ScalarReplAggregates.cpp revision 4cc576bc5ea27951f3bb15ccefbe483293bf8eaf
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/Pass.h" 32#include "llvm/Analysis/Dominators.h" 33#include "llvm/Target/TargetData.h" 34#include "llvm/Transforms/Utils/PromoteMemToReg.h" 35#include "llvm/Transforms/Utils/Local.h" 36#include "llvm/Support/Debug.h" 37#include "llvm/Support/ErrorHandling.h" 38#include "llvm/Support/GetElementPtrTypeIterator.h" 39#include "llvm/Support/IRBuilder.h" 40#include "llvm/Support/MathExtras.h" 41#include "llvm/Support/raw_ostream.h" 42#include "llvm/ADT/SmallVector.h" 43#include "llvm/ADT/Statistic.h" 44using namespace llvm; 45 46STATISTIC(NumReplaced, "Number of allocas broken up"); 47STATISTIC(NumPromoted, "Number of allocas promoted"); 48STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 49STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 50 51namespace { 52 struct SROA : public FunctionPass { 53 static char ID; // Pass identification, replacement for typeid 54 explicit SROA(signed T = -1) : FunctionPass(&ID) { 55 if (T == -1) 56 SRThreshold = 128; 57 else 58 SRThreshold = T; 59 } 60 61 bool runOnFunction(Function &F); 62 63 bool performScalarRepl(Function &F); 64 bool performPromotion(Function &F); 65 66 // getAnalysisUsage - This pass does not require any passes, but we know it 67 // will not alter the CFG, so say so. 68 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 69 AU.addRequired<DominatorTree>(); 70 AU.addRequired<DominanceFrontier>(); 71 AU.setPreservesCFG(); 72 } 73 74 private: 75 TargetData *TD; 76 77 /// DeadInsts - Keep track of instructions we have made dead, so that 78 /// we can remove them after we are done working. 79 SmallVector<Value*, 32> DeadInsts; 80 81 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 82 /// information about the uses. All these fields are initialized to false 83 /// and set to true when something is learned. 84 struct AllocaInfo { 85 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 86 bool isUnsafe : 1; 87 88 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 89 bool isMemCpySrc : 1; 90 91 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 92 bool isMemCpyDst : 1; 93 94 AllocaInfo() 95 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false) {} 96 }; 97 98 unsigned SRThreshold; 99 100 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 101 102 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 103 104 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 105 AllocaInfo &Info); 106 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset, 107 AllocaInfo &Info); 108 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 109 const Type *MemOpType, bool isStore, AllocaInfo &Info); 110 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 111 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 112 const Type *&IdxTy); 113 114 void DoScalarReplacement(AllocaInst *AI, 115 std::vector<AllocaInst*> &WorkList); 116 void DeleteDeadInstructions(); 117 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocaInst *Base); 118 119 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 120 SmallVector<AllocaInst*, 32> &NewElts); 121 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 122 SmallVector<AllocaInst*, 32> &NewElts); 123 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 124 SmallVector<AllocaInst*, 32> &NewElts); 125 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 126 AllocaInst *AI, 127 SmallVector<AllocaInst*, 32> &NewElts); 128 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 129 SmallVector<AllocaInst*, 32> &NewElts); 130 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 131 SmallVector<AllocaInst*, 32> &NewElts); 132 133 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 134 }; 135} 136 137char SROA::ID = 0; 138static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 139 140// Public interface to the ScalarReplAggregates pass 141FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 142 return new SROA(Threshold); 143} 144 145 146//===----------------------------------------------------------------------===// 147// Convert To Scalar Optimization. 148//===----------------------------------------------------------------------===// 149 150namespace { 151/// ConvertToScalarInfo - This struct is used by CanConvertToScalar 152class ConvertToScalarInfo { 153 /// AllocaSize - The size of the alloca being considered. 154 unsigned AllocaSize; 155 const TargetData &TD; 156 157 bool IsNotTrivial; 158 const Type *VectorTy; 159 bool HadAVector; 160 161public: 162 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 163 : AllocaSize(Size), TD(td) { 164 IsNotTrivial = false; 165 VectorTy = 0; 166 HadAVector = false; 167 } 168 169 AllocaInst *TryConvert(AllocaInst *AI) { 170 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 171 // out. 172 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 173 // FIXME: In the trivial case, just use mem2reg. 174 return 0; 175 176 // If we were able to find a vector type that can handle this with 177 // insert/extract elements, and if there was at least one use that had 178 // a vector type, promote this to a vector. We don't want to promote 179 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 180 // we just get a lot of insert/extracts. If at least one vector is 181 // involved, then we probably really do have a union of vector/array. 182 const Type *NewTy; 183 if (VectorTy && VectorTy->isVectorTy() && HadAVector) { 184 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 185 << *VectorTy << '\n'); 186 NewTy = VectorTy; // Use the vector type. 187 } else { 188 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 189 // Create and insert the integer alloca. 190 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 191 } 192 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 193 ConvertUsesToScalar(AI, NewAI, 0); 194 return NewAI; 195 } 196 197private: 198 bool CanConvertToScalar(Value *V, uint64_t Offset); 199 void MergeInType(const Type *In, uint64_t Offset); 200 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 201 202 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 203 uint64_t Offset, IRBuilder<> &Builder); 204 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 205 uint64_t Offset, IRBuilder<> &Builder); 206}; 207} // end anonymous namespace. 208 209/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 210/// the offset specified by Offset (which is specified in bytes). 211/// 212/// There are two cases we handle here: 213/// 1) A union of vector types of the same size and potentially its elements. 214/// Here we turn element accesses into insert/extract element operations. 215/// This promotes a <4 x float> with a store of float to the third element 216/// into a <4 x float> that uses insert element. 217/// 2) A fully general blob of memory, which we turn into some (potentially 218/// large) integer type with extract and insert operations where the loads 219/// and stores would mutate the memory. 220void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) { 221 // Remember if we saw a vector type. 222 HadAVector |= In->isVectorTy(); 223 224 if (VectorTy && VectorTy->isVoidTy()) 225 return; 226 227 // If this could be contributing to a vector, analyze it. 228 229 // If the In type is a vector that is the same size as the alloca, see if it 230 // matches the existing VecTy. 231 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 232 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 233 // If we're storing/loading a vector of the right size, allow it as a 234 // vector. If this the first vector we see, remember the type so that 235 // we know the element size. 236 if (VectorTy == 0) 237 VectorTy = VInTy; 238 return; 239 } 240 } else if (In->isFloatTy() || In->isDoubleTy() || 241 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 242 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 243 // If we're accessing something that could be an element of a vector, see 244 // if the implied vector agrees with what we already have and if Offset is 245 // compatible with it. 246 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 247 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 248 (VectorTy == 0 || 249 cast<VectorType>(VectorTy)->getElementType() 250 ->getPrimitiveSizeInBits()/8 == EltSize)) { 251 if (VectorTy == 0) 252 VectorTy = VectorType::get(In, AllocaSize/EltSize); 253 return; 254 } 255 } 256 257 // Otherwise, we have a case that we can't handle with an optimized vector 258 // form. We can still turn this into a large integer. 