ScalarReplAggregates.cpp revision 90747e34e6ca7162eaf8dde032649071045f161d
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/DebugInfo.h" 34#include "llvm/Analysis/DIBuilder.h" 35#include "llvm/Analysis/Dominators.h" 36#include "llvm/Analysis/Loads.h" 37#include "llvm/Analysis/ValueTracking.h" 38#include "llvm/Target/TargetData.h" 39#include "llvm/Transforms/Utils/PromoteMemToReg.h" 40#include "llvm/Transforms/Utils/Local.h" 41#include "llvm/Transforms/Utils/SSAUpdater.h" 42#include "llvm/Support/CallSite.h" 43#include "llvm/Support/Debug.h" 44#include "llvm/Support/ErrorHandling.h" 45#include "llvm/Support/GetElementPtrTypeIterator.h" 46#include "llvm/Support/IRBuilder.h" 47#include "llvm/Support/MathExtras.h" 48#include "llvm/Support/raw_ostream.h" 49#include "llvm/ADT/SetVector.h" 50#include "llvm/ADT/SmallVector.h" 51#include "llvm/ADT/Statistic.h" 52using namespace llvm; 53 54STATISTIC(NumReplaced, "Number of allocas broken up"); 55STATISTIC(NumPromoted, "Number of allocas promoted"); 56STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); 57STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 58STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 59 60namespace { 61 struct SROA : public FunctionPass { 62 SROA(int T, bool hasDT, char &ID) 63 : FunctionPass(ID), HasDomTree(hasDT) { 64 if (T == -1) 65 SRThreshold = 128; 66 else 67 SRThreshold = T; 68 } 69 70 bool runOnFunction(Function &F); 71 72 bool performScalarRepl(Function &F); 73 bool performPromotion(Function &F); 74 75 private: 76 bool HasDomTree; 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 /// The alloca to promote. 88 AllocaInst *AI; 89 90 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite 91 /// looping and avoid redundant work. 92 SmallPtrSet<PHINode*, 8> CheckedPHIs; 93 94 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 95 bool isUnsafe : 1; 96 97 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 98 bool isMemCpySrc : 1; 99 100 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 101 bool isMemCpyDst : 1; 102 103 /// hasSubelementAccess - This is true if a subelement of the alloca is 104 /// ever accessed, or false if the alloca is only accessed with mem 105 /// intrinsics or load/store that only access the entire alloca at once. 106 bool hasSubelementAccess : 1; 107 108 /// hasALoadOrStore - This is true if there are any loads or stores to it. 109 /// The alloca may just be accessed with memcpy, for example, which would 110 /// not set this. 111 bool hasALoadOrStore : 1; 112 113 explicit AllocaInfo(AllocaInst *ai) 114 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 115 hasSubelementAccess(false), hasALoadOrStore(false) {} 116 }; 117 118 unsigned SRThreshold; 119 120 void MarkUnsafe(AllocaInfo &I, Instruction *User) { 121 I.isUnsafe = true; 122 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); 123 } 124 125 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 126 127 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info); 128 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset, 129 AllocaInfo &Info); 130 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info); 131 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 132 Type *MemOpType, bool isStore, AllocaInfo &Info, 133 Instruction *TheAccess, bool AllowWholeAccess); 134 bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size); 135 uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset, 136 Type *&IdxTy); 137 138 void DoScalarReplacement(AllocaInst *AI, 139 std::vector<AllocaInst*> &WorkList); 140 void DeleteDeadInstructions(); 141 142 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 143 SmallVector<AllocaInst*, 32> &NewElts); 144 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 145 SmallVector<AllocaInst*, 32> &NewElts); 146 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 147 SmallVector<AllocaInst*, 32> &NewElts); 148 void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 149 uint64_t Offset, 150 SmallVector<AllocaInst*, 32> &NewElts); 151 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 152 AllocaInst *AI, 153 SmallVector<AllocaInst*, 32> &NewElts); 154 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 155 SmallVector<AllocaInst*, 32> &NewElts); 156 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 157 SmallVector<AllocaInst*, 32> &NewElts); 158 159 static MemTransferInst *isOnlyCopiedFromConstantGlobal( 160 AllocaInst *AI, SmallVector<Instruction*, 4> &ToDelete); 161 }; 162 163 // SROA_DT - SROA that uses DominatorTree. 164 struct SROA_DT : public SROA { 165 static char ID; 166 public: 167 SROA_DT(int T = -1) : SROA(T, true, ID) { 168 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 169 } 170 171 // getAnalysisUsage - This pass does not require any passes, but we know it 172 // will not alter the CFG, so say so. 173 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 174 AU.addRequired<DominatorTree>(); 175 AU.setPreservesCFG(); 176 } 177 }; 178 179 // SROA_SSAUp - SROA that uses SSAUpdater. 180 struct SROA_SSAUp : public SROA { 181 static char ID; 182 public: 183 SROA_SSAUp(int T = -1) : SROA(T, false, ID) { 184 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 185 } 186 187 // getAnalysisUsage - This pass does not require any passes, but we know it 188 // will not alter the CFG, so say so. 189 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 190 AU.setPreservesCFG(); 191 } 192 }; 193 194} 195 196char SROA_DT::ID = 0; 197char SROA_SSAUp::ID = 0; 198 199INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 200 "Scalar Replacement of Aggregates (DT)", false, false) 201INITIALIZE_PASS_DEPENDENCY(DominatorTree) 202INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 203 "Scalar Replacement of Aggregates (DT)", false, false) 204 205INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 206 "Scalar Replacement of Aggregates (SSAUp)", false, false) 207INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 208 "Scalar Replacement of Aggregates (SSAUp)", false, false) 209 210// Public interface to the ScalarReplAggregates pass 211FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 212 bool UseDomTree) { 213 if (UseDomTree) 214 return new SROA_DT(Threshold); 215 return new SROA_SSAUp(Threshold); 216} 217 218 219//===----------------------------------------------------------------------===// 220// Convert To Scalar Optimization. 