 774 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 775 } else { 776 ShAmt = Offset; 777 } 778 779 // Note: we support negative bitwidths (with shr) which are not defined. 780 // We do this to support (f.e.) stores off the end of a structure where 781 // only some bits in the structure are set. 782 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 783 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 784 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 785 ShAmt), "tmp"); 786 Mask <<= ShAmt; 787 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 788 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 789 -ShAmt), "tmp"); 790 Mask = Mask.lshr(-ShAmt); 791 } 792 793 // Mask out the bits we are about to insert from the old value, and or 794 // in the new bits. 795 if (SrcWidth != DestWidth) { 796 assert(DestWidth > SrcWidth); 797 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 798 SV = Builder.CreateOr(Old, SV, "ins"); 799 } 800 return SV; 801} 802 803 804//===----------------------------------------------------------------------===// 805// SRoA Driver 806//===----------------------------------------------------------------------===// 807 808 809bool SROA::runOnFunction(Function &F) { 810 TD = getAnalysisIfAvailable<TargetData>(); 811 812 bool Changed = performPromotion(F); 813 814 // FIXME: ScalarRepl currently depends on TargetData more than it 815 // theoretically needs to. It should be refactored in order to support 816 // target-independent IR. Until this is done, just skip the actual 817 // scalar-replacement portion of this pass. 818 if (!TD) return Changed; 819 820 while (1) { 821 bool LocalChange = performScalarRepl(F); 822 if (!LocalChange) break; // No need to repromote if no scalarrepl 823 Changed = true; 824 LocalChange = performPromotion(F); 825 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 826 } 827 828 return Changed; 829} 830 831namespace { 832class AllocaPromoter : public LoadAndStorePromoter { 833 AllocaInst *AI; 834public: 835 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S) 836 : LoadAndStorePromoter(Insts, S), AI(0) {} 837 838 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 839 // Remember which alloca we're promoting (for isInstInList). 840 this->AI = AI; 841 LoadAndStorePromoter::run(Insts); 842 AI->eraseFromParent(); 843 } 844 845 virtual bool isInstInList(Instruction *I, 846 const SmallVectorImpl<Instruction*> &Insts) const { 847 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 848 return LI->getOperand(0) == AI; 849 return cast<StoreInst>(I)->getPointerOperand() == AI; 850 } 851}; 852} // end anon namespace 853 854bool SROA::performPromotion(Function &F) { 855 std::vector<AllocaInst*> Allocas; 856 DominatorTree *DT = 0; 857 if (HasDomTree) 858 DT = &getAnalysis<DominatorTree>(); 859 860 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 861 862 bool Changed = false; 863 SmallVector<Instruction*, 64> Insts; 864 while (1) { 865 Allocas.clear(); 866 867 // Find allocas that are safe to promote, by looking at all instructions in 868 // the entry node 869 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 870 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 871 if (isAllocaPromotable(AI)) 872 Allocas.push_back(AI); 873 874 if (Allocas.empty()) break; 875 876 if (HasDomTree) 877 PromoteMemToReg(Allocas, *DT); 878 else { 879 SSAUpdater SSA; 880 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 881 AllocaInst *AI = Allocas[i]; 882 883 // Build list of instructions to promote. 884 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 885 UI != E; ++UI) 886 Insts.push_back(cast<Instruction>(*UI)); 887 888 AllocaPromoter(Insts, SSA).run(AI, Insts); 889 Insts.clear(); 890 } 891 } 892 NumPromoted += Allocas.size(); 893 Changed = true; 894 } 895 896 return Changed; 897} 898 899 900/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 901/// SROA. It must be a struct or array type with a small number of elements. 902static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 903 const Type *T = AI->getAllocatedType(); 904 // Do not promote any struct into more than 32 separate vars. 905 if (const StructType *ST = dyn_cast<StructType>(T)) 906 return ST->getNumElements() <= 32; 907 // Arrays are much less likely to be safe for SROA; only consider 908 // them if they are very small. 909 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 910 return AT->getNumElements() <= 8; 911 return false; 912} 913 914 915// performScalarRepl - This algorithm is a simple worklist driven algorithm, 916// which runs on all of the malloc/alloca instructions in the function, removing 917// them if they are only used by getelementptr instructions. 918// 919bool SROA::performScalarRepl(Function &F) { 920 std::vector<AllocaInst*> WorkList; 921 922 // Scan the entry basic block, adding allocas to the worklist. 923 BasicBlock &BB = F.getEntryBlock(); 924 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 925 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 926 WorkList.push_back(A); 927 928 // Process the worklist 929 bool Changed = false; 930 while (!WorkList.empty()) { 931 AllocaInst *AI = WorkList.back(); 932 WorkList.pop_back(); 933 934 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 935 // with unused elements. 936 if (AI->use_empty()) { 937 AI->eraseFromParent(); 938 Changed = true; 939 continue; 940 } 941 942 // If this alloca is impossible for us to promote, reject it early. 