259 VectorTy = Type::getVoidTy(In->getContext()); 260} 261 262/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 263/// its accesses to a single vector type, return true and set VecTy to 264/// the new type. If we could convert the alloca into a single promotable 265/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 266/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 267/// is the current offset from the base of the alloca being analyzed. 268/// 269/// If we see at least one access to the value that is as a vector type, set the 270/// SawVec flag. 271bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 272 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 273 Instruction *User = cast<Instruction>(*UI); 274 275 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 276 // Don't break volatile loads. 277 if (LI->isVolatile()) 278 return false; 279 MergeInType(LI->getType(), Offset); 280 continue; 281 } 282 283 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 284 // Storing the pointer, not into the value? 285 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 286 MergeInType(SI->getOperand(0)->getType(), Offset); 287 continue; 288 } 289 290 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 291 if (!CanConvertToScalar(BCI, Offset)) 292 return false; 293 IsNotTrivial = true; 294 continue; 295 } 296 297 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 298 // If this is a GEP with a variable indices, we can't handle it. 299 if (!GEP->hasAllConstantIndices()) 300 return false; 301 302 // Compute the offset that this GEP adds to the pointer. 303 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 304 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 305 &Indices[0], Indices.size()); 306 // See if all uses can be converted. 307 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 308 return false; 309 IsNotTrivial = true; 310 continue; 311 } 312 313 // If this is a constant sized memset of a constant value (e.g. 0) we can 314 // handle it. 315 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 316 // Store of constant value and constant size. 317 if (isa<ConstantInt>(MSI->getValue()) && 318 isa<ConstantInt>(MSI->getLength())) { 319 IsNotTrivial = true; 320 continue; 321 } 322 } 323 324 // If this is a memcpy or memmove into or out of the whole allocation, we 325 // can handle it like a load or store of the scalar type. 326 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 327 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 328 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 329 IsNotTrivial = true; 330 continue; 331 } 332 } 333 334 // Otherwise, we cannot handle this! 335 return false; 336 } 337 338 return true; 339} 340 341/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 342/// directly. This happens when we are converting an "integer union" to a 343/// single integer scalar, or when we are converting a "vector union" to a 344/// vector with insert/extractelement instructions. 345/// 346/// Offset is an offset from the original alloca, in bits that need to be 347/// shifted to the right. By the end of this, there should be no uses of Ptr. 348void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 349 uint64_t Offset) { 350 while (!Ptr->use_empty()) { 351 Instruction *User = cast<Instruction>(Ptr->use_back()); 352 353 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 354 ConvertUsesToScalar(CI, NewAI, Offset); 355 CI->eraseFromParent(); 356 continue; 357 } 358 359 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 360 // Compute the offset that this GEP adds to the pointer. 361 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 362 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 363 &Indices[0], Indices.size()); 364 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 365 GEP->eraseFromParent(); 366 continue; 367 } 368 369 IRBuilder<> Builder(User->getParent(), User); 370 371 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 372 // The load is a bit extract from NewAI shifted right by Offset bits. 373 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 374 Value *NewLoadVal 375 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 376 LI->replaceAllUsesWith(NewLoadVal); 377 LI->eraseFromParent(); 378 continue; 379 } 380 381 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 382 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 383 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 384 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 385 Builder); 386 Builder.CreateStore(New, NewAI); 387 SI->eraseFromParent(); 388 389 // If the load we just inserted is now dead, then the inserted store 390 // overwrote the entire thing. 391 if (Old->use_empty()) 392 Old->eraseFromParent(); 393 continue; 394 } 395 396 // If this is a constant sized memset of a constant value (e.g. 0) we can 397 // transform it into a store of the expanded constant value. 398 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 399 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 400 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 401 if (NumBytes != 0) { 402 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 403 404 // Compute the value replicated the right number of times. 405 APInt APVal(NumBytes*8, Val); 406 407 // Splat the value if non-zero. 408 if (Val) 409 for (unsigned i = 1; i != NumBytes; ++i) 410 APVal |= APVal << 8; 411 412 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 413 Value *New = ConvertScalar_InsertValue( 414 ConstantInt::get(User->getContext(), APVal), 415 Old, Offset, Builder); 416 Builder.CreateStore(New, NewAI); 417 418 // If the load we just inserted is now dead, then the memset overwrote 419 // the entire thing. 420 if (Old->use_empty()) 421 Old->eraseFromParent(); 422 } 423 MSI->eraseFromParent(); 424 continue; 425 } 426 427 // If this is a memcpy or memmove into or out of the whole allocation, we 428 // can handle it like a load or store of the scalar type. 429 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 430 assert(Offset == 0 && "must be store to start of alloca"); 431 432 // If the source and destination are both to the same alloca, then this is 433 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 434 // as appropriate. 435 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject(0)); 436 437 if (MTI->getSource()->getUnderlyingObject(0) != OrigAI) { 438 // Dest must be OrigAI, change this to be a load from the original 439 // pointer (bitcasted), then a store to our new alloca. 440 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 441 Value *SrcPtr = MTI->getSource(); 442 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 443 444 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 445 SrcVal->setAlignment(MTI->getAlignment()); 446 Builder.CreateStore(SrcVal, NewAI); 447 } else if (MTI->getDest()->getUnderlyingObject(0) != OrigAI) { 448 // Src must be OrigAI, change this to be a load from NewAI then a store 449 // through the original dest pointer (bitcasted). 450 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 451 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 452 453 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 454 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 455 NewStore->setAlignment(MTI->getAlignment()); 456 } else { 457 // Noop transfer. Src == Dst 458 } 459 460 MTI->eraseFromParent(); 461 continue; 462 } 463 464 llvm_unreachable("Unsupported operation!"); 465 } 466} 467 468/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 469/// or vector value FromVal, extracting the bits from the offset specified by 470/// Offset. This returns the value, which is of type ToType. 471/// 472/// This happens when we are converting an "integer union" to a single 473/// integer scalar, or when we are converting a "vector union" to a vector with 474/// insert/extractelement instructions. 475/// 476/// Offset is an offset from the original alloca, in bits that need to be 477/// shifted to the right. 478Value *ConvertToScalarInfo:: 479ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 480 uint64_t Offset, IRBuilder<> &Builder) { 481 // If the load is of the whole new alloca, no conversion is needed. 