221//===----------------------------------------------------------------------===// 222 223namespace { 224/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 225/// optimization, which scans the uses of an alloca and determines if it can 226/// rewrite it in terms of a single new alloca that can be mem2reg'd. 227class ConvertToScalarInfo { 228 /// AllocaSize - The size of the alloca being considered in bytes. 229 unsigned AllocaSize; 230 const TargetData &TD; 231 232 /// IsNotTrivial - This is set to true if there is some access to the object 233 /// which means that mem2reg can't promote it. 234 bool IsNotTrivial; 235 236 /// ScalarKind - Tracks the kind of alloca being considered for promotion, 237 /// computed based on the uses of the alloca rather than the LLVM type system. 238 enum { 239 Unknown, 240 241 // Accesses via GEPs that are consistent with element access of a vector 242 // type. This will not be converted into a vector unless there is a later 243 // access using an actual vector type. 244 ImplicitVector, 245 246 // Accesses via vector operations and GEPs that are consistent with the 247 // layout of a vector type. 248 Vector, 249 250 // An integer bag-of-bits with bitwise operations for insertion and 251 // extraction. Any combination of types can be converted into this kind 252 // of scalar. 253 Integer 254 } ScalarKind; 255 256 /// VectorTy - This tracks the type that we should promote the vector to if 257 /// it is possible to turn it into a vector. This starts out null, and if it 258 /// isn't possible to turn into a vector type, it gets set to VoidTy. 259 VectorType *VectorTy; 260 261 /// HadNonMemTransferAccess - True if there is at least one access to the 262 /// alloca that is not a MemTransferInst. We don't want to turn structs into 263 /// large integers unless there is some potential for optimization. 264 bool HadNonMemTransferAccess; 265 266public: 267 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 268 : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown), 269 VectorTy(0), HadNonMemTransferAccess(false) { } 270 271 AllocaInst *TryConvert(AllocaInst *AI); 272 273private: 274 bool CanConvertToScalar(Value *V, uint64_t Offset); 275 void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset); 276 bool MergeInVectorType(VectorType *VInTy, uint64_t Offset); 277 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 278 279 Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType, 280 uint64_t Offset, IRBuilder<> &Builder); 281 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 282 uint64_t Offset, IRBuilder<> &Builder); 283}; 284} // end anonymous namespace. 285 286 287/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 288/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 289/// alloca if possible or null if not. 290AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 291 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 292 // out. 293 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 294 return 0; 295 296 // If an alloca has only memset / memcpy uses, it may still have an Unknown 297 // ScalarKind. Treat it as an Integer below. 298 if (ScalarKind == Unknown) 299 ScalarKind = Integer; 300 301 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8) 302 ScalarKind = Integer; 303 304 // If we were able to find a vector type that can handle this with 305 // insert/extract elements, and if there was at least one use that had 306 // a vector type, promote this to a vector. We don't want to promote 307 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 308 // we just get a lot of insert/extracts. If at least one vector is 309 // involved, then we probably really do have a union of vector/array. 310 Type *NewTy; 311 if (ScalarKind == Vector) { 312 assert(VectorTy && "Missing type for vector scalar."); 313 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 314 << *VectorTy << '\n'); 315 NewTy = VectorTy; // Use the vector type. 316 } else { 317 unsigned BitWidth = AllocaSize * 8; 318 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) && 319 !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth)) 320 return 0; 321 322 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 323 // Create and insert the integer alloca. 324 NewTy = IntegerType::get(AI->getContext(), BitWidth); 325 } 326 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 327 ConvertUsesToScalar(AI, NewAI, 0); 328 return NewAI; 329} 330 331/// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type 332/// (VectorTy) so far at the offset specified by Offset (which is specified in 333/// bytes). 334/// 335/// There are two cases we handle here: 336/// 1) A union of vector types of the same size and potentially its elements. 337/// Here we turn element accesses into insert/extract element operations. 338/// This promotes a <4 x float> with a store of float to the third element 339/// into a <4 x float> that uses insert element. 340/// 2) A fully general blob of memory, which we turn into some (potentially 341/// large) integer type with extract and insert operations where the loads 342/// and stores would mutate the memory. We mark this by setting VectorTy 343/// to VoidTy. 344void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In, 345 uint64_t Offset) { 346 // If we already decided to turn this into a blob of integer memory, there is 347 // nothing to be done. 348 if (ScalarKind == Integer) 349 return; 350 351 // If this could be contributing to a vector, analyze it. 352 353 // If the In type is a vector that is the same size as the alloca, see if it 354 // matches the existing VecTy. 355 if (VectorType *VInTy = dyn_cast<VectorType>(In)) { 356 if (MergeInVectorType(VInTy, Offset)) 357 return; 358 } else if (In->isFloatTy() || In->isDoubleTy() || 359 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 360 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 361 // Full width accesses can be ignored, because they can always be turned 362 // into bitcasts. 