943 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 944 continue; 945 946 // Check to see if this allocation is only modified by a memcpy/memmove from 947 // a constant global. If this is the case, we can change all users to use 948 // the constant global instead. This is commonly produced by the CFE by 949 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 950 // is only subsequently read. 951 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 952 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 953 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 954 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 955 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 956 TheCopy->eraseFromParent(); // Don't mutate the global. 957 AI->eraseFromParent(); 958 ++NumGlobals; 959 Changed = true; 960 continue; 961 } 962 963 // Check to see if we can perform the core SROA transformation. We cannot 964 // transform the allocation instruction if it is an array allocation 965 // (allocations OF arrays are ok though), and an allocation of a scalar 966 // value cannot be decomposed at all. 967 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 968 969 // Do not promote [0 x %struct]. 970 if (AllocaSize == 0) continue; 971 972 // Do not promote any struct whose size is too big. 973 if (AllocaSize > SRThreshold) continue; 974 975 // If the alloca looks like a good candidate for scalar replacement, and if 976 // all its users can be transformed, then split up the aggregate into its 977 // separate elements. 978 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 979 DoScalarReplacement(AI, WorkList); 980 Changed = true; 981 continue; 982 } 983 984 // If we can turn this aggregate value (potentially with casts) into a 985 // simple scalar value that can be mem2reg'd into a register value. 986 // IsNotTrivial tracks whether this is something that mem2reg could have 987 // promoted itself. If so, we don't want to transform it needlessly. Note 988 // that we can't just check based on the type: the alloca may be of an i32 989 // but that has pointer arithmetic to set byte 3 of it or something. 990 if (AllocaInst *NewAI = 991 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 992 NewAI->takeName(AI); 993 AI->eraseFromParent(); 994 ++NumConverted; 995 Changed = true; 996 continue; 997 } 998 999 // Otherwise, couldn't process this alloca. 1000 } 1001 1002 return Changed; 1003} 1004 1005/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1006/// predicate, do SROA now. 1007void SROA::DoScalarReplacement(AllocaInst *AI, 1008 std::vector<AllocaInst*> &WorkList) { 1009 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1010 SmallVector<AllocaInst*, 32> ElementAllocas; 1011 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1012 ElementAllocas.reserve(ST->getNumContainedTypes()); 1013 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1014 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1015 AI->getAlignment(), 1016 AI->getName() + "." + Twine(i), AI); 1017 ElementAllocas.push_back(NA); 1018 WorkList.push_back(NA); // Add to worklist for recursive processing 1019 } 1020 } else { 1021 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1022 ElementAllocas.reserve(AT->getNumElements()); 1023 const Type *ElTy = AT->getElementType(); 1024 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1025 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1026 AI->getName() + "." + Twine(i), AI); 1027 ElementAllocas.push_back(NA); 1028 WorkList.push_back(NA); // Add to worklist for recursive processing 1029 } 1030 } 1031 1032 // Now that we have created the new alloca instructions, rewrite all the 1033 // uses of the old alloca. 1034 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1035 1036 // Now erase any instructions that were made dead while rewriting the alloca. 1037 DeleteDeadInstructions(); 1038 AI->eraseFromParent(); 1039 1040 ++NumReplaced; 1041} 1042 1043/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1044/// recursively including all their operands that become trivially dead. 1045void SROA::DeleteDeadInstructions() { 1046 while (!DeadInsts.empty()) { 1047 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1048 1049 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1050 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1051 // Zero out the operand and see if it becomes trivially dead. 1052 // (But, don't add allocas to the dead instruction list -- they are 1053 // already on the worklist and will be deleted separately.) 1054 *OI = 0; 1055 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1056 DeadInsts.push_back(U); 1057 } 1058 1059 I->eraseFromParent(); 1060 } 1061} 1062 1063/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1064/// performing scalar replacement of alloca AI. The results are flagged in 1065/// the Info parameter. Offset indicates the position within AI that is 1066/// referenced by this instruction. 1067void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1068 AllocaInfo &Info) { 1069 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1070 Instruction *User = cast<Instruction>(*UI); 1071 1072 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1073 isSafeForScalarRepl(BC, AI, Offset, Info); 1074 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1075 uint64_t GEPOffset = Offset; 1076 isSafeGEP(GEPI, AI, GEPOffset, Info); 1077 if (!Info.isUnsafe) 1078 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 1079 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1080 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1081 if (Length) 1082 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 1083 UI.getOperandNo() == 0, Info); 1084 else 1085 MarkUnsafe(Info); 1086 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1087 if (!LI->isVolatile()) { 1088 const Type *LIType = LI->getType(); 1089 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 1090 LIType, false, Info); 1091 Info.hasALoadOrStore = true; 1092 } else 1093 MarkUnsafe(Info); 1094 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1095 // Store is ok if storing INTO the pointer, not storing the pointer 1096 if (!SI->isVolatile() && SI->getOperand(0) != I) { 1097 const Type *SIType = SI->getOperand(0)->getType(); 1098 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 1099 SIType, true, Info); 1100 Info.hasALoadOrStore = true; 1101 } else 1102 MarkUnsafe(Info); 1103 } else { 1104 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 1105 MarkUnsafe(Info); 1106 } 1107 if (Info.isUnsafe) return; 1108 } 1109} 1110 1111/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1112/// replacement. It is safe when all the indices are constant, in-bounds 1113/// references, and when the resulting offset corresponds to an element within 1114/// the alloca type. The results are flagged in the Info parameter. Upon 1115/// return, Offset is adjusted as specified by the GEP indices. 1116void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 1117 uint64_t &Offset, AllocaInfo &Info) { 1118 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1119 if (GEPIt == E) 1120 return; 1121 1122 // Walk through the GEP type indices, checking the types that this indexes 1123 // into. 1124 for (; GEPIt != E; ++GEPIt) { 1125 // Ignore struct elements, no extra checking needed for these. 1126 if ((*GEPIt)->isStructTy()) 1127 continue; 1128 1129 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1130 if (!IdxVal) 1131 return MarkUnsafe(Info); 1132 } 1133 1134 // Compute the offset due to this GEP and check if the alloca has a 1135 // component element at that offset. 1136 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1137 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1138 &Indices[0], Indices.size()); 1139 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 1140 MarkUnsafe(Info); 1141} 1142 1143/// isHomogeneousAggregate - Check if type T is a struct or array containing 1144/// elements of the same type (which is always true for arrays). If so, 1145/// return true with NumElts and EltTy set to the number of elements and the 1146/// element type, respectively. 1147static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1148 const Type *&EltTy) { 1149 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1150 NumElts = AT->getNumElements(); 1151 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1152 return true; 1153 } 1154 if (const StructType *ST = dyn_cast<StructType>(T)) { 1155 NumElts = ST->getNumContainedTypes(); 1156 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1157 for (unsigned n = 1; n < NumElts; ++n) { 1158 if (ST->getContainedType(n) != EltTy) 1159 return false; 1160 } 1161 return true; 1162 } 1163 return false; 1164} 1165 1166/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1167/// "homogeneous" aggregates with the same element type and number of elements. 1168static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1169 if (T1 == T2) 1170 return true; 1171 1172 unsigned NumElts1, NumElts2; 1173 const Type *EltTy1, *EltTy2; 1174 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1175 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1176 NumElts1 == NumElts2 && 1177 EltTy1 == EltTy2) 1178 return true; 1179 1180 return false; 1181} 1182 1183/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1184/// alloca or has an offset and size that corresponds to a component element 1185/// within it. The offset checked here may have been formed from a GEP with a 1186/// pointer bitcasted to a different type. 1187void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 1188 const Type *MemOpType, bool isStore, 1189 AllocaInfo &Info) { 1190 // Check if this is a load/store of the entire alloca. 1191 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 1192 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1193 // loads/stores (which are essentially the same as the MemIntrinsics with 1194 // regard to copying padding between elements). But, if an alloca is 1195 // flagged as both a source and destination of such operations, we'll need 1196 // to check later for padding between elements. 1197 if (!MemOpType || MemOpType->isIntegerTy()) { 1198 if (isStore) 1199 Info.isMemCpyDst = true; 1200 else 1201 Info.isMemCpySrc = true; 1202 return; 1203 } 1204 // This is also safe for references using a type that is compatible with 1205 // the type of the alloca, so that loads/stores can be rewritten using 1206 // insertvalue/extractvalue. 1207 if (isCompatibleAggregate(MemOpType, AI->getAllocatedType())) { 1208 Info.hasSubelementAccess = true; 1209 return; 1210 } 1211 } 1212 // Check if the offset/size correspond to a component within the alloca type. 1213 const Type *T = AI->getAllocatedType(); 1214 if (TypeHasComponent(T, Offset, MemSize)) { 1215 Info.hasSubelementAccess = true; 1216 return; 1217 } 1218 1219 return MarkUnsafe(Info); 1220} 1221 1222/// TypeHasComponent - Return true if T has a component type with the 1223/// specified offset and size. If Size is zero, do not check the size. 1224bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1225 const Type *EltTy; 1226 uint64_t EltSize; 1227 if (const StructType *ST = dyn_cast<StructType>(T)) { 1228 const StructLayout *Layout = TD->getStructLayout(ST); 1229 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1230 EltTy = ST->getContainedType(EltIdx); 1231 EltSize = TD->getTypeAllocSize(EltTy); 1232 Offset -= Layout->getElementOffset(EltIdx); 1233 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1234 EltTy = AT->getElementType(); 1235 EltSize = TD->getTypeAllocSize(EltTy); 1236 if (Offset >= AT->getNumElements() * EltSize) 1237 return false; 1238 Offset %= EltSize; 1239 } else { 1240 return false; 1241 } 1242 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1243 return true; 1244 // Check if the component spans multiple elements. 