ScalarReplAggregates.cpp revision 081c34b725980f995be9080eaec24cd3dfaaf065
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/Target/TargetData.h" 35#include "llvm/Transforms/Utils/PromoteMemToReg.h" 36#include "llvm/Transforms/Utils/Local.h" 37#include "llvm/Support/Debug.h" 38#include "llvm/Support/ErrorHandling.h" 39#include "llvm/Support/GetElementPtrTypeIterator.h" 40#include "llvm/Support/IRBuilder.h" 41#include "llvm/Support/MathExtras.h" 42#include "llvm/Support/raw_ostream.h" 43#include "llvm/ADT/SmallVector.h" 44#include "llvm/ADT/Statistic.h" 45using namespace llvm; 46 47STATISTIC(NumReplaced, "Number of allocas broken up"); 48STATISTIC(NumPromoted, "Number of allocas promoted"); 49STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 50STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 51 52namespace { 53 struct SROA : public FunctionPass { 54 static char ID; // Pass identification, replacement for typeid 55 explicit SROA(signed T = -1) : FunctionPass(ID) { 56 initializeSROAPass(*PassRegistry::getPassRegistry()); 57 if (T == -1) 58 SRThreshold = 128; 59 else 60 SRThreshold = T; 61 } 62 63 bool runOnFunction(Function &F); 64 65 bool performScalarRepl(Function &F); 66 bool performPromotion(Function &F); 67 68 // getAnalysisUsage - This pass does not require any passes, but we know it 69 // will not alter the CFG, so say so. 70 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 71 AU.addRequired<DominatorTree>(); 72 AU.addRequired<DominanceFrontier>(); 73 AU.setPreservesCFG(); 74 } 75 76 private: 77 TargetData *TD; 78 79 /// DeadInsts - Keep track of instructions we have made dead, so that 80 /// we can remove them after we are done working. 81 SmallVector<Value*, 32> DeadInsts; 82 83 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 84 /// information about the uses. All these fields are initialized to false 85 /// and set to true when something is learned. 86 struct AllocaInfo { 87 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 88 bool isUnsafe : 1; 89 90 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 91 bool isMemCpySrc : 1; 92 93 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 94 bool isMemCpyDst : 1; 95 96 AllocaInfo() 97 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false) {} 98 }; 99 100 unsigned SRThreshold; 101 102 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 103 104 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 105 106 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 107 AllocaInfo &Info); 108 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset, 109 AllocaInfo &Info); 110 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 111 const Type *MemOpType, bool isStore, AllocaInfo &Info); 112 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 113 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 114 const Type *&IdxTy); 115 116 void DoScalarReplacement(AllocaInst *AI, 117 std::vector<AllocaInst*> &WorkList); 118 void DeleteDeadInstructions(); 119 120 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 121 SmallVector<AllocaInst*, 32> &NewElts); 122 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 123 SmallVector<AllocaInst*, 32> &NewElts); 124 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 125 SmallVector<AllocaInst*, 32> &NewElts); 126 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 127 AllocaInst *AI, 128 SmallVector<AllocaInst*, 32> &NewElts); 129 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 130 SmallVector<AllocaInst*, 32> &NewElts); 131 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 132 SmallVector<AllocaInst*, 32> &NewElts); 133 134 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 135 }; 136} 137 138char SROA::ID = 0; 139INITIALIZE_PASS_BEGIN(SROA, "scalarrepl", 140 "Scalar Replacement of Aggregates", false, false) 141INITIALIZE_PASS_DEPENDENCY(DominatorTree) 142INITIALIZE_PASS_DEPENDENCY(DominanceFrontier) 143INITIALIZE_PASS_END(SROA, "scalarrepl", 144 "Scalar Replacement of Aggregates", false, false) 145 146// Public interface to the ScalarReplAggregates pass 147FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 148 return new SROA(Threshold); 149} 150 151 152//===----------------------------------------------------------------------===// 153// Convert To Scalar Optimization. 154//===----------------------------------------------------------------------===// 155 156namespace { 157/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 158/// optimization, which scans the uses of an alloca and determines if it can 159/// rewrite it in terms of a single new alloca that can be mem2reg'd. 160class ConvertToScalarInfo { 161 /// AllocaSize - The size of the alloca being considered. 162 unsigned AllocaSize; 163 const TargetData &TD; 164 165 /// IsNotTrivial - This is set to true if there is some access to the object 166 /// which means that mem2reg can't promote it. 167 bool IsNotTrivial; 168 169 /// VectorTy - This tracks the type that we should promote the vector to if 170 /// it is possible to turn it into a vector. This starts out null, and if it 171 /// isn't possible to turn into a vector type, it gets set to VoidTy. 172 const Type *VectorTy; 173 174 /// HadAVector - True if there is at least one vector access to the alloca. 175 /// We don't want to turn random arrays into vectors and use vector element 176 /// insert/extract, but if there are element accesses to something that is 177 /// also declared as a vector, we do want to promote to a vector. 178 bool HadAVector; 179 180public: 181 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 182 : AllocaSize(Size), TD(td) { 183 IsNotTrivial = false; 184 VectorTy = 0; 185 HadAVector = false; 186 } 187 188 AllocaInst *TryConvert(AllocaInst *AI); 189 190private: 191 bool CanConvertToScalar(Value *V, uint64_t Offset); 192 void MergeInType(const Type *In, uint64_t Offset); 193 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 194 195 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 196 uint64_t Offset, IRBuilder<> &Builder); 197 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 198 uint64_t Offset, IRBuilder<> &Builder); 199}; 200} // end anonymous namespace. 201 202 203/// IsVerbotenVectorType - Return true if this is a vector type ScalarRepl isn't 204/// allowed to form. We do this to avoid MMX types, which is a complete hack, 205/// but is required until the backend is fixed. 206static bool IsVerbotenVectorType(const VectorType *VTy, const Instruction *I) { 207 StringRef Triple(I->getParent()->getParent()->getParent()->getTargetTriple()); 208 if (!Triple.startswith("i386") && 209 !Triple.startswith("x86_64")) 210 return false; 211 212 // Reject all the MMX vector types. 213 switch (VTy->getNumElements()) { 214 default: return false; 215 case 1: return VTy->getElementType()->isIntegerTy(64); 216 case 2: return VTy->getElementType()->isIntegerTy(32); 217 case 4: return VTy->getElementType()->isIntegerTy(16); 218 case 8: return VTy->getElementType()->isIntegerTy(8); 219 } 220} 221 222 223/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 224/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 225/// alloca if possible or null if not. 226AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 227 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 228 // out. 229 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 230 return 0; 231 232 // If we were able to find a vector type that can handle this with 233 // insert/extract elements, and if there was at least one use that had 234 // a vector type, promote this to a vector. We don't want to promote 235 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 236 // we just get a lot of insert/extracts. If at least one vector is 237 // involved, then we probably really do have a union of vector/array. 238 const Type *NewTy; 239 if (VectorTy && VectorTy->isVectorTy() && HadAVector && 240 !IsVerbotenVectorType(cast<VectorType>(VectorTy), AI)) { 241 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 242 << *VectorTy << '\n'); 243 NewTy = VectorTy; // Use the vector type. 244 } else { 245 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 246 // Create and insert the integer alloca. 247 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 248 } 249 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 250 ConvertUsesToScalar(AI, NewAI, 0); 251 return NewAI; 252} 253 254/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy) 255/// so far at the offset specified by Offset (which is specified in bytes). 256/// 257/// There are two cases we handle here: 258/// 1) A union of vector types of the same size and potentially its elements. 259/// Here we turn element accesses into insert/extract element operations. 