363 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 364 if (EltSize == AllocaSize) 365 return; 366 367 // If we're accessing something that could be an element of a vector, see 368 // if the implied vector agrees with what we already have and if Offset is 369 // compatible with it. 370 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 371 (!VectorTy || EltSize == VectorTy->getElementType() 372 ->getPrimitiveSizeInBits()/8)) { 373 if (!VectorTy) { 374 ScalarKind = ImplicitVector; 375 VectorTy = VectorType::get(In, AllocaSize/EltSize); 376 } 377 return; 378 } 379 } 380 381 // Otherwise, we have a case that we can't handle with an optimized vector 382 // form. We can still turn this into a large integer. 383 ScalarKind = Integer; 384} 385 386/// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore, 387/// returning true if the type was successfully merged and false otherwise. 388bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy, 389 uint64_t Offset) { 390 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 391 // If we're storing/loading a vector of the right size, allow it as a 392 // vector. If this the first vector we see, remember the type so that 393 // we know the element size. If this is a subsequent access, ignore it 394 // even if it is a differing type but the same size. Worst case we can 395 // bitcast the resultant vectors. 396 if (!VectorTy) 397 VectorTy = VInTy; 398 ScalarKind = Vector; 399 return true; 400 } 401 402 return false; 403} 404 405/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 406/// its accesses to a single vector type, return true and set VecTy to 407/// the new type. If we could convert the alloca into a single promotable 408/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 409/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 410/// is the current offset from the base of the alloca being analyzed. 411/// 412/// If we see at least one access to the value that is as a vector type, set the 413/// SawVec flag. 414bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 415 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 416 Instruction *User = cast<Instruction>(*UI); 417 418 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 419 // Don't break volatile loads. 420 if (!LI->isSimple()) 421 return false; 422 // Don't touch MMX operations. 423 if (LI->getType()->isX86_MMXTy()) 424 return false; 425 HadNonMemTransferAccess = true; 426 MergeInTypeForLoadOrStore(LI->getType(), Offset); 427 continue; 428 } 429 430 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 431 // Storing the pointer, not into the value? 432 if (SI->getOperand(0) == V || !SI->isSimple()) return false; 433 // Don't touch MMX operations. 434 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 435 return false; 436 HadNonMemTransferAccess = true; 437 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset); 438 continue; 439 } 440 441 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 442 if (!onlyUsedByLifetimeMarkers(BCI)) 443 IsNotTrivial = true; // Can't be mem2reg'd. 444 if (!CanConvertToScalar(BCI, Offset)) 445 return false; 446 continue; 447 } 448 449 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 450 // If this is a GEP with a variable indices, we can't handle it. 451 if (!GEP->hasAllConstantIndices()) 452 return false; 453 454 // Compute the offset that this GEP adds to the pointer. 455 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 456 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 457 Indices); 458 // See if all uses can be converted. 459 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 460 return false; 461 IsNotTrivial = true; // Can't be mem2reg'd. 462 HadNonMemTransferAccess = true; 463 continue; 464 } 465 466 // If this is a constant sized memset of a constant value (e.g. 0) we can 467 // handle it. 468 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 469 // Store of constant value. 470 if (!isa<ConstantInt>(MSI->getValue())) 471 return false; 472 473 // Store of constant size. 474 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength()); 475 if (!Len) 476 return false; 477 478 // If the size differs from the alloca, we can only convert the alloca to 479 // an integer bag-of-bits. 480 // FIXME: This should handle all of the cases that are currently accepted 481 // as vector element insertions. 482 if (Len->getZExtValue() != AllocaSize || Offset != 0) 483 ScalarKind = Integer; 484 485 IsNotTrivial = true; // Can't be mem2reg'd. 486 HadNonMemTransferAccess = true; 487 continue; 488 } 489 490 // If this is a memcpy or memmove into or out of the whole allocation, we 491 // can handle it like a load or store of the scalar type. 492 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 493 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 494 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 495 return false; 496 497 IsNotTrivial = true; // Can't be mem2reg'd. 498 continue; 499 } 500 501 // If this is a lifetime intrinsic, we can handle it. 502 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 503 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 504 II->getIntrinsicID() == Intrinsic::lifetime_end) { 505 continue; 506 } 507 } 508 509 // Otherwise, we cannot handle this! 510 return false; 511 } 512 513 return true; 514} 515 516/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 517/// directly. This happens when we are converting an "integer union" to a 518/// single integer scalar, or when we are converting a "vector union" to a 519/// vector with insert/extractelement instructions. 520/// 521/// Offset is an offset from the original alloca, in bits that need to be 522/// shifted to the right. By the end of this, there should be no uses of Ptr. 523void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 524 uint64_t Offset) { 525 while (!Ptr->use_empty()) { 526 Instruction *User = cast<Instruction>(Ptr->use_back()); 527 528 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 529 ConvertUsesToScalar(CI, NewAI, Offset); 530 CI->eraseFromParent(); 531 continue; 532 } 533 534 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 535 // Compute the offset that this GEP adds to the pointer. 