1245 if (Offset + Size > EltSize) 1246 return false; 1247 return TypeHasComponent(EltTy, Offset, Size); 1248} 1249 1250/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1251/// the instruction I, which references it, to use the separate elements. 1252/// Offset indicates the position within AI that is referenced by this 1253/// instruction. 1254void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1255 SmallVector<AllocaInst*, 32> &NewElts) { 1256 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1257 Instruction *User = cast<Instruction>(*UI); 1258 1259 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1260 RewriteBitCast(BC, AI, Offset, NewElts); 1261 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1262 RewriteGEP(GEPI, AI, Offset, NewElts); 1263 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1264 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1265 uint64_t MemSize = Length->getZExtValue(); 1266 if (Offset == 0 && 1267 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1268 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1269 // Otherwise the intrinsic can only touch a single element and the 1270 // address operand will be updated, so nothing else needs to be done. 1271 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1272 const Type *LIType = LI->getType(); 1273 1274 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1275 // Replace: 1276 // %res = load { i32, i32 }* %alloc 1277 // with: 1278 // %load.0 = load i32* %alloc.0 1279 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1280 // %load.1 = load i32* %alloc.1 1281 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1282 // (Also works for arrays instead of structs) 1283 Value *Insert = UndefValue::get(LIType); 1284 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1285 Value *Load = new LoadInst(NewElts[i], "load", LI); 1286 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1287 } 1288 LI->replaceAllUsesWith(Insert); 1289 DeadInsts.push_back(LI); 1290 } else if (LIType->isIntegerTy() && 1291 TD->getTypeAllocSize(LIType) == 1292 TD->getTypeAllocSize(AI->getAllocatedType())) { 1293 // If this is a load of the entire alloca to an integer, rewrite it. 1294 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1295 } 1296 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1297 Value *Val = SI->getOperand(0); 1298 const Type *SIType = Val->getType(); 1299 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1300 // Replace: 1301 // store { i32, i32 } %val, { i32, i32 }* %alloc 1302 // with: 1303 // %val.0 = extractvalue { i32, i32 } %val, 0 1304 // store i32 %val.0, i32* %alloc.0 1305 // %val.1 = extractvalue { i32, i32 } %val, 1 1306 // store i32 %val.1, i32* %alloc.1 1307 // (Also works for arrays instead of structs) 1308 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1309 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1310 new StoreInst(Extract, NewElts[i], SI); 1311 } 1312 DeadInsts.push_back(SI); 1313 } else if (SIType->isIntegerTy() && 1314 TD->getTypeAllocSize(SIType) == 1315 TD->getTypeAllocSize(AI->getAllocatedType())) { 1316 // If this is a store of the entire alloca from an integer, rewrite it. 1317 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1318 } 1319 } 1320 } 1321} 1322 1323/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1324/// and recursively continue updating all of its uses. 1325void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1326 SmallVector<AllocaInst*, 32> &NewElts) { 1327 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1328 if (BC->getOperand(0) != AI) 1329 return; 1330 1331 // The bitcast references the original alloca. Replace its uses with 1332 // references to the first new element alloca. 1333 Instruction *Val = NewElts[0]; 1334 if (Val->getType() != BC->getDestTy()) { 1335 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1336 Val->takeName(BC); 1337 } 1338 BC->replaceAllUsesWith(Val); 1339 DeadInsts.push_back(BC); 1340} 1341 1342/// FindElementAndOffset - Return the index of the element containing Offset 1343/// within the specified type, which must be either a struct or an array. 1344/// Sets T to the type of the element and Offset to the offset within that 1345/// element. IdxTy is set to the type of the index result to be used in a 1346/// GEP instruction. 1347uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1348 const Type *&IdxTy) { 1349 uint64_t Idx = 0; 1350 if (const StructType *ST = dyn_cast<StructType>(T)) { 1351 const StructLayout *Layout = TD->getStructLayout(ST); 1352 Idx = Layout->getElementContainingOffset(Offset); 1353 T = ST->getContainedType(Idx); 1354 Offset -= Layout->getElementOffset(Idx); 1355 IdxTy = Type::getInt32Ty(T->getContext()); 1356 return Idx; 1357 } 1358 const ArrayType *AT = cast<ArrayType>(T); 1359 T = AT->getElementType(); 1360 uint64_t EltSize = TD->getTypeAllocSize(T); 1361 Idx = Offset / EltSize; 1362 Offset -= Idx * EltSize; 1363 IdxTy = Type::getInt64Ty(T->getContext()); 1364 return Idx; 1365} 1366 1367/// RewriteGEP - Check if this GEP instruction moves the pointer across 1368/// elements of the alloca that are being split apart, and if so, rewrite 1369/// the GEP to be relative to the new element. 