482 if (FromVal->getType() == ToType && Offset == 0) 483 return FromVal; 484 485 // If the result alloca is a vector type, this is either an element 486 // access or a bitcast to another vector type of the same size. 487 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 488 if (ToType->isVectorTy()) 489 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 490 491 // Otherwise it must be an element access. 492 unsigned Elt = 0; 493 if (Offset) { 494 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 495 Elt = Offset/EltSize; 496 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 497 } 498 // Return the element extracted out of it. 499 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 500 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 501 if (V->getType() != ToType) 502 V = Builder.CreateBitCast(V, ToType, "tmp"); 503 return V; 504 } 505 506 // If ToType is a first class aggregate, extract out each of the pieces and 507 // use insertvalue's to form the FCA. 508 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 509 const StructLayout &Layout = *TD.getStructLayout(ST); 510 Value *Res = UndefValue::get(ST); 511 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 512 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 513 Offset+Layout.getElementOffsetInBits(i), 514 Builder); 515 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 516 } 517 return Res; 518 } 519 520 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 521 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 522 Value *Res = UndefValue::get(AT); 523 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 524 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 525 Offset+i*EltSize, Builder); 526 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 527 } 528 return Res; 529 } 530 531 // Otherwise, this must be a union that was converted to an integer value. 532 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 533 534 // If this is a big-endian system and the load is narrower than the 535 // full alloca type, we need to do a shift to get the right bits. 536 int ShAmt = 0; 537 if (TD.isBigEndian()) { 538 // On big-endian machines, the lowest bit is stored at the bit offset 539 // from the pointer given by getTypeStoreSizeInBits. This matters for 540 // integers with a bitwidth that is not a multiple of 8. 541 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 542 TD.getTypeStoreSizeInBits(ToType) - Offset; 543 } else { 544 ShAmt = Offset; 545 } 546 547 // Note: we support negative bitwidths (with shl) which are not defined. 548 // We do this to support (f.e.) loads off the end of a structure where 549 // only some bits are used. 550 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 551 FromVal = Builder.CreateLShr(FromVal, 552 ConstantInt::get(FromVal->getType(), 553 ShAmt), "tmp"); 554 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 555 FromVal = Builder.CreateShl(FromVal, 556 ConstantInt::get(FromVal->getType(), 557 -ShAmt), "tmp"); 558 559 // Finally, unconditionally truncate the integer to the right width. 560 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 561 if (LIBitWidth < NTy->getBitWidth()) 562 FromVal = 563 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 564 LIBitWidth), "tmp"); 565 else if (LIBitWidth > NTy->getBitWidth()) 566 FromVal = 567 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 568 LIBitWidth), "tmp"); 569 570 // If the result is an integer, this is a trunc or bitcast. 571 if (ToType->isIntegerTy()) { 572 // Should be done. 573 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 574 // Just do a bitcast, we know the sizes match up. 575 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 576 } else { 577 // Otherwise must be a pointer. 578 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 579 } 580 assert(FromVal->getType() == ToType && "Didn't convert right?"); 581 return FromVal; 582} 583 584/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 585/// or vector value "Old" at the offset specified by Offset. 586/// 587/// This happens when we are converting an "integer union" to a 588/// single integer scalar, or when we are converting a "vector union" to a 589/// vector with insert/extractelement instructions. 590/// 591/// Offset is an offset from the original alloca, in bits that need to be 592/// shifted to the right. 593Value *ConvertToScalarInfo:: 594ConvertScalar_InsertValue(Value *SV, Value *Old, 595 uint64_t Offset, IRBuilder<> &Builder) { 596 // Convert the stored type to the actual type, shift it left to insert 597 // then 'or' into place. 598 const Type *AllocaType = Old->getType(); 599 LLVMContext &Context = Old->getContext(); 600 601 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 602 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 603 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 604 605 // Changing the whole vector with memset or with an access of a different 606 // vector type? 607 if (ValSize == VecSize) 608 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 609 610 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 611 612 // Must be an element insertion. 613 unsigned Elt = Offset/EltSize; 614 615 if (SV->getType() != VTy->getElementType()) 616 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 617 618 SV = Builder.CreateInsertElement(Old, SV, 619 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 620 "tmp"); 621 return SV; 622 } 623 624 // If SV is a first-class aggregate value, insert each value recursively. 625 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 626 const StructLayout &Layout = *TD.getStructLayout(ST); 627 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 628 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 629 Old = ConvertScalar_InsertValue(Elt, Old, 630 Offset+Layout.getElementOffsetInBits(i), 631 Builder); 632 } 633 return Old; 634 } 635 636 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 637 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 638 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 639 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 640 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 641 } 642 return Old; 643 } 644 645 // If SV is a float, convert it to the appropriate integer type. 646 // If it is a pointer, do the same. 647 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 648 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 649 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 650 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 651 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 652 SV = Builder.CreateBitCast(SV, 653 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 654 else if (SV->getType()->isPointerTy()) 655 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 656 657 // Zero extend or truncate the value if needed. 658 if (SV->getType() != AllocaType) { 659 if (SV->getType()->getPrimitiveSizeInBits() < 660 AllocaType->getPrimitiveSizeInBits()) 661 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 662 else { 663 // Truncation may be needed if storing more than the alloca can hold 664 // (undefined behavior). 665 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 666 SrcWidth = DestWidth; 667 SrcStoreWidth = DestStoreWidth; 668 } 669 } 670 671 // If this is a big-endian system and the store is narrower than the 672 // full alloca type, we need to do a shift to get the right bits. 673 int ShAmt = 0; 674 if (TD.isBigEndian()) { 675 // On big-endian machines, the lowest bit is stored at the bit offset 676 // from the pointer given by getTypeStoreSizeInBits. This matters for 677 // integers with a bitwidth that is not a multiple of 8. 678 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 679 } else { 680 ShAmt = Offset; 681 } 682 683 // Note: we support negative bitwidths (with shr) which are not defined. 