536 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 537 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 538 Indices); 539 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 540 GEP->eraseFromParent(); 541 continue; 542 } 543 544 IRBuilder<> Builder(User); 545 546 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 547 // The load is a bit extract from NewAI shifted right by Offset bits. 548 Value *LoadedVal = Builder.CreateLoad(NewAI); 549 Value *NewLoadVal 550 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 551 LI->replaceAllUsesWith(NewLoadVal); 552 LI->eraseFromParent(); 553 continue; 554 } 555 556 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 557 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 558 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 559 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 560 Builder); 561 Builder.CreateStore(New, NewAI); 562 SI->eraseFromParent(); 563 564 // If the load we just inserted is now dead, then the inserted store 565 // overwrote the entire thing. 566 if (Old->use_empty()) 567 Old->eraseFromParent(); 568 continue; 569 } 570 571 // If this is a constant sized memset of a constant value (e.g. 0) we can 572 // transform it into a store of the expanded constant value. 573 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 574 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 575 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 576 if (NumBytes != 0) { 577 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 578 579 // Compute the value replicated the right number of times. 580 APInt APVal(NumBytes*8, Val); 581 582 // Splat the value if non-zero. 583 if (Val) 584 for (unsigned i = 1; i != NumBytes; ++i) 585 APVal |= APVal << 8; 586 587 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 588 Value *New = ConvertScalar_InsertValue( 589 ConstantInt::get(User->getContext(), APVal), 590 Old, Offset, Builder); 591 Builder.CreateStore(New, NewAI); 592 593 // If the load we just inserted is now dead, then the memset overwrote 594 // the entire thing. 595 if (Old->use_empty()) 596 Old->eraseFromParent(); 597 } 598 MSI->eraseFromParent(); 599 continue; 600 } 601 602 // If this is a memcpy or memmove into or out of the whole allocation, we 603 // can handle it like a load or store of the scalar type. 604 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 605 assert(Offset == 0 && "must be store to start of alloca"); 606 607 // If the source and destination are both to the same alloca, then this is 608 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 609 // as appropriate. 610 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 611 612 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 613 // Dest must be OrigAI, change this to be a load from the original 614 // pointer (bitcasted), then a store to our new alloca. 615 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 616 Value *SrcPtr = MTI->getSource(); 617 PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 618 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 619 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 620 AIPTy = PointerType::get(AIPTy->getElementType(), 621 SPTy->getAddressSpace()); 622 } 623 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 624 625 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 626 SrcVal->setAlignment(MTI->getAlignment()); 627 Builder.CreateStore(SrcVal, NewAI); 628 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 629 // Src must be OrigAI, change this to be a load from NewAI then a store 630 // through the original dest pointer (bitcasted). 631 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 632 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 633 634 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 635 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 636 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 637 AIPTy = PointerType::get(AIPTy->getElementType(), 638 DPTy->getAddressSpace()); 639 } 640 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 641 642 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 643 NewStore->setAlignment(MTI->getAlignment()); 644 } else { 645 // Noop transfer. Src == Dst 646 } 647 648 MTI->eraseFromParent(); 649 continue; 650 } 651 652 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 653 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 654 II->getIntrinsicID() == Intrinsic::lifetime_end) { 655 // There's no need to preserve these, as the resulting alloca will be 656 // converted to a register anyways. 657 II->eraseFromParent(); 658 continue; 659 } 660 } 661 662 llvm_unreachable("Unsupported operation!"); 663 } 664} 665 666/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 667/// or vector value FromVal, extracting the bits from the offset specified by 668/// Offset. This returns the value, which is of type ToType. 669/// 670/// This happens when we are converting an "integer union" to a single 671/// integer scalar, or when we are converting a "vector union" to a vector with 672/// insert/extractelement instructions. 673/// 674/// Offset is an offset from the original alloca, in bits that need to be 675/// shifted to the right. 676Value *ConvertToScalarInfo:: 677ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, 678 uint64_t Offset, IRBuilder<> &Builder) { 679 // If the load is of the whole new alloca, no conversion is needed. 680 Type *FromType = FromVal->getType(); 681 if (FromType == ToType && Offset == 0) 682 return FromVal; 683 684 // If the result alloca is a vector type, this is either an element 685 // access or a bitcast to another vector type of the same size. 686 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) { 687 unsigned FromTypeSize = TD.getTypeAllocSize(FromType); 688 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 689 if (FromTypeSize == ToTypeSize) 690 return Builder.CreateBitCast(FromVal, ToType); 691 692 // Otherwise it must be an element access. 