1370void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1371 SmallVector<AllocaInst*, 32> &NewElts) { 1372 uint64_t OldOffset = Offset; 1373 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1374 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1375 &Indices[0], Indices.size()); 1376 1377 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1378 1379 const Type *T = AI->getAllocatedType(); 1380 const Type *IdxTy; 1381 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1382 if (GEPI->getOperand(0) == AI) 1383 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1384 1385 T = AI->getAllocatedType(); 1386 uint64_t EltOffset = Offset; 1387 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1388 1389 // If this GEP does not move the pointer across elements of the alloca 1390 // being split, then it does not needs to be rewritten. 1391 if (Idx == OldIdx) 1392 return; 1393 1394 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1395 SmallVector<Value*, 8> NewArgs; 1396 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1397 while (EltOffset != 0) { 1398 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1399 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1400 } 1401 Instruction *Val = NewElts[Idx]; 1402 if (NewArgs.size() > 1) { 1403 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1404 NewArgs.end(), "", GEPI); 1405 Val->takeName(GEPI); 1406 } 1407 if (Val->getType() != GEPI->getType()) 1408 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1409 GEPI->replaceAllUsesWith(Val); 1410 DeadInsts.push_back(GEPI); 1411} 1412 1413/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1414/// Rewrite it to copy or set the elements of the scalarized memory. 1415void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1416 AllocaInst *AI, 1417 SmallVector<AllocaInst*, 32> &NewElts) { 1418 // If this is a memcpy/memmove, construct the other pointer as the 1419 // appropriate type. The "Other" pointer is the pointer that goes to memory 1420 // that doesn't have anything to do with the alloca that we are promoting. For 1421 // memset, this Value* stays null. 1422 Value *OtherPtr = 0; 1423 unsigned MemAlignment = MI->getAlignment(); 1424 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1425 if (Inst == MTI->getRawDest()) 1426 OtherPtr = MTI->getRawSource(); 1427 else { 1428 assert(Inst == MTI->getRawSource()); 1429 OtherPtr = MTI->getRawDest(); 1430 } 1431 } 1432 1433 // If there is an other pointer, we want to convert it to the same pointer 1434 // type as AI has, so we can GEP through it safely. 1435 if (OtherPtr) { 1436 unsigned AddrSpace = 1437 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 1438 1439 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1440 // optimization, but it's also required to detect the corner case where 1441 // both pointer operands are referencing the same memory, and where 1442 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1443 // function is only called for mem intrinsics that access the whole 1444 // aggregate, so non-zero GEPs are not an issue here.) 1445 OtherPtr = OtherPtr->stripPointerCasts(); 1446 1447 // Copying the alloca to itself is a no-op: just delete it. 1448 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1449 // This code will run twice for a no-op memcpy -- once for each operand. 1450 // Put only one reference to MI on the DeadInsts list. 1451 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1452 E = DeadInsts.end(); I != E; ++I) 1453 if (*I == MI) return; 1454 DeadInsts.push_back(MI); 1455 return; 1456 } 1457 1458 // If the pointer is not the right type, insert a bitcast to the right 1459 // type. 1460 const Type *NewTy = 1461 PointerType::get(AI->getType()->getElementType(), AddrSpace); 1462 1463 if (OtherPtr->getType() != NewTy) 1464 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 1465 } 1466 1467 // Process each element of the aggregate. 1468 bool SROADest = MI->getRawDest() == Inst; 1469 1470 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1471 1472 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1473 // If this is a memcpy/memmove, emit a GEP of the other element address. 1474 Value *OtherElt = 0; 1475 unsigned OtherEltAlign = MemAlignment; 1476 1477 if (OtherPtr) { 1478 Value *Idx[2] = { Zero, 1479 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1480 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1481 OtherPtr->getName()+"."+Twine(i), 1482 MI); 1483 uint64_t EltOffset; 1484 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1485 const Type *OtherTy = OtherPtrTy->getElementType(); 1486 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1487 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1488 } else { 1489 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 1490 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1491 } 1492 1493 // The alignment of the other pointer is the guaranteed alignment of the 1494 // element, which is affected by both the known alignment of the whole 1495 // mem intrinsic and the alignment of the element. If the alignment of 1496 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 1497 // known alignment is just 4 bytes. 1498 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 1499 } 1500 1501 Value *EltPtr = NewElts[i]; 1502 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 1503 1504 // If we got down to a scalar, insert a load or store as appropriate. 1505 if (EltTy->isSingleValueType()) { 1506 if (isa<MemTransferInst>(MI)) { 1507 if (SROADest) { 1508 // From Other to Alloca. 1509 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 1510 new StoreInst(Elt, EltPtr, MI); 1511 } else { 1512 // From Alloca to Other. 