684 // We do this to support (f.e.) stores off the end of a structure where 685 // only some bits in the structure are set. 686 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 687 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 688 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 689 ShAmt), "tmp"); 690 Mask <<= ShAmt; 691 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 692 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 693 -ShAmt), "tmp"); 694 Mask = Mask.lshr(-ShAmt); 695 } 696 697 // Mask out the bits we are about to insert from the old value, and or 698 // in the new bits. 699 if (SrcWidth != DestWidth) { 700 assert(DestWidth > SrcWidth); 701 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 702 SV = Builder.CreateOr(Old, SV, "ins"); 703 } 704 return SV; 705} 706 707 708//===----------------------------------------------------------------------===// 709// SRoA Driver 710//===----------------------------------------------------------------------===// 711 712 713bool SROA::runOnFunction(Function &F) { 714 TD = getAnalysisIfAvailable<TargetData>(); 715 716 bool Changed = performPromotion(F); 717 718 // FIXME: ScalarRepl currently depends on TargetData more than it 719 // theoretically needs to. It should be refactored in order to support 720 // target-independent IR. Until this is done, just skip the actual 721 // scalar-replacement portion of this pass. 722 if (!TD) return Changed; 723 724 while (1) { 725 bool LocalChange = performScalarRepl(F); 726 if (!LocalChange) break; // No need to repromote if no scalarrepl 727 Changed = true; 728 LocalChange = performPromotion(F); 729 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 730 } 731 732 return Changed; 733} 734 735 736bool SROA::performPromotion(Function &F) { 737 std::vector<AllocaInst*> Allocas; 738 DominatorTree &DT = getAnalysis<DominatorTree>(); 739 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 740 741 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 742 743 bool Changed = false; 744 745 while (1) { 746 Allocas.clear(); 747 748 // Find allocas that are safe to promote, by looking at all instructions in 749 // the entry node 750 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 751 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 752 if (isAllocaPromotable(AI)) 753 Allocas.push_back(AI); 754 755 if (Allocas.empty()) break; 756 757 PromoteMemToReg(Allocas, DT, DF); 758 NumPromoted += Allocas.size(); 759 Changed = true; 760 } 761 762 return Changed; 763} 764 765 766/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 767/// SROA. It must be a struct or array type with a small number of elements. 768static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 769 const Type *T = AI->getAllocatedType(); 770 // Do not promote any struct into more than 32 separate vars. 771 if (const StructType *ST = dyn_cast<StructType>(T)) 772 return ST->getNumElements() <= 32; 773 // Arrays are much less likely to be safe for SROA; only consider 774 // them if they are very small. 775 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 776 return AT->getNumElements() <= 8; 777 return false; 778} 779 780 781// performScalarRepl - This algorithm is a simple worklist driven algorithm, 782// which runs on all of the malloc/alloca instructions in the function, removing 783// them if they are only used by getelementptr instructions. 784// 785bool SROA::performScalarRepl(Function &F) { 786 std::vector<AllocaInst*> WorkList; 787 788 // Scan the entry basic block, adding allocas to the worklist. 789 BasicBlock &BB = F.getEntryBlock(); 790 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 791 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 792 WorkList.push_back(A); 793 794 // Process the worklist 795 bool Changed = false; 796 while (!WorkList.empty()) { 797 AllocaInst *AI = WorkList.back(); 798 WorkList.pop_back(); 799 800 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 801 // with unused elements. 802 if (AI->use_empty()) { 803 AI->eraseFromParent(); 804 Changed = true; 805 continue; 806 } 807 808 // If this alloca is impossible for us to promote, reject it early. 809 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 810 continue; 811 812 // Check to see if this allocation is only modified by a memcpy/memmove from 813 // a constant global. If this is the case, we can change all users to use 814 // the constant global instead. This is commonly produced by the CFE by 815 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 816 // is only subsequently read. 817 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 818 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 819 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 820 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 821 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 822 TheCopy->eraseFromParent(); // Don't mutate the global. 823 AI->eraseFromParent(); 824 ++NumGlobals; 825 Changed = true; 826 continue; 827 } 828 829 // Check to see if we can perform the core SROA transformation. We cannot 830 // transform the allocation instruction if it is an array allocation 831 // (allocations OF arrays are ok though), and an allocation of a scalar 832 // value cannot be decomposed at all. 833 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 834 835 // Do not promote [0 x %struct]. 836 if (AllocaSize == 0) continue; 837 838 // Do not promote any struct whose size is too big. 839 if (AllocaSize > SRThreshold) continue; 840 841 // If the alloca looks like a good candidate for scalar replacement, and if 842 // all its users can be transformed, then split up the aggregate into its 843 // separate elements. 844 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 845 DoScalarReplacement(AI, WorkList); 846 Changed = true; 847 continue; 848 } 849 850 // If we can turn this aggregate value (potentially with casts) into a 851 // simple scalar value that can be mem2reg'd into a register value. 852 // IsNotTrivial tracks whether this is something that mem2reg could have 853 // promoted itself. If so, we don't want to transform it needlessly. Note 854 // that we can't just check based on the type: the alloca may be of an i32 855 // but that has pointer arithmetic to set byte 3 of it or something. 856 if (AllocaInst *NewAI = 857 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 858 NewAI->takeName(AI); 859 AI->eraseFromParent(); 860 ++NumConverted; 861 Changed = true; 862 continue; 863 } 864 865 // Otherwise, couldn't process this alloca. 866 } 867 868 return Changed; 869} 870 871/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 872/// predicate, do SROA now. 873void SROA::DoScalarReplacement(AllocaInst *AI, 874 std::vector<AllocaInst*> &WorkList) { 875 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 876 SmallVector<AllocaInst*, 32> ElementAllocas; 877 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 878 ElementAllocas.reserve(ST->getNumContainedTypes()); 879 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 880 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 881 AI->getAlignment(), 882 AI->getName() + "." + Twine(i), AI); 883 ElementAllocas.push_back(NA); 884 WorkList.push_back(NA); // Add to worklist for recursive processing 885 } 886 } else { 887 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 888 ElementAllocas.reserve(AT->getNumElements()); 889 const Type *ElTy = AT->getElementType(); 890 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 891 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 892 AI->getName() + "." + Twine(i), AI); 893 ElementAllocas.push_back(NA); 894 WorkList.push_back(NA); // Add to worklist for recursive processing 895 } 896 } 897 898 // Now that we have created the new alloca instructions, rewrite all the 899 // uses of the old alloca. 900 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 901 902 // Now erase any instructions that were made dead while rewriting the alloca. 903 DeleteDeadInstructions(); 904 AI->eraseFromParent(); 905 906 NumReplaced++; 907} 908 909/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 910/// recursively including all their operands that become trivially dead. 911void SROA::DeleteDeadInstructions() { 912 while (!DeadInsts.empty()) { 913 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 914 915 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 916 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 917 // Zero out the operand and see if it becomes trivially dead. 918 // (But, don't add allocas to the dead instruction list -- they are 919 // already on the worklist and will be deleted separately.) 920 *OI = 0; 921 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 922 DeadInsts.push_back(U); 923 } 924 925 I->eraseFromParent(); 926 } 927} 928 929/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 930/// performing scalar replacement of alloca AI. The results are flagged in 931/// the Info parameter. Offset indicates the position within AI that is 932/// referenced by this instruction. 