260/// This promotes a <4 x float> with a store of float to the third element 261/// into a <4 x float> that uses insert element. 262/// 2) A fully general blob of memory, which we turn into some (potentially 263/// large) integer type with extract and insert operations where the loads 264/// and stores would mutate the memory. We mark this by setting VectorTy 265/// to VoidTy. 266void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) { 267 // If we already decided to turn this into a blob of integer memory, there is 268 // nothing to be done. 269 if (VectorTy && VectorTy->isVoidTy()) 270 return; 271 272 // If this could be contributing to a vector, analyze it. 273 274 // If the In type is a vector that is the same size as the alloca, see if it 275 // matches the existing VecTy. 276 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 277 // Remember if we saw a vector type. 278 HadAVector = true; 279 280 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 281 // If we're storing/loading a vector of the right size, allow it as a 282 // vector. If this the first vector we see, remember the type so that 283 // we know the element size. If this is a subsequent access, ignore it 284 // even if it is a differing type but the same size. Worst case we can 285 // bitcast the resultant vectors. 286 if (VectorTy == 0) 287 VectorTy = VInTy; 288 return; 289 } 290 } else if (In->isFloatTy() || In->isDoubleTy() || 291 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 292 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 293 // If we're accessing something that could be an element of a vector, see 294 // if the implied vector agrees with what we already have and if Offset is 295 // compatible with it. 296 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 297 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 298 (VectorTy == 0 || 299 cast<VectorType>(VectorTy)->getElementType() 300 ->getPrimitiveSizeInBits()/8 == EltSize)) { 301 if (VectorTy == 0) 302 VectorTy = VectorType::get(In, AllocaSize/EltSize); 303 return; 304 } 305 } 306 307 // Otherwise, we have a case that we can't handle with an optimized vector 308 // form. We can still turn this into a large integer. 309 VectorTy = Type::getVoidTy(In->getContext()); 310} 311 312/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 313/// its accesses to a single vector type, return true and set VecTy to 314/// the new type. If we could convert the alloca into a single promotable 315/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 316/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 317/// is the current offset from the base of the alloca being analyzed. 318/// 319/// If we see at least one access to the value that is as a vector type, set the 320/// SawVec flag. 321bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 322 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 323 Instruction *User = cast<Instruction>(*UI); 324 325 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 326 // Don't break volatile loads. 327 if (LI->isVolatile()) 328 return false; 329 // Don't touch MMX operations. 330 if (LI->getType()->isX86_MMXTy()) 331 return false; 332 MergeInType(LI->getType(), Offset); 333 continue; 334 } 335 336 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 337 // Storing the pointer, not into the value? 338 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 339 // Don't touch MMX operations. 340 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 341 return false; 342 MergeInType(SI->getOperand(0)->getType(), Offset); 343 continue; 344 } 345 346 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 347 IsNotTrivial = true; // Can't be mem2reg'd. 348 if (!CanConvertToScalar(BCI, Offset)) 349 return false; 350 continue; 351 } 352 353 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 354 // If this is a GEP with a variable indices, we can't handle it. 355 if (!GEP->hasAllConstantIndices()) 356 return false; 357 358 // Compute the offset that this GEP adds to the pointer. 359 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 360 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 361 &Indices[0], Indices.size()); 362 // See if all uses can be converted. 363 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 364 return false; 365 IsNotTrivial = true; // Can't be mem2reg'd. 366 continue; 367 } 368 369 // If this is a constant sized memset of a constant value (e.g. 0) we can 370 // handle it. 371 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 372 // Store of constant value and constant size. 373 if (!isa<ConstantInt>(MSI->getValue()) || 374 !isa<ConstantInt>(MSI->getLength())) 375 return false; 376 IsNotTrivial = true; // Can't be mem2reg'd. 377 continue; 378 } 379 380 // If this is a memcpy or memmove into or out of the whole allocation, we 381 // can handle it like a load or store of the scalar type. 382 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 383 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 384 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 385 return false; 386 387 IsNotTrivial = true; // Can't be mem2reg'd. 388 continue; 389 } 390 391 // Otherwise, we cannot handle this! 392 return false; 393 } 394 395 return true; 396} 397 398/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 399/// directly. This happens when we are converting an "integer union" to a 400/// single integer scalar, or when we are converting a "vector union" to a 401/// vector with insert/extractelement instructions. 402/// 403/// Offset is an offset from the original alloca, in bits that need to be 404/// shifted to the right. By the end of this, there should be no uses of Ptr. 405void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 406 uint64_t Offset) { 407 while (!Ptr->use_empty()) { 408 Instruction *User = cast<Instruction>(Ptr->use_back()); 409 410 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 411 ConvertUsesToScalar(CI, NewAI, Offset); 412 CI->eraseFromParent(); 413 continue; 414 } 415 416 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 417 // Compute the offset that this GEP adds to the pointer. 418 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 419 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 420 &Indices[0], Indices.size()); 421 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 422 GEP->eraseFromParent(); 423 continue; 424 } 425 426 IRBuilder<> Builder(User->getParent(), User); 427 428 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 429 // The load is a bit extract from NewAI shifted right by Offset bits. 430 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 431 Value *NewLoadVal 432 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 433 LI->replaceAllUsesWith(NewLoadVal); 434 LI->eraseFromParent(); 435 continue; 436 } 437 438 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 439 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 440 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 441 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 442 Builder); 443 Builder.CreateStore(New, NewAI); 444 SI->eraseFromParent(); 445 446 // If the load we just inserted is now dead, then the inserted store 447 // overwrote the entire thing. 448 if (Old->use_empty()) 449 Old->eraseFromParent(); 450 continue; 451 } 452 453 // If this is a constant sized memset of a constant value (e.g. 0) we can 454 // transform it into a store of the expanded constant value. 455 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 456 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 457 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 458 if (NumBytes != 0) { 459 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 460 461 // Compute the value replicated the right number of times. 462 APInt APVal(NumBytes*8, Val); 463 464 // Splat the value if non-zero. 465 if (Val) 466 for (unsigned i = 1; i != NumBytes; ++i) 467 APVal |= APVal << 8; 468 469 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 470 Value *New = ConvertScalar_InsertValue( 471 ConstantInt::get(User->getContext(), APVal), 472 Old, Offset, Builder); 473 Builder.CreateStore(New, NewAI); 474 475 // If the load we just inserted is now dead, then the memset overwrote 476 // the entire thing. 477 if (Old->use_empty()) 478 Old->eraseFromParent(); 479 } 480 MSI->eraseFromParent(); 481 continue; 482 } 483 484 // If this is a memcpy or memmove into or out of the whole allocation, we 485 // can handle it like a load or store of the scalar type. 486 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 487 assert(Offset == 0 && "must be store to start of alloca"); 488 489 // If the source and destination are both to the same alloca, then this is 490 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 491 // as appropriate. 