693 unsigned Elt = 0; 694 if (Offset) { 695 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 696 Elt = Offset/EltSize; 697 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 698 } 699 // Return the element extracted out of it. 700 Value *V = Builder.CreateExtractElement(FromVal, Builder.getInt32(Elt)); 701 if (V->getType() != ToType) 702 V = Builder.CreateBitCast(V, ToType); 703 return V; 704 } 705 706 // If ToType is a first class aggregate, extract out each of the pieces and 707 // use insertvalue's to form the FCA. 708 if (StructType *ST = dyn_cast<StructType>(ToType)) { 709 const StructLayout &Layout = *TD.getStructLayout(ST); 710 Value *Res = UndefValue::get(ST); 711 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 712 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 713 Offset+Layout.getElementOffsetInBits(i), 714 Builder); 715 Res = Builder.CreateInsertValue(Res, Elt, i); 716 } 717 return Res; 718 } 719 720 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 721 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 722 Value *Res = UndefValue::get(AT); 723 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 724 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 725 Offset+i*EltSize, Builder); 726 Res = Builder.CreateInsertValue(Res, Elt, i); 727 } 728 return Res; 729 } 730 731 // Otherwise, this must be a union that was converted to an integer value. 732 IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 733 734 // If this is a big-endian system and the load is narrower than the 735 // full alloca type, we need to do a shift to get the right bits. 736 int ShAmt = 0; 737 if (TD.isBigEndian()) { 738 // On big-endian machines, the lowest bit is stored at the bit offset 739 // from the pointer given by getTypeStoreSizeInBits. This matters for 740 // integers with a bitwidth that is not a multiple of 8. 741 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 742 TD.getTypeStoreSizeInBits(ToType) - Offset; 743 } else { 744 ShAmt = Offset; 745 } 746 747 // Note: we support negative bitwidths (with shl) which are not defined. 748 // We do this to support (f.e.) loads off the end of a structure where 749 // only some bits are used. 750 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 751 FromVal = Builder.CreateLShr(FromVal, 752 ConstantInt::get(FromVal->getType(), ShAmt)); 753 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 754 FromVal = Builder.CreateShl(FromVal, 755 ConstantInt::get(FromVal->getType(), -ShAmt)); 756 757 // Finally, unconditionally truncate the integer to the right width. 758 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 759 if (LIBitWidth < NTy->getBitWidth()) 760 FromVal = 761 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 762 LIBitWidth)); 763 else if (LIBitWidth > NTy->getBitWidth()) 764 FromVal = 765 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 766 LIBitWidth)); 767 768 // If the result is an integer, this is a trunc or bitcast. 769 if (ToType->isIntegerTy()) { 770 // Should be done. 771 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 772 // Just do a bitcast, we know the sizes match up. 773 FromVal = Builder.CreateBitCast(FromVal, ToType); 774 } else { 775 // Otherwise must be a pointer. 776 FromVal = Builder.CreateIntToPtr(FromVal, ToType); 777 } 778 assert(FromVal->getType() == ToType && "Didn't convert right?"); 779 return FromVal; 780} 781 782/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 783/// or vector value "Old" at the offset specified by Offset. 784/// 785/// This happens when we are converting an "integer union" to a 786/// single integer scalar, or when we are converting a "vector union" to a 787/// vector with insert/extractelement instructions. 788/// 789/// Offset is an offset from the original alloca, in bits that need to be 790/// shifted to the right. 791Value *ConvertToScalarInfo:: 792ConvertScalar_InsertValue(Value *SV, Value *Old, 793 uint64_t Offset, IRBuilder<> &Builder) { 794 // Convert the stored type to the actual type, shift it left to insert 795 // then 'or' into place. 796 Type *AllocaType = Old->getType(); 797 LLVMContext &Context = Old->getContext(); 798 799 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 800 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 801 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 802 803 // Changing the whole vector with memset or with an access of a different 804 // vector type? 805 if (ValSize == VecSize) 806 return Builder.CreateBitCast(SV, AllocaType); 807 808 // Must be an element insertion. 809 Type *EltTy = VTy->getElementType(); 810 if (SV->getType() != EltTy) 811 SV = Builder.CreateBitCast(SV, EltTy); 812 uint64_t EltSize = TD.getTypeAllocSizeInBits(EltTy); 813 unsigned Elt = Offset/EltSize; 814 return Builder.CreateInsertElement(Old, SV, Builder.getInt32(Elt)); 815 } 816 817 // If SV is a first-class aggregate value, insert each value recursively. 818 if (StructType *ST = dyn_cast<StructType>(SV->getType())) { 819 const StructLayout &Layout = *TD.getStructLayout(ST); 820 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 821 Value *Elt = Builder.CreateExtractValue(SV, i); 822 Old = ConvertScalar_InsertValue(Elt, Old, 823 Offset+Layout.getElementOffsetInBits(i), 824 Builder); 825 } 826 return Old; 827 } 828 829 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 830 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 831 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 832 Value *Elt = Builder.CreateExtractValue(SV, i); 833 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 834 } 835 return Old; 836 } 837 838 // If SV is a float, convert it to the appropriate integer type. 839 // If it is a pointer, do the same. 840 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 841 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 842 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 843 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 844 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 845 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth)); 846 else if (SV->getType()->isPointerTy()) 847 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext())); 848 849 // Zero extend or truncate the value if needed. 850 if (SV->getType() != AllocaType) { 851 if (SV->getType()->getPrimitiveSizeInBits() < 852 AllocaType->getPrimitiveSizeInBits()) 853 SV = Builder.CreateZExt(SV, AllocaType); 854 else { 855 // Truncation may be needed if storing more than the alloca can hold 856 // (undefined behavior). 