933void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 934 AllocaInfo &Info) { 935 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 936 Instruction *User = cast<Instruction>(*UI); 937 938 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 939 isSafeForScalarRepl(BC, AI, Offset, Info); 940 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 941 uint64_t GEPOffset = Offset; 942 isSafeGEP(GEPI, AI, GEPOffset, Info); 943 if (!Info.isUnsafe) 944 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 945 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 946 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 947 if (Length) 948 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 949 UI.getOperandNo() == 0, Info); 950 else 951 MarkUnsafe(Info); 952 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 953 if (!LI->isVolatile()) { 954 const Type *LIType = LI->getType(); 955 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 956 LIType, false, Info); 957 } else 958 MarkUnsafe(Info); 959 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 960 // Store is ok if storing INTO the pointer, not storing the pointer 961 if (!SI->isVolatile() && SI->getOperand(0) != I) { 962 const Type *SIType = SI->getOperand(0)->getType(); 963 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 964 SIType, true, Info); 965 } else 966 MarkUnsafe(Info); 967 } else { 968 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 969 MarkUnsafe(Info); 970 } 971 if (Info.isUnsafe) return; 972 } 973} 974 975/// isSafeGEP - Check if a GEP instruction can be handled for scalar 976/// replacement. It is safe when all the indices are constant, in-bounds 977/// references, and when the resulting offset corresponds to an element within 978/// the alloca type. The results are flagged in the Info parameter. Upon 979/// return, Offset is adjusted as specified by the GEP indices. 980void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 981 uint64_t &Offset, AllocaInfo &Info) { 982 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 983 if (GEPIt == E) 984 return; 985 986 // Walk through the GEP type indices, checking the types that this indexes 987 // into. 988 for (; GEPIt != E; ++GEPIt) { 989 // Ignore struct elements, no extra checking needed for these. 990 if ((*GEPIt)->isStructTy()) 991 continue; 992 993 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 994 if (!IdxVal) 995 return MarkUnsafe(Info); 996 } 997 998 // Compute the offset due to this GEP and check if the alloca has a 999 // component element at that offset. 1000 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1001 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1002 &Indices[0], Indices.size()); 1003 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 1004 MarkUnsafe(Info); 1005} 1006 1007/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1008/// alloca or has an offset and size that corresponds to a component element 1009/// within it. The offset checked here may have been formed from a GEP with a 1010/// pointer bitcasted to a different type. 1011void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 1012 const Type *MemOpType, bool isStore, 1013 AllocaInfo &Info) { 1014 // Check if this is a load/store of the entire alloca. 1015 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 1016 bool UsesAggregateType = (MemOpType == AI->getAllocatedType()); 1017 // This is safe for MemIntrinsics (where MemOpType is 0), integer types 1018 // (which are essentially the same as the MemIntrinsics, especially with 1019 // regard to copying padding between elements), or references using the 1020 // aggregate type of the alloca. 1021 if (!MemOpType || MemOpType->isIntegerTy() || UsesAggregateType) { 1022 if (!UsesAggregateType) { 1023 if (isStore) 1024 Info.isMemCpyDst = true; 1025 else 1026 Info.isMemCpySrc = true; 1027 } 1028 return; 1029 } 1030 } 1031 // Check if the offset/size correspond to a component within the alloca type. 1032 const Type *T = AI->getAllocatedType(); 1033 if (TypeHasComponent(T, Offset, MemSize)) 1034 return; 1035 1036 return MarkUnsafe(Info); 1037} 1038 1039/// TypeHasComponent - Return true if T has a component type with the 1040/// specified offset and size. If Size is zero, do not check the size. 1041bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1042 const Type *EltTy; 1043 uint64_t EltSize; 1044 if (const StructType *ST = dyn_cast<StructType>(T)) { 1045 const StructLayout *Layout = TD->getStructLayout(ST); 1046 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1047 EltTy = ST->getContainedType(EltIdx); 1048 EltSize = TD->getTypeAllocSize(EltTy); 1049 Offset -= Layout->getElementOffset(EltIdx); 1050 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1051 EltTy = AT->getElementType(); 1052 EltSize = TD->getTypeAllocSize(EltTy); 1053 if (Offset >= AT->getNumElements() * EltSize) 1054 return false; 1055 Offset %= EltSize; 1056 } else { 1057 return false; 1058 } 1059 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1060 return true; 1061 // Check if the component spans multiple elements. 1062 if (Offset + Size > EltSize) 1063 return false; 1064 return TypeHasComponent(EltTy, Offset, Size); 1065} 1066 1067/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1068/// the instruction I, which references it, to use the separate elements. 1069/// Offset indicates the position within AI that is referenced by this 1070/// instruction. 1071void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1072 SmallVector<AllocaInst*, 32> &NewElts) { 1073 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1074 Instruction *User = cast<Instruction>(*UI); 1075 1076 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1077 RewriteBitCast(BC, AI, Offset, NewElts); 1078 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1079 RewriteGEP(GEPI, AI, Offset, NewElts); 1080 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1081 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1082 uint64_t MemSize = Length->getZExtValue(); 1083 if (Offset == 0 && 1084 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1085 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1086 // Otherwise the intrinsic can only touch a single element and the 1087 // address operand will be updated, so nothing else needs to be done. 1088 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1089 const Type *LIType = LI->getType(); 1090 if (LIType == AI->getAllocatedType()) { 1091 // Replace: 1092 // %res = load { i32, i32 }* %alloc 1093 // with: 1094 // %load.0 = load i32* %alloc.0 1095 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1096 // %load.1 = load i32* %alloc.1 1097 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1098 // (Also works for arrays instead of structs) 1099 Value *Insert = UndefValue::get(LIType); 1100 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1101 Value *Load = new LoadInst(NewElts[i], "load", LI); 1102 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 1103 } 1104 LI->replaceAllUsesWith(Insert); 1105 DeadInsts.push_back(LI); 1106 } else if (LIType->isIntegerTy() && 1107 TD->getTypeAllocSize(LIType) == 1108 TD->getTypeAllocSize(AI->getAllocatedType())) { 1109 // If this is a load of the entire alloca to an integer, rewrite it. 1110 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1111 } 1112 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1113 Value *Val = SI->getOperand(0); 1114 const Type *SIType = Val->getType(); 1115 if (SIType == AI->getAllocatedType()) { 1116 // Replace: 1117 // store { i32, i32 } %val, { i32, i32 }* %alloc 1118 // with: 1119 // %val.0 = extractvalue { i32, i32 } %val, 0 1120 // store i32 %val.0, i32* %alloc.0 1121 // %val.1 = extractvalue { i32, i32 } %val, 1 1122 // store i32 %val.1, i32* %alloc.1 1123 // (Also works for arrays instead of structs) 1124 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1125 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 1126 new StoreInst(Extract, NewElts[i], SI); 1127 } 1128 DeadInsts.push_back(SI); 1129 } else if (SIType->isIntegerTy() && 1130 TD->getTypeAllocSize(SIType) == 1131 TD->getTypeAllocSize(AI->getAllocatedType())) { 1132 // If this is a store of the entire alloca from an integer, rewrite it. 1133 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1134 } 1135 } 1136 } 1137} 1138 1139/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1140/// and recursively continue updating all of its uses. 1141void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1142 SmallVector<AllocaInst*, 32> &NewElts) { 1143 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1144 if (BC->getOperand(0) != AI) 1145 return; 1146 1147 // The bitcast references the original alloca. Replace its uses with 1148 // references to the first new element alloca. 