492 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject(0)); 493 494 if (MTI->getSource()->getUnderlyingObject(0) != OrigAI) { 495 // Dest must be OrigAI, change this to be a load from the original 496 // pointer (bitcasted), then a store to our new alloca. 497 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 498 Value *SrcPtr = MTI->getSource(); 499 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 500 501 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 502 SrcVal->setAlignment(MTI->getAlignment()); 503 Builder.CreateStore(SrcVal, NewAI); 504 } else if (MTI->getDest()->getUnderlyingObject(0) != OrigAI) { 505 // Src must be OrigAI, change this to be a load from NewAI then a store 506 // through the original dest pointer (bitcasted). 507 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 508 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 509 510 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 511 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 512 NewStore->setAlignment(MTI->getAlignment()); 513 } else { 514 // Noop transfer. Src == Dst 515 } 516 517 MTI->eraseFromParent(); 518 continue; 519 } 520 521 llvm_unreachable("Unsupported operation!"); 522 } 523} 524 525/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 526/// or vector value FromVal, extracting the bits from the offset specified by 527/// Offset. This returns the value, which is of type ToType. 528/// 529/// This happens when we are converting an "integer union" to a single 530/// integer scalar, or when we are converting a "vector union" to a vector with 531/// insert/extractelement instructions. 532/// 533/// Offset is an offset from the original alloca, in bits that need to be 534/// shifted to the right. 535Value *ConvertToScalarInfo:: 536ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 537 uint64_t Offset, IRBuilder<> &Builder) { 538 // If the load is of the whole new alloca, no conversion is needed. 539 if (FromVal->getType() == ToType && Offset == 0) 540 return FromVal; 541 542 // If the result alloca is a vector type, this is either an element 543 // access or a bitcast to another vector type of the same size. 544 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 545 if (ToType->isVectorTy()) 546 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 547 548 // Otherwise it must be an element access. 549 unsigned Elt = 0; 550 if (Offset) { 551 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 552 Elt = Offset/EltSize; 553 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 554 } 555 // Return the element extracted out of it. 556 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 557 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 558 if (V->getType() != ToType) 559 V = Builder.CreateBitCast(V, ToType, "tmp"); 560 return V; 561 } 562 563 // If ToType is a first class aggregate, extract out each of the pieces and 564 // use insertvalue's to form the FCA. 565 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 566 const StructLayout &Layout = *TD.getStructLayout(ST); 567 Value *Res = UndefValue::get(ST); 568 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 569 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 570 Offset+Layout.getElementOffsetInBits(i), 571 Builder); 572 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 573 } 574 return Res; 575 } 576 577 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 578 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 579 Value *Res = UndefValue::get(AT); 580 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 581 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 582 Offset+i*EltSize, Builder); 583 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 584 } 585 return Res; 586 } 587 588 // Otherwise, this must be a union that was converted to an integer value. 589 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 590 591 // If this is a big-endian system and the load is narrower than the 592 // full alloca type, we need to do a shift to get the right bits. 593 int ShAmt = 0; 594 if (TD.isBigEndian()) { 595 // On big-endian machines, the lowest bit is stored at the bit offset 596 // from the pointer given by getTypeStoreSizeInBits. This matters for 597 // integers with a bitwidth that is not a multiple of 8. 598 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 599 TD.getTypeStoreSizeInBits(ToType) - Offset; 600 } else { 601 ShAmt = Offset; 602 } 603 604 // Note: we support negative bitwidths (with shl) which are not defined. 605 // We do this to support (f.e.) loads off the end of a structure where 606 // only some bits are used. 607 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 608 FromVal = Builder.CreateLShr(FromVal, 609 ConstantInt::get(FromVal->getType(), 610 ShAmt), "tmp"); 611 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 612 FromVal = Builder.CreateShl(FromVal, 613 ConstantInt::get(FromVal->getType(), 614 -ShAmt), "tmp"); 615 616 // Finally, unconditionally truncate the integer to the right width. 617 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 618 if (LIBitWidth < NTy->getBitWidth()) 619 FromVal = 620 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 621 LIBitWidth), "tmp"); 622 else if (LIBitWidth > NTy->getBitWidth()) 623 FromVal = 624 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 625 LIBitWidth), "tmp"); 626 627 // If the result is an integer, this is a trunc or bitcast. 628 if (ToType->isIntegerTy()) { 629 // Should be done. 630 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 631 // Just do a bitcast, we know the sizes match up. 632 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 633 } else { 634 // Otherwise must be a pointer. 635 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 636 } 637 assert(FromVal->getType() == ToType && "Didn't convert right?"); 638 return FromVal; 639} 640 641/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 642/// or vector value "Old" at the offset specified by Offset. 643/// 644/// This happens when we are converting an "integer union" to a 645/// single integer scalar, or when we are converting a "vector union" to a 646/// vector with insert/extractelement instructions. 647/// 648/// Offset is an offset from the original alloca, in bits that need to be 649/// shifted to the right. 650Value *ConvertToScalarInfo:: 651ConvertScalar_InsertValue(Value *SV, Value *Old, 652 uint64_t Offset, IRBuilder<> &Builder) { 653 // Convert the stored type to the actual type, shift it left to insert 654 // then 'or' into place. 655 const Type *AllocaType = Old->getType(); 656 LLVMContext &Context = Old->getContext(); 657 658 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 659 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 660 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 661 662 // Changing the whole vector with memset or with an access of a different 663 // vector type? 664 if (ValSize == VecSize) 665 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 666 667 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 668 669 // Must be an element insertion. 670 unsigned Elt = Offset/EltSize; 671 672 if (SV->getType() != VTy->getElementType()) 673 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 674 675 SV = Builder.CreateInsertElement(Old, SV, 676 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 677 "tmp"); 678 return SV; 679 } 680 681 // If SV is a first-class aggregate value, insert each value recursively. 682 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 683 const StructLayout &Layout = *TD.getStructLayout(ST); 684 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 685 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 686 Old = ConvertScalar_InsertValue(Elt, Old, 687 Offset+Layout.getElementOffsetInBits(i), 688 Builder); 689 } 690 return Old; 691 } 692 693 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 694 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 695 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 696 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 697 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 698 } 699 return Old; 700 } 701 702 // If SV is a float, convert it to the appropriate integer type. 703 // If it is a pointer, do the same. 704 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 705 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 706 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 707 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 708 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 709 SV = Builder.CreateBitCast(SV, 710 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 711 else if (SV->getType()->isPointerTy()) 712 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 713 714 // Zero extend or truncate the value if needed. 