857 SV = Builder.CreateTrunc(SV, AllocaType); 858 SrcWidth = DestWidth; 859 SrcStoreWidth = DestStoreWidth; 860 } 861 } 862 863 // If this is a big-endian system and the store is narrower than the 864 // full alloca type, we need to do a shift to get the right bits. 865 int ShAmt = 0; 866 if (TD.isBigEndian()) { 867 // On big-endian machines, the lowest bit is stored at the bit offset 868 // from the pointer given by getTypeStoreSizeInBits. This matters for 869 // integers with a bitwidth that is not a multiple of 8. 870 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 871 } else { 872 ShAmt = Offset; 873 } 874 875 // Note: we support negative bitwidths (with shr) which are not defined. 876 // We do this to support (f.e.) stores off the end of a structure where 877 // only some bits in the structure are set. 878 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 879 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 880 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt)); 881 Mask <<= ShAmt; 882 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 883 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt)); 884 Mask = Mask.lshr(-ShAmt); 885 } 886 887 // Mask out the bits we are about to insert from the old value, and or 888 // in the new bits. 889 if (SrcWidth != DestWidth) { 890 assert(DestWidth > SrcWidth); 891 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 892 SV = Builder.CreateOr(Old, SV, "ins"); 893 } 894 return SV; 895} 896 897 898//===----------------------------------------------------------------------===// 899// SRoA Driver 900//===----------------------------------------------------------------------===// 901 902 903bool SROA::runOnFunction(Function &F) { 904 TD = getAnalysisIfAvailable<TargetData>(); 905 906 bool Changed = performPromotion(F); 907 908 // FIXME: ScalarRepl currently depends on TargetData more than it 909 // theoretically needs to. It should be refactored in order to support 910 // target-independent IR. Until this is done, just skip the actual 911 // scalar-replacement portion of this pass. 912 if (!TD) return Changed; 913 914 while (1) { 915 bool LocalChange = performScalarRepl(F); 916 if (!LocalChange) break; // No need to repromote if no scalarrepl 917 Changed = true; 918 LocalChange = performPromotion(F); 919 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 920 } 921 922 return Changed; 923} 924 925namespace { 926class AllocaPromoter : public LoadAndStorePromoter { 927 AllocaInst *AI; 928 DIBuilder *DIB; 929 SmallVector<DbgDeclareInst *, 4> DDIs; 930 SmallVector<DbgValueInst *, 4> DVIs; 931public: 932 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 933 DIBuilder *DB) 934 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {} 935 936 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 937 // Remember which alloca we're promoting (for isInstInList). 938 this->AI = AI; 939 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) 940 for (Value::use_iterator UI = DebugNode->use_begin(), 941 E = DebugNode->use_end(); UI != E; ++UI) 942 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI)) 943 DDIs.push_back(DDI); 944 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI)) 945 DVIs.push_back(DVI); 946 947 LoadAndStorePromoter::run(Insts); 948 AI->eraseFromParent(); 949 for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(), 950 E = DDIs.end(); I != E; ++I) { 951 DbgDeclareInst *DDI = *I; 952 DDI->eraseFromParent(); 953 } 954 for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(), 955 E = DVIs.end(); I != E; ++I) { 956 DbgValueInst *DVI = *I; 957 DVI->eraseFromParent(); 958 } 959 } 960 961 virtual bool isInstInList(Instruction *I, 962 const SmallVectorImpl<Instruction*> &Insts) const { 963 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 964 return LI->getOperand(0) == AI; 965 return cast<StoreInst>(I)->getPointerOperand() == AI; 966 } 967 968 virtual void updateDebugInfo(Instruction *Inst) const { 969 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(), 970 E = DDIs.end(); I != E; ++I) { 971 DbgDeclareInst *DDI = *I; 972 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 973 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB); 974 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 975 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB); 976 } 977 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(), 978 E = DVIs.end(); I != E; ++I) { 979 DbgValueInst *DVI = *I; 980 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 981 Instruction *DbgVal = NULL; 982 // If an argument is zero extended then use argument directly. The ZExt 983 // may be zapped by an optimization pass in future. 984 Argument *ExtendedArg = NULL; 985 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 986 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0)); 987 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 988 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0)); 989 if (ExtendedArg) 990 DbgVal = DIB->insertDbgValueIntrinsic(ExtendedArg, 0, 991 DIVariable(DVI->getVariable()), 992 SI); 993 else 994 DbgVal = DIB->insertDbgValueIntrinsic(SI->getOperand(0), 0, 995 DIVariable(DVI->getVariable()), 996 SI); 997 DbgVal->setDebugLoc(DVI->getDebugLoc()); 998 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 999 Instruction *DbgVal = 1000 DIB->insertDbgValueIntrinsic(LI->getOperand(0), 0, 1001 DIVariable(DVI->getVariable()), LI); 1002 DbgVal->setDebugLoc(DVI->getDebugLoc()); 1003 } 1004 } 1005 } 1006}; 1007} // end anon namespace 1008 1009/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1010/// subsequently loaded can be rewritten to load both input pointers and then 1011/// select between the result, allowing the load of the alloca to be promoted. 1012/// From this: 1013/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1014/// %V = load i32* %P2 1015/// to: 1016/// %V1 = load i32* %Alloca -> will be mem2reg'd 1017/// %V2 = load i32* %Other 1018/// %V = select i1 %cond, i32 %V1, i32 %V2 1019/// 1020/// We can do this to a select if its only uses are loads and if the operand to 1021/// the select can be loaded unconditionally. 