1149 Instruction *Val = NewElts[0]; 1150 if (Val->getType() != BC->getDestTy()) { 1151 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1152 Val->takeName(BC); 1153 } 1154 BC->replaceAllUsesWith(Val); 1155 DeadInsts.push_back(BC); 1156} 1157 1158/// FindElementAndOffset - Return the index of the element containing Offset 1159/// within the specified type, which must be either a struct or an array. 1160/// Sets T to the type of the element and Offset to the offset within that 1161/// element. IdxTy is set to the type of the index result to be used in a 1162/// GEP instruction. 1163uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1164 const Type *&IdxTy) { 1165 uint64_t Idx = 0; 1166 if (const StructType *ST = dyn_cast<StructType>(T)) { 1167 const StructLayout *Layout = TD->getStructLayout(ST); 1168 Idx = Layout->getElementContainingOffset(Offset); 1169 T = ST->getContainedType(Idx); 1170 Offset -= Layout->getElementOffset(Idx); 1171 IdxTy = Type::getInt32Ty(T->getContext()); 1172 return Idx; 1173 } 1174 const ArrayType *AT = cast<ArrayType>(T); 1175 T = AT->getElementType(); 1176 uint64_t EltSize = TD->getTypeAllocSize(T); 1177 Idx = Offset / EltSize; 1178 Offset -= Idx * EltSize; 1179 IdxTy = Type::getInt64Ty(T->getContext()); 1180 return Idx; 1181} 1182 1183/// RewriteGEP - Check if this GEP instruction moves the pointer across 1184/// elements of the alloca that are being split apart, and if so, rewrite 1185/// the GEP to be relative to the new element. 1186void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1187 SmallVector<AllocaInst*, 32> &NewElts) { 1188 uint64_t OldOffset = Offset; 1189 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1190 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1191 &Indices[0], Indices.size()); 1192 1193 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1194 1195 const Type *T = AI->getAllocatedType(); 1196 const Type *IdxTy; 1197 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1198 if (GEPI->getOperand(0) == AI) 1199 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1200 1201 T = AI->getAllocatedType(); 1202 uint64_t EltOffset = Offset; 1203 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1204 1205 // If this GEP does not move the pointer across elements of the alloca 1206 // being split, then it does not needs to be rewritten. 1207 if (Idx == OldIdx) 1208 return; 1209 1210 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1211 SmallVector<Value*, 8> NewArgs; 1212 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1213 while (EltOffset != 0) { 1214 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1215 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1216 } 1217 Instruction *Val = NewElts[Idx]; 1218 if (NewArgs.size() > 1) { 1219 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 1220 NewArgs.end(), "", GEPI); 1221 Val->takeName(GEPI); 1222 } 1223 if (Val->getType() != GEPI->getType()) 1224 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1225 GEPI->replaceAllUsesWith(Val); 1226 DeadInsts.push_back(GEPI); 1227} 1228 1229/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 1230/// Rewrite it to copy or set the elements of the scalarized memory. 1231void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 1232 AllocaInst *AI, 1233 SmallVector<AllocaInst*, 32> &NewElts) { 1234 // If this is a memcpy/memmove, construct the other pointer as the 1235 // appropriate type. The "Other" pointer is the pointer that goes to memory 1236 // that doesn't have anything to do with the alloca that we are promoting. For 1237 // memset, this Value* stays null. 1238 Value *OtherPtr = 0; 1239 LLVMContext &Context = MI->getContext(); 1240 unsigned MemAlignment = MI->getAlignment(); 1241 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 1242 if (Inst == MTI->getRawDest()) 1243 OtherPtr = MTI->getRawSource(); 1244 else { 1245 assert(Inst == MTI->getRawSource()); 1246 OtherPtr = MTI->getRawDest(); 1247 } 1248 } 1249 1250 // If there is an other pointer, we want to convert it to the same pointer 1251 // type as AI has, so we can GEP through it safely. 1252 if (OtherPtr) { 1253 1254 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 1255 // optimization, but it's also required to detect the corner case where 1256 // both pointer operands are referencing the same memory, and where 1257 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 1258 // function is only called for mem intrinsics that access the whole 1259 // aggregate, so non-zero GEPs are not an issue here.) 1260 while (1) { 1261 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) { 1262 OtherPtr = BC->getOperand(0); 1263 continue; 1264 } 1265 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) { 1266 // All zero GEPs are effectively bitcasts. 1267 if (GEP->hasAllZeroIndices()) { 1268 OtherPtr = GEP->getOperand(0); 1269 continue; 1270 } 1271 } 1272 break; 1273 } 1274 // Copying the alloca to itself is a no-op: just delete it. 1275 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 1276 // This code will run twice for a no-op memcpy -- once for each operand. 1277 // Put only one reference to MI on the DeadInsts list. 1278 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 1279 E = DeadInsts.end(); I != E; ++I) 1280 if (*I == MI) return; 1281 DeadInsts.push_back(MI); 1282 return; 1283 } 1284 1285 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 1286 if (BCE->getOpcode() == Instruction::BitCast) 1287 OtherPtr = BCE->getOperand(0); 1288 1289 // If the pointer is not the right type, insert a bitcast to the right 1290 // type. 1291 if (OtherPtr->getType() != AI->getType()) 1292 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 1293 MI); 1294 } 1295 1296 // Process each element of the aggregate. 1297 Value *TheFn = MI->getCalledValue(); 1298 const Type *BytePtrTy = MI->getRawDest()->getType(); 1299 bool SROADest = MI->getRawDest() == Inst; 1300 1301 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 1302 1303 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1304 // If this is a memcpy/memmove, emit a GEP of the other element address. 1305 Value *OtherElt = 0; 1306 unsigned OtherEltAlign = MemAlignment; 1307 1308 if (OtherPtr) { 1309 Value *Idx[2] = { Zero, 1310 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 1311 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 1312 OtherPtr->getName()+"."+Twine(i), 1313 MI); 1314 uint64_t EltOffset; 1315 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 1316 if (const StructType *ST = 1317 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 1318 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 1319 } else { 1320 const Type *EltTy = 1321 cast<SequentialType>(OtherPtr->getType())->getElementType(); 1322 EltOffset = TD->getTypeAllocSize(EltTy)*i; 1323 } 1324 1325 // The alignment of the other pointer is the guaranteed alignment of the 1326 // element, which is affected by both the known alignment of the whole 1327 // mem intrinsic and the alignment of the element. If the alignment of 1328 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 1329 // known alignment is just 4 bytes. 1330 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 1331 } 1332 1333 Value *EltPtr = NewElts[i]; 1334 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 1335 1336 // If we got down to a scalar, insert a load or store as appropriate. 1337 if (EltTy->isSingleValueType()) { 1338 if (isa<MemTransferInst>(MI)) { 1339 if (SROADest) { 1340 // From Other to Alloca. 1341 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 1342 new StoreInst(Elt, EltPtr, MI); 1343 } else { 1344 // From Alloca to Other. 1345 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 1346 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 1347 } 1348 continue; 1349 } 1350 assert(isa<MemSetInst>(MI)); 1351 1352 // If the stored element is zero (common case), just store a null 1353 // constant. 