1513 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 1514 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 1515 } 1516 continue; 1517 } 1518 assert(isa<MemSetInst>(MI)); 1519 1520 // If the stored element is zero (common case), just store a null 1521 // constant. 1522 Constant *StoreVal; 1523 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 1524 if (CI->isZero()) { 1525 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 1526 } else { 1527 // If EltTy is a vector type, get the element type. 1528 const Type *ValTy = EltTy->getScalarType(); 1529 1530 // Construct an integer with the right value. 1531 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 1532 APInt OneVal(EltSize, CI->getZExtValue()); 1533 APInt TotalVal(OneVal); 1534 // Set each byte. 1535 for (unsigned i = 0; 8*i < EltSize; ++i) { 1536 TotalVal = TotalVal.shl(8); 1537 TotalVal |= OneVal; 1538 } 1539 1540 // Convert the integer value to the appropriate type. 1541 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 1542 if (ValTy->isPointerTy()) 1543 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 1544 else if (ValTy->isFloatingPointTy()) 1545 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 1546 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 1547 1548 // If the requested value was a vector constant, create it. 1549 if (EltTy != ValTy) { 1550 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 1551 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 1552 StoreVal = ConstantVector::get(&Elts[0], NumElts); 1553 } 1554 } 1555 new StoreInst(StoreVal, EltPtr, MI); 1556 continue; 1557 } 1558 // Otherwise, if we're storing a byte variable, use a memset call for 1559 // this element. 1560 } 1561 1562 unsigned EltSize = TD->getTypeAllocSize(EltTy); 1563 1564 IRBuilder<> Builder(MI); 1565 1566 // Finally, insert the meminst for this element. 1567 if (isa<MemSetInst>(MI)) { 1568 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 1569 MI->isVolatile()); 1570 } else { 1571 assert(isa<MemTransferInst>(MI)); 1572 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 1573 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 1574 1575 if (isa<MemCpyInst>(MI)) 1576 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 1577 else 1578 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 1579 } 1580 } 1581 DeadInsts.push_back(MI); 1582} 1583 1584/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 1585/// overwrites the entire allocation. Extract out the pieces of the stored 1586/// integer and store them individually. 1587void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 1588 SmallVector<AllocaInst*, 32> &NewElts){ 1589 // Extract each element out of the integer according to its structure offset 1590 // and store the element value to the individual alloca. 1591 Value *SrcVal = SI->getOperand(0); 1592 const Type *AllocaEltTy = AI->getAllocatedType(); 1593 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1594 1595 IRBuilder<> Builder(SI); 1596 1597 // Handle tail padding by extending the operand 1598 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 1599 SrcVal = Builder.CreateZExt(SrcVal, 1600 IntegerType::get(SI->getContext(), AllocaSizeBits)); 1601 1602 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 1603 << '\n'); 1604 1605 // There are two forms here: AI could be an array or struct. Both cases 1606 // have different ways to compute the element offset. 1607 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1608 const StructLayout *Layout = TD->getStructLayout(EltSTy); 1609 1610 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1611 // Get the number of bits to shift SrcVal to get the value. 1612 const Type *FieldTy = EltSTy->getElementType(i); 1613 uint64_t Shift = Layout->getElementOffsetInBits(i); 1614 1615 if (TD->isBigEndian()) 1616 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 1617 1618 Value *EltVal = SrcVal; 1619 if (Shift) { 1620 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1621 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 1622 } 1623 1624 // Truncate down to an integer of the right size. 1625 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1626 1627 // Ignore zero sized fields like {}, they obviously contain no data. 1628 if (FieldSizeBits == 0) continue; 1629 1630 if (FieldSizeBits != AllocaSizeBits) 1631 EltVal = Builder.CreateTrunc(EltVal, 1632 IntegerType::get(SI->getContext(), FieldSizeBits)); 1633 Value *DestField = NewElts[i]; 1634 if (EltVal->getType() == FieldTy) { 1635 // Storing to an integer field of this size, just do it. 1636 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 1637 // Bitcast to the right element type (for fp/vector values). 1638 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 1639 } else { 1640 // Otherwise, bitcast the dest pointer (for aggregates). 1641 DestField = Builder.CreateBitCast(DestField, 1642 PointerType::getUnqual(EltVal->getType())); 1643 } 1644 new StoreInst(EltVal, DestField, SI); 1645 } 1646 1647 } else { 1648 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1649 const Type *ArrayEltTy = ATy->getElementType(); 1650 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1651 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1652 1653 uint64_t Shift; 1654 1655 if (TD->isBigEndian()) 1656 Shift = AllocaSizeBits-ElementOffset; 1657 else 1658 Shift = 0; 1659 1660 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1661 // Ignore zero sized fields like {}, they obviously contain no data. 1662 if (ElementSizeBits == 0) continue; 1663 1664 Value *EltVal = SrcVal; 1665 if (Shift) { 1666 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1667 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 1668 } 1669 1670 // Truncate down to an integer of the right size. 1671 if (ElementSizeBits != AllocaSizeBits) 1672 EltVal = Builder.CreateTrunc(EltVal, 1673 IntegerType::get(SI->getContext(), 1674 ElementSizeBits)); 1675 Value *DestField = NewElts[i]; 1676 if (EltVal->getType() == ArrayEltTy) { 1677 // Storing to an integer field of this size, just do it. 1678 } else if (ArrayEltTy->isFloatingPointTy() || 1679 ArrayEltTy->isVectorTy()) { 1680 // Bitcast to the right element type (for fp/vector values). 