1022static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1023 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1024 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1025 1026 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1027 UI != UE; ++UI) { 1028 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1029 if (LI == 0 || !LI->isSimple()) return false; 1030 1031 // Both operands to the select need to be dereferencable, either absolutely 1032 // (e.g. allocas) or at this point because we can see other accesses to it. 1033 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1034 LI->getAlignment(), TD)) 1035 return false; 1036 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1037 LI->getAlignment(), TD)) 1038 return false; 1039 } 1040 1041 return true; 1042} 1043 1044/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1045/// subsequently loaded can be rewritten to load both input pointers in the pred 1046/// blocks and then PHI the results, allowing the load of the alloca to be 1047/// promoted. 1048/// From this: 1049/// %P2 = phi [i32* %Alloca, i32* %Other] 1050/// %V = load i32* %P2 1051/// to: 1052/// %V1 = load i32* %Alloca -> will be mem2reg'd 1053/// ... 1054/// %V2 = load i32* %Other 1055/// ... 1056/// %V = phi [i32 %V1, i32 %V2] 1057/// 1058/// We can do this to a select if its only uses are loads and if the operand to 1059/// the select can be loaded unconditionally. 1060static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1061 // For now, we can only do this promotion if the load is in the same block as 1062 // the PHI, and if there are no stores between the phi and load. 1063 // TODO: Allow recursive phi users. 1064 // TODO: Allow stores. 1065 BasicBlock *BB = PN->getParent(); 1066 unsigned MaxAlign = 0; 1067 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1068 UI != UE; ++UI) { 1069 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1070 if (LI == 0 || !LI->isSimple()) return false; 1071 1072 // For now we only allow loads in the same block as the PHI. This is a 1073 // common case that happens when instcombine merges two loads through a PHI. 1074 if (LI->getParent() != BB) return false; 1075 1076 // Ensure that there are no instructions between the PHI and the load that 1077 // could store. 1078 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1079 if (BBI->mayWriteToMemory()) 1080 return false; 1081 1082 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1083 } 1084 1085 // Okay, we know that we have one or more loads in the same block as the PHI. 1086 // We can transform this if it is safe to push the loads into the predecessor 1087 // blocks. The only thing to watch out for is that we can't put a possibly 1088 // trapping load in the predecessor if it is a critical edge. 1089 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1090 BasicBlock *Pred = PN->getIncomingBlock(i); 1091 Value *InVal = PN->getIncomingValue(i); 1092 1093 // If the terminator of the predecessor has side-effects (an invoke), 1094 // there is no safe place to put a load in the predecessor. 1095 if (Pred->getTerminator()->mayHaveSideEffects()) 1096 return false; 1097 1098 // If the value is produced by the terminator of the predecessor 1099 // (an invoke), there is no valid place to put a load in the predecessor. 1100 if (Pred->getTerminator() == InVal) 1101 return false; 1102 1103 // If the predecessor has a single successor, then the edge isn't critical. 1104 if (Pred->getTerminator()->getNumSuccessors() == 1) 1105 continue; 1106 1107 // If this pointer is always safe to load, or if we can prove that there is 1108 // already a load in the block, then we can move the load to the pred block. 1109 if (InVal->isDereferenceablePointer() || 1110 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1111 continue; 1112 1113 return false; 1114 } 1115 1116 return true; 1117} 1118 1119 1120/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1121/// direct (non-volatile) loads and stores to it. If the alloca is close but 1122/// not quite there, this will transform the code to allow promotion. As such, 1123/// it is a non-pure predicate. 1124static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1125 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1126 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1127 1128 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1129 UI != UE; ++UI) { 1130 User *U = *UI; 1131 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1132 if (!LI->isSimple()) 1133 return false; 1134 continue; 1135 } 1136 1137 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1138 if (SI->getOperand(0) == AI || !SI->isSimple()) 1139 return false; // Don't allow a store OF the AI, only INTO the AI. 1140 continue; 1141 } 1142 1143 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1144 // If the condition being selected on is a constant, fold the select, yes 1145 // this does (rarely) happen early on. 1146 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1147 Value *Result = SI->getOperand(1+CI->isZero()); 1148 SI->replaceAllUsesWith(Result); 1149 SI->eraseFromParent(); 1150 1151 // This is very rare and we just scrambled the use list of AI, start 1152 // over completely. 1153 return tryToMakeAllocaBePromotable(AI, TD); 1154 } 1155 1156 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1157 // loads, then we can transform this by rewriting the select. 1158 if (!isSafeSelectToSpeculate(SI, TD)) 1159 return false; 1160 1161 InstsToRewrite.insert(SI); 1162 continue; 1163 } 1164 1165 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1166 if (PN->use_empty()) { // Dead PHIs can be stripped. 1167 InstsToRewrite.insert(PN); 1168 continue; 1169 } 1170 1171 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1172 // in the pred blocks, then we can transform this by rewriting the PHI. 1173 if (!isSafePHIToSpeculate(PN, TD)) 1174 return false; 1175 1176 InstsToRewrite.insert(PN); 1177 continue; 1178 } 1179 1180 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1181 if (onlyUsedByLifetimeMarkers(BCI)) { 1182 InstsToRewrite.insert(BCI); 1183 continue; 1184 } 1185 } 1186 1187 return false; 1188 } 1189 1190 // If there are no instructions to rewrite, then all uses are load/stores and 1191 // we're done! 1192 if (InstsToRewrite.empty()) 1193 return true; 1194 1195 // If we have instructions that need to be rewritten for this to be promotable 1196 // take care of it now. 1197 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1198 if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) { 1199 // This could only be a bitcast used by nothing but lifetime intrinsics. 