715 if (SV->getType() != AllocaType) { 716 if (SV->getType()->getPrimitiveSizeInBits() < 717 AllocaType->getPrimitiveSizeInBits()) 718 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 719 else { 720 // Truncation may be needed if storing more than the alloca can hold 721 // (undefined behavior). 722 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 723 SrcWidth = DestWidth; 724 SrcStoreWidth = DestStoreWidth; 725 } 726 } 727 728 // If this is a big-endian system and the store is narrower than the 729 // full alloca type, we need to do a shift to get the right bits. 730 int ShAmt = 0; 731 if (TD.isBigEndian()) { 732 // On big-endian machines, the lowest bit is stored at the bit offset 733 // from the pointer given by getTypeStoreSizeInBits. This matters for 734 // integers with a bitwidth that is not a multiple of 8. 735 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 736 } else { 737 ShAmt = Offset; 738 } 739 740 // Note: we support negative bitwidths (with shr) which are not defined. 741 // We do this to support (f.e.) stores off the end of a structure where 742 // only some bits in the structure are set. 743 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 744 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 745 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 746 ShAmt), "tmp"); 747 Mask <<= ShAmt; 748 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 749 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 750 -ShAmt), "tmp"); 751 Mask = Mask.lshr(-ShAmt); 752 } 753 754 // Mask out the bits we are about to insert from the old value, and or 755 // in the new bits. 756 if (SrcWidth != DestWidth) { 757 assert(DestWidth > SrcWidth); 758 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 759 SV = Builder.CreateOr(Old, SV, "ins"); 760 } 761 return SV; 762} 763 764 765//===----------------------------------------------------------------------===// 766// SRoA Driver 767//===----------------------------------------------------------------------===// 768 769 770bool SROA::runOnFunction(Function &F) { 771 TD = getAnalysisIfAvailable<TargetData>(); 772 773 bool Changed = performPromotion(F); 774 775 // FIXME: ScalarRepl currently depends on TargetData more than it 776 // theoretically needs to. It should be refactored in order to support 777 // target-independent IR. Until this is done, just skip the actual 778 // scalar-replacement portion of this pass. 779 if (!TD) return Changed; 780 781 while (1) { 782 bool LocalChange = performScalarRepl(F); 783 if (!LocalChange) break; // No need to repromote if no scalarrepl 784 Changed = true; 785 LocalChange = performPromotion(F); 786 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 787 } 788 789 return Changed; 790} 791 792 793bool SROA::performPromotion(Function &F) { 794 std::vector<AllocaInst*> Allocas; 795 DominatorTree &DT = getAnalysis<DominatorTree>(); 796 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 797 798 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 799 800 bool Changed = false; 801 802 while (1) { 803 Allocas.clear(); 804 805 // Find allocas that are safe to promote, by looking at all instructions in 806 // the entry node 807 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 808 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 809 if (isAllocaPromotable(AI)) 810 Allocas.push_back(AI); 811 812 if (Allocas.empty()) break; 813 814 PromoteMemToReg(Allocas, DT, DF); 815 NumPromoted += Allocas.size(); 816 Changed = true; 817 } 818 819 return Changed; 820} 821 822 823/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 824/// SROA. It must be a struct or array type with a small number of elements. 825static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 826 const Type *T = AI->getAllocatedType(); 827 // Do not promote any struct into more than 32 separate vars. 828 if (const StructType *ST = dyn_cast<StructType>(T)) 829 return ST->getNumElements() <= 32; 830 // Arrays are much less likely to be safe for SROA; only consider 831 // them if they are very small. 832 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 833 return AT->getNumElements() <= 8; 834 return false; 835} 836 837 838// performScalarRepl - This algorithm is a simple worklist driven algorithm, 839// which runs on all of the malloc/alloca instructions in the function, removing 840// them if they are only used by getelementptr instructions. 841// 842bool SROA::performScalarRepl(Function &F) { 843 std::vector<AllocaInst*> WorkList; 844 845 // Scan the entry basic block, adding allocas to the worklist. 846 BasicBlock &BB = F.getEntryBlock(); 847 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 848 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 849 WorkList.push_back(A); 850 851 // Process the worklist 852 bool Changed = false; 853 while (!WorkList.empty()) { 854 AllocaInst *AI = WorkList.back(); 855 WorkList.pop_back(); 856 857 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 858 // with unused elements. 859 if (AI->use_empty()) { 860 AI->eraseFromParent(); 861 Changed = true; 862 continue; 863 } 864 865 // If this alloca is impossible for us to promote, reject it early. 866 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 867 continue; 868 869 // Check to see if this allocation is only modified by a memcpy/memmove from 870 // a constant global. If this is the case, we can change all users to use 871 // the constant global instead. This is commonly produced by the CFE by 872 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 873 // is only subsequently read. 874 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 875 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 876 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 877 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 878 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 879 TheCopy->eraseFromParent(); // Don't mutate the global. 880 AI->eraseFromParent(); 881 ++NumGlobals; 882 Changed = true; 883 continue; 884 } 885 886 // Check to see if we can perform the core SROA transformation. We cannot 887 // transform the allocation instruction if it is an array allocation 888 // (allocations OF arrays are ok though), and an allocation of a scalar 889 // value cannot be decomposed at all. 890 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 891 892 // Do not promote [0 x %struct]. 893 if (AllocaSize == 0) continue; 894 895 // Do not promote any struct whose size is too big. 896 if (AllocaSize > SRThreshold) continue; 897 898 // If the alloca looks like a good candidate for scalar replacement, and if 899 // all its users can be transformed, then split up the aggregate into its 900 // separate elements. 901 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 902 DoScalarReplacement(AI, WorkList); 903 Changed = true; 904 continue; 905 } 906 907 // If we can turn this aggregate value (potentially with casts) into a 908 // simple scalar value that can be mem2reg'd into a register value. 909 // IsNotTrivial tracks whether this is something that mem2reg could have 910 // promoted itself. If so, we don't want to transform it needlessly. Note 911 // that we can't just check based on the type: the alloca may be of an i32 912 // but that has pointer arithmetic to set byte 3 of it or something. 913 if (AllocaInst *NewAI = 914 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 915 NewAI->takeName(AI); 916 AI->eraseFromParent(); 917 ++NumConverted; 918 Changed = true; 919 continue; 920 } 921 922 // Otherwise, couldn't process this alloca. 923 } 924 925 return Changed; 926} 927 928/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 929/// predicate, do SROA now. 930void SROA::DoScalarReplacement(AllocaInst *AI, 931 std::vector<AllocaInst*> &WorkList) { 932 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 933 SmallVector<AllocaInst*, 32> ElementAllocas; 934 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 935 ElementAllocas.reserve(ST->getNumContainedTypes()); 936 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 937 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 938 AI->getAlignment(), 939 AI->getName() + "." + Twine(i), AI); 940 ElementAllocas.push_back(NA); 941 WorkList.push_back(NA); // Add to worklist for recursive processing 942 } 943 } else { 944 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 945 ElementAllocas.reserve(AT->getNumElements()); 946 const Type *ElTy = AT->getElementType(); 947 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 948 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 949 AI->getName() + "." + Twine(i), AI); 950 ElementAllocas.push_back(NA); 951 WorkList.push_back(NA); // Add to worklist for recursive processing 952 } 953 } 954 955 // Now that we have created the new alloca instructions, rewrite all the 956 // uses of the old alloca. 