1354 Constant *StoreVal; 1355 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(1))) { 1356 if (CI->isZero()) { 1357 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 1358 } else { 1359 // If EltTy is a vector type, get the element type. 1360 const Type *ValTy = EltTy->getScalarType(); 1361 1362 // Construct an integer with the right value. 1363 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 1364 APInt OneVal(EltSize, CI->getZExtValue()); 1365 APInt TotalVal(OneVal); 1366 // Set each byte. 1367 for (unsigned i = 0; 8*i < EltSize; ++i) { 1368 TotalVal = TotalVal.shl(8); 1369 TotalVal |= OneVal; 1370 } 1371 1372 // Convert the integer value to the appropriate type. 1373 StoreVal = ConstantInt::get(Context, TotalVal); 1374 if (ValTy->isPointerTy()) 1375 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 1376 else if (ValTy->isFloatingPointTy()) 1377 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 1378 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 1379 1380 // If the requested value was a vector constant, create it. 1381 if (EltTy != ValTy) { 1382 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 1383 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 1384 StoreVal = ConstantVector::get(&Elts[0], NumElts); 1385 } 1386 } 1387 new StoreInst(StoreVal, EltPtr, MI); 1388 continue; 1389 } 1390 // Otherwise, if we're storing a byte variable, use a memset call for 1391 // this element. 1392 } 1393 1394 // Cast the element pointer to BytePtrTy. 1395 if (EltPtr->getType() != BytePtrTy) 1396 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI); 1397 1398 // Cast the other pointer (if we have one) to BytePtrTy. 1399 if (OtherElt && OtherElt->getType() != BytePtrTy) { 1400 // Preserve address space of OtherElt 1401 const PointerType* OtherPTy = cast<PointerType>(OtherElt->getType()); 1402 const PointerType* PTy = cast<PointerType>(BytePtrTy); 1403 if (OtherPTy->getElementType() != PTy->getElementType()) { 1404 Type *NewOtherPTy = PointerType::get(PTy->getElementType(), 1405 OtherPTy->getAddressSpace()); 1406 OtherElt = new BitCastInst(OtherElt, NewOtherPTy, 1407 OtherElt->getNameStr(), MI); 1408 } 1409 } 1410 1411 unsigned EltSize = TD->getTypeAllocSize(EltTy); 1412 1413 // Finally, insert the meminst for this element. 1414 if (isa<MemTransferInst>(MI)) { 1415 Value *Ops[] = { 1416 SROADest ? EltPtr : OtherElt, // Dest ptr 1417 SROADest ? OtherElt : EltPtr, // Src ptr 1418 ConstantInt::get(MI->getOperand(2)->getType(), EltSize), // Size 1419 // Align 1420 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign), 1421 MI->getVolatileCst() 1422 }; 1423 // In case we fold the address space overloaded memcpy of A to B 1424 // with memcpy of B to C, change the function to be a memcpy of A to C. 1425 const Type *Tys[] = { Ops[0]->getType(), Ops[1]->getType(), 1426 Ops[2]->getType() }; 1427 Module *M = MI->getParent()->getParent()->getParent(); 1428 TheFn = Intrinsic::getDeclaration(M, MI->getIntrinsicID(), Tys, 3); 1429 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1430 } else { 1431 assert(isa<MemSetInst>(MI)); 1432 Value *Ops[] = { 1433 EltPtr, MI->getOperand(1), // Dest, Value, 1434 ConstantInt::get(MI->getOperand(2)->getType(), EltSize), // Size 1435 Zero, // Align 1436 ConstantInt::get(Type::getInt1Ty(MI->getContext()), 0) // isVolatile 1437 }; 1438 const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() }; 1439 Module *M = MI->getParent()->getParent()->getParent(); 1440 TheFn = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2); 1441 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 1442 } 1443 } 1444 DeadInsts.push_back(MI); 1445} 1446 1447/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 1448/// overwrites the entire allocation. Extract out the pieces of the stored 1449/// integer and store them individually. 1450void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 1451 SmallVector<AllocaInst*, 32> &NewElts){ 1452 // Extract each element out of the integer according to its structure offset 1453 // and store the element value to the individual alloca. 1454 Value *SrcVal = SI->getOperand(0); 1455 const Type *AllocaEltTy = AI->getAllocatedType(); 1456 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1457 1458 // Handle tail padding by extending the operand 1459 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 1460 SrcVal = new ZExtInst(SrcVal, 1461 IntegerType::get(SI->getContext(), AllocaSizeBits), 1462 "", SI); 1463 1464 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 1465 << '\n'); 1466 1467 // There are two forms here: AI could be an array or struct. Both cases 1468 // have different ways to compute the element offset. 1469 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1470 const StructLayout *Layout = TD->getStructLayout(EltSTy); 1471 1472 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1473 // Get the number of bits to shift SrcVal to get the value. 1474 const Type *FieldTy = EltSTy->getElementType(i); 1475 uint64_t Shift = Layout->getElementOffsetInBits(i); 1476 1477 if (TD->isBigEndian()) 1478 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 1479 1480 Value *EltVal = SrcVal; 1481 if (Shift) { 1482 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1483 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1484 "sroa.store.elt", SI); 1485 } 1486 1487 // Truncate down to an integer of the right size. 1488 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1489 1490 // Ignore zero sized fields like {}, they obviously contain no data. 1491 if (FieldSizeBits == 0) continue; 1492 1493 if (FieldSizeBits != AllocaSizeBits) 1494 EltVal = new TruncInst(EltVal, 1495 IntegerType::get(SI->getContext(), FieldSizeBits), 1496 "", SI); 1497 Value *DestField = NewElts[i]; 1498 if (EltVal->getType() == FieldTy) { 1499 // Storing to an integer field of this size, just do it. 1500 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 1501 // Bitcast to the right element type (for fp/vector values). 1502 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 1503 } else { 1504 // Otherwise, bitcast the dest pointer (for aggregates). 1505 DestField = new BitCastInst(DestField, 1506 PointerType::getUnqual(EltVal->getType()), 1507 "", SI); 1508 } 1509 new StoreInst(EltVal, DestField, SI); 1510 } 1511 1512 } else { 1513 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1514 const Type *ArrayEltTy = ATy->getElementType(); 1515 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1516 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1517 1518 uint64_t Shift; 1519 1520 if (TD->isBigEndian()) 1521 Shift = AllocaSizeBits-ElementOffset; 1522 else 1523 Shift = 0; 1524 1525 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1526 // Ignore zero sized fields like {}, they obviously contain no data. 1527 if (ElementSizeBits == 0) continue; 1528 1529 Value *EltVal = SrcVal; 1530 if (Shift) { 1531 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1532 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1533 "sroa.store.elt", SI); 1534 } 1535 1536 // Truncate down to an integer of the right size. 1537 if (ElementSizeBits != AllocaSizeBits) 1538 EltVal = new TruncInst(EltVal, 1539 IntegerType::get(SI->getContext(), 1540 ElementSizeBits),"",SI); 1541 Value *DestField = NewElts[i]; 1542 if (EltVal->getType() == ArrayEltTy) { 1543 // Storing to an integer field of this size, just do it. 1544 } else if (ArrayEltTy->isFloatingPointTy() || 1545 ArrayEltTy->isVectorTy()) { 1546 // Bitcast to the right element type (for fp/vector values). 1547 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1548 } else { 1549 // Otherwise, bitcast the dest pointer (for aggregates). 1550 DestField = new BitCastInst(DestField, 1551 PointerType::getUnqual(EltVal->getType()), 1552 "", SI); 1553 } 1554 new StoreInst(EltVal, DestField, SI); 1555 1556 if (TD->isBigEndian()) 1557 Shift -= ElementOffset; 1558 else 1559 Shift += ElementOffset; 1560 } 1561 } 1562 1563 DeadInsts.push_back(SI); 1564} 1565 1566/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1567/// an integer. Load the individual pieces to form the aggregate value. 1568void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1569 SmallVector<AllocaInst*, 32> &NewElts) { 1570 // Extract each element out of the NewElts according to its structure offset 1571 // and form the result value. 