1681 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 1682 } else { 1683 // Otherwise, bitcast the dest pointer (for aggregates). 1684 DestField = Builder.CreateBitCast(DestField, 1685 PointerType::getUnqual(EltVal->getType())); 1686 } 1687 new StoreInst(EltVal, DestField, SI); 1688 1689 if (TD->isBigEndian()) 1690 Shift -= ElementOffset; 1691 else 1692 Shift += ElementOffset; 1693 } 1694 } 1695 1696 DeadInsts.push_back(SI); 1697} 1698 1699/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1700/// an integer. Load the individual pieces to form the aggregate value. 1701void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1702 SmallVector<AllocaInst*, 32> &NewElts) { 1703 // Extract each element out of the NewElts according to its structure offset 1704 // and form the result value. 1705 const Type *AllocaEltTy = AI->getAllocatedType(); 1706 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1707 1708 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1709 << '\n'); 1710 1711 // There are two forms here: AI could be an array or struct. Both cases 1712 // have different ways to compute the element offset. 1713 const StructLayout *Layout = 0; 1714 uint64_t ArrayEltBitOffset = 0; 1715 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1716 Layout = TD->getStructLayout(EltSTy); 1717 } else { 1718 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1719 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1720 } 1721 1722 Value *ResultVal = 1723 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1724 1725 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1726 // Load the value from the alloca. If the NewElt is an aggregate, cast 1727 // the pointer to an integer of the same size before doing the load. 1728 Value *SrcField = NewElts[i]; 1729 const Type *FieldTy = 1730 cast<PointerType>(SrcField->getType())->getElementType(); 1731 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1732 1733 // Ignore zero sized fields like {}, they obviously contain no data. 1734 if (FieldSizeBits == 0) continue; 1735 1736 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1737 FieldSizeBits); 1738 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1739 !FieldTy->isVectorTy()) 1740 SrcField = new BitCastInst(SrcField, 1741 PointerType::getUnqual(FieldIntTy), 1742 "", LI); 1743 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1744 1745 // If SrcField is a fp or vector of the right size but that isn't an 1746 // integer type, bitcast to an integer so we can shift it. 1747 if (SrcField->getType() != FieldIntTy) 1748 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1749 1750 // Zero extend the field to be the same size as the final alloca so that 1751 // we can shift and insert it. 1752 if (SrcField->getType() != ResultVal->getType()) 1753 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1754 1755 // Determine the number of bits to shift SrcField. 1756 uint64_t Shift; 1757 if (Layout) // Struct case. 1758 Shift = Layout->getElementOffsetInBits(i); 1759 else // Array case. 1760 Shift = i*ArrayEltBitOffset; 1761 1762 if (TD->isBigEndian()) 1763 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1764 1765 if (Shift) { 1766 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1767 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1768 } 1769 1770 // Don't create an 'or x, 0' on the first iteration. 1771 if (!isa<Constant>(ResultVal) || 1772 !cast<Constant>(ResultVal)->isNullValue()) 1773 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1774 else 1775 ResultVal = SrcField; 1776 } 1777 1778 // Handle tail padding by truncating the result 1779 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1780 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1781 1782 LI->replaceAllUsesWith(ResultVal); 1783 DeadInsts.push_back(LI); 1784} 1785 1786/// HasPadding - Return true if the specified type has any structure or 1787/// alignment padding in between the elements that would be split apart 1788/// by SROA; return false otherwise. 1789static bool HasPadding(const Type *Ty, const TargetData &TD) { 1790 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1791 Ty = ATy->getElementType(); 1792 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1793 } 1794 1795 // SROA currently handles only Arrays and Structs. 1796 const StructType *STy = cast<StructType>(Ty); 1797 const StructLayout *SL = TD.getStructLayout(STy); 1798 unsigned PrevFieldBitOffset = 0; 1799 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1800 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1801 1802 // Check to see if there is any padding between this element and the 1803 // previous one. 1804 if (i) { 1805 unsigned PrevFieldEnd = 1806 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1807 if (PrevFieldEnd < FieldBitOffset) 1808 return true; 1809 } 1810 PrevFieldBitOffset = FieldBitOffset; 1811 } 1812 // Check for tail padding. 