957 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 958 959 // Now erase any instructions that were made dead while rewriting the alloca. 960 DeleteDeadInstructions(); 961 AI->eraseFromParent(); 962 963 ++NumReplaced; 964} 965 966/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 967/// recursively including all their operands that become trivially dead. 968void SROA::DeleteDeadInstructions() { 969 while (!DeadInsts.empty()) { 970 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 971 972 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 973 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 974 // Zero out the operand and see if it becomes trivially dead. 975 // (But, don't add allocas to the dead instruction list -- they are 976 // already on the worklist and will be deleted separately.) 977 *OI = 0; 978 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 979 DeadInsts.push_back(U); 980 } 981 982 I->eraseFromParent(); 983 } 984} 985 986/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 987/// performing scalar replacement of alloca AI. The results are flagged in 988/// the Info parameter. Offset indicates the position within AI that is 989/// referenced by this instruction. 990void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 991 AllocaInfo &Info) { 992 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 993 Instruction *User = cast<Instruction>(*UI); 994 995 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 996 isSafeForScalarRepl(BC, AI, Offset, Info); 997 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 998 uint64_t GEPOffset = Offset; 999 isSafeGEP(GEPI, AI, GEPOffset, Info); 1000 if (!Info.isUnsafe) 1001 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 1002 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1003 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1004 if (Length) 1005 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 1006 UI.getOperandNo() == 0, Info); 1007 else 1008 MarkUnsafe(Info); 1009 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1010 if (!LI->isVolatile()) { 1011 const Type *LIType = LI->getType(); 1012 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 1013 LIType, false, Info); 1014 } else 1015 MarkUnsafe(Info); 1016 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1017 // Store is ok if storing INTO the pointer, not storing the pointer 1018 if (!SI->isVolatile() && SI->getOperand(0) != I) { 1019 const Type *SIType = SI->getOperand(0)->getType(); 1020 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 1021 SIType, true, Info); 1022 } else 1023 MarkUnsafe(Info); 1024 } else { 1025 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 1026 MarkUnsafe(Info); 1027 } 1028 if (Info.isUnsafe) return; 1029 } 1030} 1031 1032/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1033/// replacement. It is safe when all the indices are constant, in-bounds 1034/// references, and when the resulting offset corresponds to an element within 1035/// the alloca type. The results are flagged in the Info parameter. Upon 1036/// return, Offset is adjusted as specified by the GEP indices. 1037void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 1038 uint64_t &Offset, AllocaInfo &Info) { 1039 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1040 if (GEPIt == E) 1041 return; 1042 1043 // Walk through the GEP type indices, checking the types that this indexes 1044 // into. 1045 for (; GEPIt != E; ++GEPIt) { 1046 // Ignore struct elements, no extra checking needed for these. 1047 if ((*GEPIt)->isStructTy()) 1048 continue; 1049 1050 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1051 if (!IdxVal) 1052 return MarkUnsafe(Info); 1053 } 1054 1055 // Compute the offset due to this GEP and check if the alloca has a 1056 // component element at that offset. 1057 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1058 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1059 &Indices[0], Indices.size()); 1060 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 1061 MarkUnsafe(Info); 1062} 1063 1064/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1065/// alloca or has an offset and size that corresponds to a component element 1066/// within it. The offset checked here may have been formed from a GEP with a 1067/// pointer bitcasted to a different type. 1068void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 1069 const Type *MemOpType, bool isStore, 1070 AllocaInfo &Info) { 1071 // Check if this is a load/store of the entire alloca. 1072 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 1073 bool UsesAggregateType = (MemOpType == AI->getAllocatedType()); 1074 // This is safe for MemIntrinsics (where MemOpType is 0), integer types 1075 // (which are essentially the same as the MemIntrinsics, especially with 1076 // regard to copying padding between elements), or references using the 1077 // aggregate type of the alloca. 1078 if (!MemOpType || MemOpType->isIntegerTy() || UsesAggregateType) { 1079 if (!UsesAggregateType) { 1080 if (isStore) 1081 Info.isMemCpyDst = true; 1082 else 1083 Info.isMemCpySrc = true; 1084 } 1085 return; 1086 } 1087 } 1088 // Check if the offset/size correspond to a component within the alloca type. 1089 const Type *T = AI->getAllocatedType(); 1090 if (TypeHasComponent(T, Offset, MemSize)) 1091 return; 1092 1093 return MarkUnsafe(Info); 1094} 1095 1096/// TypeHasComponent - Return true if T has a component type with the 1097/// specified offset and size. If Size is zero, do not check the size. 1098bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1099 const Type *EltTy; 1100 uint64_t EltSize; 1101 if (const StructType *ST = dyn_cast<StructType>(T)) { 1102 const StructLayout *Layout = TD->getStructLayout(ST); 1103 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1104 EltTy = ST->getContainedType(EltIdx); 1105 EltSize = TD->getTypeAllocSize(EltTy); 1106 Offset -= Layout->getElementOffset(EltIdx); 1107 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1108 EltTy = AT->getElementType(); 1109 EltSize = TD->getTypeAllocSize(EltTy); 1110 if (Offset >= AT->getNumElements() * EltSize) 1111 return false; 1112 Offset %= EltSize; 1113 } else { 1114 return false; 1115 } 1116 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1117 return true; 1118 // Check if the component spans multiple elements. 1119 if (Offset + Size > EltSize) 1120 return false; 1121 return TypeHasComponent(EltTy, Offset, Size); 1122} 1123 1124/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1125/// the instruction I, which references it, to use the separate elements. 1126/// Offset indicates the position within AI that is referenced by this 1127/// instruction. 1128void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1129 SmallVector<AllocaInst*, 32> &NewElts) { 1130 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1131 Instruction *User = cast<Instruction>(*UI); 1132 1133 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1134 RewriteBitCast(BC, AI, Offset, NewElts); 1135 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1136 RewriteGEP(GEPI, AI, Offset, NewElts); 1137 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1138 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1139 uint64_t MemSize = Length->getZExtValue(); 1140 if (Offset == 0 && 1141 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1142 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1143 // Otherwise the intrinsic can only touch a single element and the 1144 // address operand will be updated, so nothing else needs to be done. 1145 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1146 const Type *LIType = LI->getType(); 1147 if (LIType == AI->getAllocatedType()) { 1148 // Replace: 1149 // %res = load { i32, i32 }* %alloc 1150 // with: 1151 // %load.0 = load i32* %alloc.0 1152 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1153 // %load.1 = load i32* %alloc.1 1154 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1155 // (Also works for arrays instead of structs) 1156 Value *Insert = UndefValue::get(LIType); 1157 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1158 Value *Load = new LoadInst(NewElts[i], "load", LI); 1159 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1160 } 1161 LI->replaceAllUsesWith(Insert); 1162 DeadInsts.push_back(LI); 1163 } else if (LIType->isIntegerTy() && 1164 TD->getTypeAllocSize(LIType) == 