1572 const Type *AllocaEltTy = AI->getAllocatedType(); 1573 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1574 1575 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1576 << '\n'); 1577 1578 // There are two forms here: AI could be an array or struct. Both cases 1579 // have different ways to compute the element offset. 1580 const StructLayout *Layout = 0; 1581 uint64_t ArrayEltBitOffset = 0; 1582 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1583 Layout = TD->getStructLayout(EltSTy); 1584 } else { 1585 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1586 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1587 } 1588 1589 Value *ResultVal = 1590 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1591 1592 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1593 // Load the value from the alloca. If the NewElt is an aggregate, cast 1594 // the pointer to an integer of the same size before doing the load. 1595 Value *SrcField = NewElts[i]; 1596 const Type *FieldTy = 1597 cast<PointerType>(SrcField->getType())->getElementType(); 1598 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1599 1600 // Ignore zero sized fields like {}, they obviously contain no data. 1601 if (FieldSizeBits == 0) continue; 1602 1603 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1604 FieldSizeBits); 1605 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1606 !FieldTy->isVectorTy()) 1607 SrcField = new BitCastInst(SrcField, 1608 PointerType::getUnqual(FieldIntTy), 1609 "", LI); 1610 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1611 1612 // If SrcField is a fp or vector of the right size but that isn't an 1613 // integer type, bitcast to an integer so we can shift it. 1614 if (SrcField->getType() != FieldIntTy) 1615 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1616 1617 // Zero extend the field to be the same size as the final alloca so that 1618 // we can shift and insert it. 1619 if (SrcField->getType() != ResultVal->getType()) 1620 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1621 1622 // Determine the number of bits to shift SrcField. 1623 uint64_t Shift; 1624 if (Layout) // Struct case. 1625 Shift = Layout->getElementOffsetInBits(i); 1626 else // Array case. 1627 Shift = i*ArrayEltBitOffset; 1628 1629 if (TD->isBigEndian()) 1630 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1631 1632 if (Shift) { 1633 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1634 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1635 } 1636 1637 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1638 } 1639 1640 // Handle tail padding by truncating the result 1641 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1642 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1643 1644 LI->replaceAllUsesWith(ResultVal); 1645 DeadInsts.push_back(LI); 1646} 1647 1648/// HasPadding - Return true if the specified type has any structure or 1649/// alignment padding, false otherwise. 1650static bool HasPadding(const Type *Ty, const TargetData &TD) { 1651 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1652 const StructLayout *SL = TD.getStructLayout(STy); 1653 unsigned PrevFieldBitOffset = 0; 1654 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1655 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1656 1657 // Padding in sub-elements? 1658 if (HasPadding(STy->getElementType(i), TD)) 1659 return true; 1660 1661 // Check to see if there is any padding between this element and the 1662 // previous one. 1663 if (i) { 1664 unsigned PrevFieldEnd = 1665 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1666 if (PrevFieldEnd < FieldBitOffset) 1667 return true; 1668 } 1669 1670 PrevFieldBitOffset = FieldBitOffset; 1671 } 1672 1673 // Check for tail padding. 1674 if (unsigned EltCount = STy->getNumElements()) { 1675 unsigned PrevFieldEnd = PrevFieldBitOffset + 1676 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1677 if (PrevFieldEnd < SL->getSizeInBits()) 1678 return true; 1679 } 1680 1681 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1682 return HasPadding(ATy->getElementType(), TD); 1683 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1684 return HasPadding(VTy->getElementType(), TD); 1685 } 1686 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1687} 1688 1689/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1690/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1691/// or 1 if safe after canonicalization has been performed. 1692bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1693 // Loop over the use list of the alloca. We can only transform it if all of 1694 // the users are safe to transform. 1695 AllocaInfo Info; 1696 1697 isSafeForScalarRepl(AI, AI, 0, Info); 1698 if (Info.isUnsafe) { 1699 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1700 return false; 1701 } 1702 1703 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1704 // source and destination, we have to be careful. In particular, the memcpy 1705 // could be moving around elements that live in structure padding of the LLVM 1706 // types, but may actually be used. In these cases, we refuse to promote the 1707 // struct. 1708 if (Info.isMemCpySrc && Info.isMemCpyDst && 1709 HasPadding(AI->getAllocatedType(), *TD)) 1710 return false; 1711 1712 return true; 1713} 1714 1715 1716 1717/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1718/// some part of a constant global variable. This intentionally only accepts 1719/// constant expressions because we don't can't rewrite arbitrary instructions. 1720static bool PointsToConstantGlobal(Value *V) { 1721 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1722 return GV->isConstant(); 1723 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1724 if (CE->getOpcode() == Instruction::BitCast || 1725 CE->getOpcode() == Instruction::GetElementPtr) 1726 return PointsToConstantGlobal(CE->getOperand(0)); 1727 return false; 1728} 1729 1730/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1731/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1732/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1733/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1734/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1735/// the alloca, and if the source pointer is a pointer to a constant global, we 1736/// can optimize this. 1737static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 1738 bool isOffset) { 1739 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1740 User *U = cast<Instruction>(*UI); 1741 1742 if (LoadInst *LI = dyn_cast<LoadInst>(U)) 1743 // Ignore non-volatile loads, they are always ok. 1744 if (!LI->isVolatile()) 1745 continue; 1746 1747 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1748 // If uses of the bitcast are ok, we are ok. 1749 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1750 return false; 1751 continue; 1752 } 1753 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 1754 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1755 // doesn't, it does. 1756 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1757 isOffset || !GEP->hasAllZeroIndices())) 1758 return false; 1759 continue; 1760 } 1761 1762 // If this is isn't our memcpy/memmove, reject it as something we can't 1763 // handle. 1764 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 1765 if (MI == 0) 1766 return false; 1767 1768 // If we already have seen a copy, reject the second one. 1769 if (TheCopy) return false; 1770 1771 // If the pointer has been offset from the start of the alloca, we can't 1772 // safely handle this. 1773 if (isOffset) return false; 1774 1775 // If the memintrinsic isn't using the alloca as the dest, reject it. 1776 if (UI.getOperandNo() != 0) return false; 1777 1778 // If the source of the memcpy/move is not a constant global, reject it. 1779 if (!PointsToConstantGlobal(MI->getSource())) 1780 return false; 1781 1782 // Otherwise, the transform is safe. Remember the copy instruction. 1783 TheCopy = MI; 1784 } 1785 return true; 1786} 1787 1788/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1789/// modified by a copy from a constant global. If we can prove this, we can 1790/// replace any uses of the alloca with uses of the global directly. 1791MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1792 MemTransferInst *TheCopy = 0; 1793 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1794 return TheCopy; 1795 return 0; 1796} 1797