1813 if (unsigned EltCount = STy->getNumElements()) { 1814 unsigned PrevFieldEnd = PrevFieldBitOffset + 1815 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1816 if (PrevFieldEnd < SL->getSizeInBits()) 1817 return true; 1818 } 1819 return false; 1820} 1821 1822/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1823/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1824/// or 1 if safe after canonicalization has been performed. 1825bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1826 // Loop over the use list of the alloca. We can only transform it if all of 1827 // the users are safe to transform. 1828 AllocaInfo Info; 1829 1830 isSafeForScalarRepl(AI, AI, 0, Info); 1831 if (Info.isUnsafe) { 1832 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1833 return false; 1834 } 1835 1836 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1837 // source and destination, we have to be careful. In particular, the memcpy 1838 // could be moving around elements that live in structure padding of the LLVM 1839 // types, but may actually be used. In these cases, we refuse to promote the 1840 // struct. 1841 if (Info.isMemCpySrc && Info.isMemCpyDst && 1842 HasPadding(AI->getAllocatedType(), *TD)) 1843 return false; 1844 1845 // If the alloca never has an access to just *part* of it, but is accessed 1846 // via loads and stores, then we should use ConvertToScalarInfo to promote 1847 // the alloca instead of promoting each piece at a time and inserting fission 1848 // and fusion code. 1849 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 1850 // If the struct/array just has one element, use basic SRoA. 1851 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1852 if (ST->getNumElements() > 1) return false; 1853 } else { 1854 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 1855 return false; 1856 } 1857 } 1858 return true; 1859} 1860 1861 1862 1863/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1864/// some part of a constant global variable. This intentionally only accepts 1865/// constant expressions because we don't can't rewrite arbitrary instructions. 1866static bool PointsToConstantGlobal(Value *V) { 1867 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1868 return GV->isConstant(); 1869 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1870 if (CE->getOpcode() == Instruction::BitCast || 1871 CE->getOpcode() == Instruction::GetElementPtr) 1872 return PointsToConstantGlobal(CE->getOperand(0)); 1873 return false; 1874} 1875 1876/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1877/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1878/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1879/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1880/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1881/// the alloca, and if the source pointer is a pointer to a constant global, we 1882/// can optimize this. 1883static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 1884 bool isOffset) { 1885 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1886 User *U = cast<Instruction>(*UI); 1887 1888 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1889 // Ignore non-volatile loads, they are always ok. 1890 if (LI->isVolatile()) return false; 1891 continue; 1892 } 1893 1894 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1895 // If uses of the bitcast are ok, we are ok. 1896 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1897 return false; 1898 continue; 1899 } 1900 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 1901 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1902 // doesn't, it does. 1903 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1904 isOffset || !GEP->hasAllZeroIndices())) 1905 return false; 1906 continue; 1907 } 1908 1909 if (CallSite CS = U) { 1910 // If this is a readonly/readnone call site, then we know it is just a 1911 // load and we can ignore it. 1912 if (CS.onlyReadsMemory()) 1913 continue; 1914 1915 // If this is the function being called then we treat it like a load and 1916 // ignore it. 1917 if (CS.isCallee(UI)) 1918 continue; 1919 1920 // If this is being passed as a byval argument, the caller is making a 1921 // copy, so it is only a read of the alloca. 1922 unsigned ArgNo = CS.getArgumentNo(UI); 1923 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 1924 continue; 1925 } 1926 1927 // If this is isn't our memcpy/memmove, reject it as something we can't 1928 // handle. 1929 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 1930 if (MI == 0) 1931 return false; 1932 1933 // If the transfer is using the alloca as a source of the transfer, then 1934 // ignore it since it is a load (unless the transfer is volatile). 1935 if (UI.getOperandNo() == 1) { 1936 if (MI->isVolatile()) return false; 1937 continue; 1938 } 1939 1940 // If we already have seen a copy, reject the second one. 1941 if (TheCopy) return false; 1942 1943 // If the pointer has been offset from the start of the alloca, we can't 1944 // safely handle this. 1945 if (isOffset) return false; 1946 1947 // If the memintrinsic isn't using the alloca as the dest, reject it. 1948 if (UI.getOperandNo() != 0) return false; 1949 1950 // If the source of the memcpy/move is not a constant global, reject it. 1951 if (!PointsToConstantGlobal(MI->getSource())) 1952 return false; 1953 1954 // Otherwise, the transform is safe. Remember the copy instruction. 1955 TheCopy = MI; 1956 } 1957 return true; 1958} 1959 1960/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1961/// modified by a copy from a constant global. If we can prove this, we can 1962/// replace any uses of the alloca with uses of the global directly. 1963MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1964 MemTransferInst *TheCopy = 0; 1965 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1966 return TheCopy; 1967 return 0; 1968} 1969