1165 TD->getTypeAllocSize(AI->getAllocatedType())) { 1166 // If this is a load of the entire alloca to an integer, rewrite it. 1167 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1168 } 1169 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1170 Value *Val = SI->getOperand(0); 1171 const Type *SIType = Val->getType(); 1172 if (SIType == AI->getAllocatedType()) { 1173 // Replace: 1174 // store { i32, i32 } %val, { i32, i32 }* %alloc 1175 // with: 1176 // %val.0 = extractvalue { i32, i32 } %val, 0 1177 // store i32 %val.0, i32* %alloc.0 1178 // %val.1 = extractvalue { i32, i32 } %val, 1 1179 // store i32 %val.1, i32* %alloc.1 1180 // (Also works for arrays instead of structs) 1181 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1182 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1183 new StoreInst(Extract, NewElts[i], SI); 1184 } 1185 DeadInsts.push_back(SI); 1186 } else if (SIType->isIntegerTy() && 1187 TD->getTypeAllocSize(SIType) == 1188 TD->getTypeAllocSize(AI->getAllocatedType())) { 1189 // If this is a store of the entire alloca from an integer, rewrite it. 1190 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1191 } 1192 } 1193 } 1194} 1195 1196/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1197/// and recursively continue updating all of its uses. 1198void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1199 SmallVector<AllocaInst*, 32> &NewElts) { 1200 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1201 if (BC->getOperand(0) != AI) 1202 return; 1203 1204 // The bitcast references the original alloca. Replace its uses with 1205 // references to the first new element alloca. 1206 Instruction *Val = NewElts[0]; 1207 if (Val->getType() != BC->getDestTy()) { 1208 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1209 Val->takeName(BC); 1210 } 1211 BC->replaceAllUsesWith(Val); 1212 DeadInsts.push_back(BC); 1213} 1214 1215/// FindElementAndOffset - Return the index of the element containing Offset 1216/// within the specified type, which must be either a struct or an array. 1217/// Sets T to the type of the element and Offset to the offset within that 1218/// element. IdxTy is set to the type of the index result to be used in a 1219/// GEP instruction. 1220uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1221 const Type *&IdxTy) { 1222 uint64_t Idx = 0; 1223 if (const StructType *ST = dyn_cast<StructType>(T)) { 1224 const StructLayout *Layout = TD->getStructLayout(ST); 1225 Idx = Layout->getElementContainingOffset(Offset); 1226 T = ST->getContainedType(Idx); 1227 Offset -= Layout->getElementOffset(Idx); 1228 IdxTy = Type::getInt32Ty(T->getContext()); 1229 return Idx; 1230 } 1231 const ArrayType *AT = cast<ArrayType>(T); 1232 T = AT->getElementType(); 1233 uint64_t EltSize = TD->getTypeAllocSize(T); 1234 Idx = Offset / EltSize; 1235 Offset -= Idx * EltSize; 1236 IdxTy = Type::getInt64Ty(T->getContext()); 1237 return Idx; 1238} 1239 1240/// RewriteGEP - Check if this GEP instruction moves the pointer across 1241/// elements of the alloca that are being split apart, and if so, rewrite 1242/// the GEP to be relative to the new element. 1243void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1244 SmallVector<AllocaInst*, 32> &NewElts) { 1245 uint64_t OldOffset = Offset; 1246 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1247 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1248 &Indices[0], Indices.size()); 1249 1250 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1251 1252 const Type *T = AI->getAllocatedType(); 1253 const Type *IdxTy; 1254 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1255 if (GEPI->getOperand(0) == AI) 1256 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1257 1258 T = AI->getAllocatedType(); 1259 uint64_t EltOffset = Offset; 1260 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1261 1262 // If this GEP does not move the pointer across elements of the alloca 1263 // being split, then it does not needs to be rewritten. 1264 if (Idx == OldIdx) 1265 return; 1266 1267 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1268 SmallVector<Value*, 8> NewArgs; 1269 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1270 while (EltOffset != 0) { 1271 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1272 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1273 } 1274 Instruction *Val = NewElts[Idx]; 1275 if (NewArgs.size() > 1) { 1276 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1277 NewArgs.end(), "", GEPI); 1278 Val->takeName(GEPI); 1279 } 1280 if (Val->getType() != GEPI->getType()) 1281 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1282 GEPI->replaceAllUsesWith(Val); 1283 DeadInsts.push_back(GEPI); 1284} 1285 1286/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1287/// Rewrite it to copy or set the elements of the scalarized memory. 1288void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1289 AllocaInst *AI, 1290 SmallVector<AllocaInst*, 32> &NewElts) { 1291 // If this is a memcpy/memmove, construct the other pointer as the 1292 // appropriate type. The "Other" pointer is the pointer that goes to memory 1293 // that doesn't have anything to do with the alloca that we are promoting. For 1294 // memset, this Value* stays null. 1295 Value *OtherPtr = 0; 1296 unsigned MemAlignment = MI->getAlignment(); 1297 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1298 if (Inst == MTI->getRawDest()) 1299 OtherPtr = MTI->getRawSource(); 1300 else { 1301 assert(Inst == MTI->getRawSource()); 1302 OtherPtr = MTI->getRawDest(); 1303 } 1304 } 1305 1306 // If there is an other pointer, we want to convert it to the same pointer 1307 // type as AI has, so we can GEP through it safely. 1308 if (OtherPtr) { 1309 unsigned AddrSpace = 1310 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 1311 1312 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1313 // optimization, but it's also required to detect the corner case where 1314 // both pointer operands are referencing the same memory, and where 1315 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1316 // function is only called for mem intrinsics that access the whole 1317 // aggregate, so non-zero GEPs are not an issue here.) 1318 OtherPtr = OtherPtr->stripPointerCasts(); 1319 1320 // Copying the alloca to itself is a no-op: just delete it. 1321 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1322 // This code will run twice for a no-op memcpy -- once for each operand. 1323 // Put only one reference to MI on the DeadInsts list. 1324 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1325 E = DeadInsts.end(); I != E; ++I) 1326 if (*I == MI) return; 1327 DeadInsts.push_back(MI); 1328 return; 1329 } 1330 1331 // If the pointer is not the right type, insert a bitcast to the right 1332 // type. 1333 const Type *NewTy = 1334 PointerType::get(AI->getType()->getElementType(), AddrSpace); 1335 1336 if (OtherPtr->getType() != NewTy) 1337 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 1338 } 1339 1340 // Process each element of the aggregate. 1341 Value *TheFn = MI->getCalledValue(); 1342 const Type *BytePtrTy = MI->getRawDest()->getType(); 1343 bool SROADest = MI->getRawDest() == Inst; 1344 1345 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1346 1347 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1348 // If this is a memcpy/memmove, emit a GEP of the other element address. 1349 Value *OtherElt = 0; 1350 unsigned OtherEltAlign = MemAlignment; 1351 1352 if (OtherPtr) { 1353 Value *Idx[2] = { Zero, 1354 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1355 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1356 OtherPtr->getName()+"."+Twine(i), 1357 MI); 1358 uint64_t EltOffset; 1359 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1360 const Type *OtherTy = OtherPtrTy->getElementType(); 1361 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 1362 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1363 } else { 1364 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 1365 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1366 } 1367 1368 // The alignment of the other pointer is the guaranteed alignment of the 1369 // element, which is affected by both the known alignment of the whole 1370 // mem intrinsic and the alignment of the element. If the alignment of 1371 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 1372 // known alignment is just 4 bytes. 1373 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 1374 } 1375 1376 Value *EltPtr = NewElts[i]; 1377 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 1378 1379 // If we got down to a scalar, insert a load or store as appropriate. 1380 if (EltTy->isSingleValueType()) { 1381 if (isa<MemTransferInst>(MI)) { 1382 if (SROADest) { 1383 // From Other to Alloca. 1384 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 1385 new StoreInst(Elt, EltPtr, MI); 1386 } else { 1387 // From Alloca to Other. 1388 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 1389 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 1390 } 1391 continue; 1392 } 1393 assert(isa<MemSetInst>(MI)); 1394 1395 // If the stored element is zero (common case), just store a null 1396 // constant. 1397 Constant *StoreVal; 1398 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 1399 if (CI->isZero()) { 1400 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 1401 } else { 1402 // If EltTy is a vector type, get the element type. 1403 const Type *ValTy = EltTy->getScalarType(); 1404 1405 // Construct an integer with the right value. 1406 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 1407 APInt OneVal(EltSize, CI->getZExtValue()); 1408 APInt TotalVal(OneVal); 1409 // Set each byte. 1410 for (unsigned i = 0; 8*i < EltSize; ++i) { 1411 TotalVal = TotalVal.shl(8); 1412 TotalVal |= OneVal; 1413 } 1414 1415 // Convert the integer value to the appropriate type. 1416 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 1417 if (ValTy->isPointerTy()) 1418 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 1419 else if (ValTy->isFloatingPointTy()) 1420 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 1421 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 1422 1423 // If the requested value was a vector constant, create it. 1424 if (EltTy != ValTy) { 1425 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 1426 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 1427 StoreVal = ConstantVector::get(&Elts[0], NumElts); 1428 } 1429 } 1430 new StoreInst(StoreVal, EltPtr, MI); 1431 continue; 1432 } 1433 // Otherwise, if we're storing a byte variable, use a memset call for 1434 // this element. 1435 } 1436 1437 // Cast the element pointer to BytePtrTy. 1438 if (EltPtr->getType() != BytePtrTy) 1439 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI); 1440 1441 // Cast the other pointer (if we have one) to BytePtrTy. 1442 if (OtherElt && OtherElt->getType() != BytePtrTy) { 1443 // Preserve address space of OtherElt 1444 const PointerType* OtherPTy = cast<PointerType>(OtherElt->getType()); 1445 const PointerType* PTy = cast<PointerType>(BytePtrTy); 1446 if (OtherPTy->getElementType() != PTy->getElementType()) { 1447 Type *NewOtherPTy = PointerType::get(PTy->getElementType(), 1448 OtherPTy->getAddressSpace()); 1449 OtherElt = new BitCastInst(OtherElt, NewOtherPTy, 1450 OtherElt->getName(), MI); 1451 } 1452 } 1453 1454 unsigned EltSize = TD->getTypeAllocSize(EltTy); 1455 1456 // Finally, insert the meminst for this element. 1457 if (isa<MemTransferInst>(MI)) { 1458 Value *Ops[] = { 1459 SROADest ? EltPtr : OtherElt, // Dest ptr 1460 SROADest ? OtherElt : EltPtr, // Src ptr 1461 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size 1462 // Align 1463 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign), 1464 MI->getVolatileCst() 1465 }; 1466 // In case we fold the address space overloaded memcpy of A to B 1467 // with memcpy of B to C, change the function to be a memcpy of A to C. 1468 const Type *Tys[] = { Ops[0]->getType(), Ops[1]->getType(), 1469 Ops[2]->getType() }; 1470 Module *M = MI->getParent()->getParent()->getParent(); 1471 TheFn = Intrinsic::getDeclaration(M, MI->getIntrinsicID(), Tys, 3); 1472 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1473 } else { 1474 assert(isa<MemSetInst>(MI)); 1475 Value *Ops[] = { 1476 EltPtr, MI->getArgOperand(1), // Dest, Value, 1477 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size 1478 Zero, // Align 1479 ConstantInt::get(Type::getInt1Ty(MI->getContext()), 0) // isVolatile 1480 }; 1481 const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() }; 1482 Module *M = MI->getParent()->getParent()->getParent(); 1483 TheFn = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2); 1484 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1485 } 1486 } 1487 DeadInsts.push_back(MI); 1488} 1489 1490/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 1491/// overwrites the entire allocation. Extract out the pieces of the stored 1492/// integer and store them individually. 1493void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 1494 SmallVector<AllocaInst*, 32> &NewElts){ 1495 // Extract each element out of the integer according to its structure offset 1496 // and store the element value to the individual alloca. 1497 Value *SrcVal = SI->getOperand(0); 1498 const Type *AllocaEltTy = AI->getAllocatedType(); 1499 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1500 1501 // Handle tail padding by extending the operand 1502 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 1503 SrcVal = new ZExtInst(SrcVal, 1504 IntegerType::get(SI->getContext(), AllocaSizeBits), 1505 "", SI); 1506 1507 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 1508 << '\n'); 1509 1510 // There are two forms here: AI could be an array or struct. Both cases 1511 // have different ways to compute the element offset. 1512 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1513 const StructLayout *Layout = TD->getStructLayout(EltSTy); 1514 1515 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1516 // Get the number of bits to shift SrcVal to get the value. 1517 const Type *FieldTy = EltSTy->getElementType(i); 1518 uint64_t Shift = Layout->getElementOffsetInBits(i); 1519 1520 if (TD->isBigEndian()) 1521 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 1522 1523 Value *EltVal = SrcVal; 1524 if (Shift) { 1525 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1526 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1527 "sroa.store.elt", SI); 1528 } 1529 1530 // Truncate down to an integer of the right size. 1531 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1532 1533 // Ignore zero sized fields like {}, they obviously contain no data. 1534 if (FieldSizeBits == 0) continue; 1535 1536 if (FieldSizeBits != AllocaSizeBits) 1537 EltVal = new TruncInst(EltVal, 1538 IntegerType::get(SI->getContext(), FieldSizeBits), 1539 "", SI); 1540 Value *DestField = NewElts[i]; 1541 if (EltVal->getType() == FieldTy) { 1542 // Storing to an integer field of this size, just do it. 1543 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 1544 // Bitcast to the right element type (for fp/vector values). 1545 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 1546 } else { 1547 // Otherwise, bitcast the dest pointer (for aggregates). 1548 DestField = new BitCastInst(DestField, 1549 PointerType::getUnqual(EltVal->getType()), 1550 "", SI); 1551 } 1552 new StoreInst(EltVal, DestField, SI); 1553 } 1554 1555 } else { 1556 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1557 const Type *ArrayEltTy = ATy->getElementType(); 1558 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1559 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1560 1561 uint64_t Shift; 1562 1563 if (TD->isBigEndian()) 1564 Shift = AllocaSizeBits-ElementOffset; 1565 else 1566 Shift = 0; 1567 1568 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1569 // Ignore zero sized fields like {}, they obviously contain no data. 1570 if (ElementSizeBits == 0) continue; 1571 1572 Value *EltVal = SrcVal; 1573 if (Shift) { 1574 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1575 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1576 "sroa.store.elt", SI); 1577 } 1578 1579 // Truncate down to an integer of the right size. 1580 if (ElementSizeBits != AllocaSizeBits) 1581 EltVal = new TruncInst(EltVal, 1582 IntegerType::get(SI->getContext(), 1583 ElementSizeBits),"",SI); 1584 Value *DestField = NewElts[i]; 1585 if (EltVal->getType() == ArrayEltTy) { 1586 // Storing to an integer field of this size, just do it. 1587 } else if (ArrayEltTy->isFloatingPointTy() || 1588 ArrayEltTy->isVectorTy()) { 1589 // Bitcast to the right element type (for fp/vector values). 1590 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1591 } else { 1592 // Otherwise, bitcast the dest pointer (for aggregates). 1593 DestField = new BitCastInst(DestField, 1594 PointerType::getUnqual(EltVal->getType()), 1595 "", SI); 1596 } 1597 new StoreInst(EltVal, DestField, SI); 1598 1599 if (TD->isBigEndian()) 1600 Shift -= ElementOffset; 1601 else 1602 Shift += ElementOffset; 1603 } 1604 } 1605 1606 DeadInsts.push_back(SI); 1607} 1608 1609/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1610/// an integer. Load the individual pieces to form the aggregate value. 1611void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1612 SmallVector<AllocaInst*, 32> &NewElts) { 1613 // Extract each element out of the NewElts according to its structure offset 1614 // and form the result value. 1615 const Type *AllocaEltTy = AI->getAllocatedType(); 1616 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1617 1618 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1619 << '\n'); 1620 1621 // There are two forms here: AI could be an array or struct. Both cases 1622 // have different ways to compute the element offset. 1623 const StructLayout *Layout = 0; 1624 uint64_t ArrayEltBitOffset = 0; 1625 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1626 Layout = TD->getStructLayout(EltSTy); 1627 } else { 1628 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1629 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1630 } 1631 1632 Value *ResultVal = 1633 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1634 1635 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1636 // Load the value from the alloca. If the NewElt is an aggregate, cast 1637 // the pointer to an integer of the same size before doing the load. 1638 Value *SrcField = NewElts[i]; 1639 const Type *FieldTy = 1640 cast<PointerType>(SrcField->getType())->getElementType(); 1641 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1642 1643 // Ignore zero sized fields like {}, they obviously contain no data. 1644 if (FieldSizeBits == 0) continue; 1645 1646 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1647 FieldSizeBits); 1648 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1649 !FieldTy->isVectorTy()) 1650 SrcField = new BitCastInst(SrcField, 1651 PointerType::getUnqual(FieldIntTy), 1652 "", LI); 1653 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1654 1655 // If SrcField is a fp or vector of the right size but that isn't an 1656 // integer type, bitcast to an integer so we can shift it. 1657 if (SrcField->getType() != FieldIntTy) 1658 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1659 1660 // Zero extend the field to be the same size as the final alloca so that 1661 // we can shift and insert it. 1662 if (SrcField->getType() != ResultVal->getType()) 1663 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1664 1665 // Determine the number of bits to shift SrcField. 1666 uint64_t Shift; 1667 if (Layout) // Struct case. 1668 Shift = Layout->getElementOffsetInBits(i); 1669 else // Array case. 1670 Shift = i*ArrayEltBitOffset; 1671 1672 if (TD->isBigEndian()) 1673 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1674 1675 if (Shift) { 1676 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1677 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1678 } 1679 1680 // Don't create an 'or x, 0' on the first iteration. 1681 if (!isa<Constant>(ResultVal) || 1682 !cast<Constant>(ResultVal)->isNullValue()) 1683 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1684 else 1685 ResultVal = SrcField; 1686 } 1687 1688 // Handle tail padding by truncating the result 1689 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1690 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1691 1692 LI->replaceAllUsesWith(ResultVal); 1693 DeadInsts.push_back(LI); 1694} 1695 1696/// HasPadding - Return true if the specified type has any structure or 1697/// alignment padding, false otherwise. 1698static bool HasPadding(const Type *Ty, const TargetData &TD) { 1699 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) 1700 return HasPadding(ATy->getElementType(), TD); 1701 1702 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) 1703 return HasPadding(VTy->getElementType(), TD); 1704 1705 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1706 const StructLayout *SL = TD.getStructLayout(STy); 1707 unsigned PrevFieldBitOffset = 0; 1708 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1709 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1710 1711 // Padding in sub-elements? 1712 if (HasPadding(STy->getElementType(i), TD)) 1713 return true; 1714 1715 // Check to see if there is any padding between this element and the 1716 // previous one. 1717 if (i) { 1718 unsigned PrevFieldEnd = 1719 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1720 if (PrevFieldEnd < FieldBitOffset) 1721 return true; 1722 } 1723 1724 PrevFieldBitOffset = FieldBitOffset; 1725 } 1726 1727 // Check for tail padding. 1728 if (unsigned EltCount = STy->getNumElements()) { 1729 unsigned PrevFieldEnd = PrevFieldBitOffset + 1730 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1731 if (PrevFieldEnd < SL->getSizeInBits()) 1732 return true; 1733 } 1734 } 1735 1736 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1737} 1738 1739/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1740/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1741/// or 1 if safe after canonicalization has been performed. 1742bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1743 // Loop over the use list of the alloca. We can only transform it if all of 1744 // the users are safe to transform. 1745 AllocaInfo Info; 1746 1747 isSafeForScalarRepl(AI, AI, 0, Info); 1748 if (Info.isUnsafe) { 1749 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1750 return false; 1751 } 1752 1753 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1754 // source and destination, we have to be careful. In particular, the memcpy 1755 // could be moving around elements that live in structure padding of the LLVM 1756 // types, but may actually be used. In these cases, we refuse to promote the 1757 // struct. 1758 if (Info.isMemCpySrc && Info.isMemCpyDst && 1759 HasPadding(AI->getAllocatedType(), *TD)) 1760 return false; 1761 1762 return true; 1763} 1764 1765 1766 1767/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1768/// some part of a constant global variable. This intentionally only accepts 1769/// constant expressions because we don't can't rewrite arbitrary instructions. 1770static bool PointsToConstantGlobal(Value *V) { 1771 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1772 return GV->isConstant(); 1773 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1774 if (CE->getOpcode() == Instruction::BitCast || 1775 CE->getOpcode() == Instruction::GetElementPtr) 1776 return PointsToConstantGlobal(CE->getOperand(0)); 1777 return false; 1778} 1779 1780/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1781/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1782/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1783/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1784/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1785/// the alloca, and if the source pointer is a pointer to a constant global, we 1786/// can optimize this. 1787static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 1788 bool isOffset) { 1789 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1790 User *U = cast<Instruction>(*UI); 1791 1792 if (LoadInst *LI = dyn_cast<LoadInst>(U)) 1793 // Ignore non-volatile loads, they are always ok. 1794 if (!LI->isVolatile()) 1795 continue; 1796 1797 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1798 // If uses of the bitcast are ok, we are ok. 1799 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1800 return false; 1801 continue; 1802 } 1803 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 1804 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1805 // doesn't, it does. 1806 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1807 isOffset || !GEP->hasAllZeroIndices())) 1808 return false; 1809 continue; 1810 } 1811 1812 // If this is isn't our memcpy/memmove, reject it as something we can't 1813 // handle. 1814 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 1815 if (MI == 0) 1816 return false; 1817 1818 // If we already have seen a copy, reject the second one. 1819 if (TheCopy) return false; 1820 1821 // If the pointer has been offset from the start of the alloca, we can't 1822 // safely handle this. 1823 if (isOffset) return false; 1824 1825 // If the memintrinsic isn't using the alloca as the dest, reject it. 1826 if (UI.getOperandNo() != 0) return false; 1827 1828 // If the source of the memcpy/move is not a constant global, reject it. 1829 if (!PointsToConstantGlobal(MI->getSource())) 1830 return false; 1831 1832 // Otherwise, the transform is safe. Remember the copy instruction. 1833 TheCopy = MI; 1834 } 1835 return true; 1836} 1837 1838/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1839/// modified by a copy from a constant global. If we can prove this, we can 1840/// replace any uses of the alloca with uses of the global directly. 1841MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1842 MemTransferInst *TheCopy = 0; 1843 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1844 return TheCopy; 1845 return 0; 1846} 1847