1200 for (BitCastInst::use_iterator I = BCI->use_begin(), E = BCI->use_end(); 1201 I != E;) { 1202 Use &U = I.getUse(); 1203 ++I; 1204 cast<Instruction>(U.getUser())->eraseFromParent(); 1205 } 1206 BCI->eraseFromParent(); 1207 continue; 1208 } 1209 1210 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1211 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1212 // loads with a new select. 1213 while (!SI->use_empty()) { 1214 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1215 1216 IRBuilder<> Builder(LI); 1217 LoadInst *TrueLoad = 1218 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1219 LoadInst *FalseLoad = 1220 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f"); 1221 1222 // Transfer alignment and TBAA info if present. 1223 TrueLoad->setAlignment(LI->getAlignment()); 1224 FalseLoad->setAlignment(LI->getAlignment()); 1225 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1226 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1227 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1228 } 1229 1230 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1231 V->takeName(LI); 1232 LI->replaceAllUsesWith(V); 1233 LI->eraseFromParent(); 1234 } 1235 1236 // Now that all the loads are gone, the select is gone too. 1237 SI->eraseFromParent(); 1238 continue; 1239 } 1240 1241 // Otherwise, we have a PHI node which allows us to push the loads into the 1242 // predecessors. 1243 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1244 if (PN->use_empty()) { 1245 PN->eraseFromParent(); 1246 continue; 1247 } 1248 1249 Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1250 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1251 PN->getName()+".ld", PN); 1252 1253 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1254 // matter which one we get and if any differ, it doesn't matter. 1255 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1256 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1257 unsigned Align = SomeLoad->getAlignment(); 1258 1259 // Rewrite all loads of the PN to use the new PHI. 1260 while (!PN->use_empty()) { 1261 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1262 LI->replaceAllUsesWith(NewPN); 1263 LI->eraseFromParent(); 1264 } 1265 1266 // Inject loads into all of the pred blocks. Keep track of which blocks we 1267 // insert them into in case we have multiple edges from the same block. 1268 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1269 1270 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1271 BasicBlock *Pred = PN->getIncomingBlock(i); 1272 LoadInst *&Load = InsertedLoads[Pred]; 1273 if (Load == 0) { 1274 Load = new LoadInst(PN->getIncomingValue(i), 1275 PN->getName() + "." + Pred->getName(), 1276 Pred->getTerminator()); 1277 Load->setAlignment(Align); 1278 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1279 } 1280 1281 NewPN->addIncoming(Load, Pred); 1282 } 1283 1284 PN->eraseFromParent(); 1285 } 1286 1287 ++NumAdjusted; 1288 return true; 1289} 1290 1291bool SROA::performPromotion(Function &F) { 1292 std::vector<AllocaInst*> Allocas; 1293 DominatorTree *DT = 0; 1294 if (HasDomTree) 1295 DT = &getAnalysis<DominatorTree>(); 1296 1297 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1298 DIBuilder DIB(*F.getParent()); 1299 bool Changed = false; 1300 SmallVector<Instruction*, 64> Insts; 1301 while (1) { 1302 Allocas.clear(); 1303 1304 // Find allocas that are safe to promote, by looking at all instructions in 1305 // the entry node 1306 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1307 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1308 if (tryToMakeAllocaBePromotable(AI, TD)) 1309 Allocas.push_back(AI); 1310 1311 if (Allocas.empty()) break; 1312 1313 if (HasDomTree) 1314 PromoteMemToReg(Allocas, *DT); 1315 else { 1316 SSAUpdater SSA; 1317 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1318 AllocaInst *AI = Allocas[i]; 1319 1320 // Build list of instructions to promote. 1321 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1322 UI != E; ++UI) 1323 Insts.push_back(cast<Instruction>(*UI)); 1324 AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts); 1325 Insts.clear(); 1326 } 1327 } 1328 NumPromoted += Allocas.size(); 1329 Changed = true; 1330 } 1331 1332 return Changed; 1333} 1334 1335 1336/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1337/// SROA. It must be a struct or array type with a small number of elements. 1338static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1339 Type *T = AI->getAllocatedType(); 1340 // Do not promote any struct into more than 32 separate vars. 1341 if (StructType *ST = dyn_cast<StructType>(T)) 1342 return ST->getNumElements() <= 32; 1343 // Arrays are much less likely to be safe for SROA; only consider 1344 // them if they are very small. 1345 if (ArrayType *AT = dyn_cast<ArrayType>(T)) 1346 return AT->getNumElements() <= 8; 1347 return false; 1348} 1349 1350 1351// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1352// which runs on all of the alloca instructions in the function, removing them 1353// if they are only used by getelementptr instructions. 1354// 1355bool SROA::performScalarRepl(Function &F) { 1356 std::vector<AllocaInst*> WorkList; 1357 1358 // Scan the entry basic block, adding allocas to the worklist. 1359 BasicBlock &BB = F.getEntryBlock(); 1360 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1361 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1362 WorkList.push_back(A); 1363 1364 // Process the worklist 1365 bool Changed = false; 1366 while (!WorkList.empty()) { 1367 AllocaInst *AI = WorkList.back(); 1368 WorkList.pop_back(); 1369 1370 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1371 // with unused elements. 1372 if (AI->use_empty()) { 1373 AI->eraseFromParent(); 1374 Changed = true; 1375 continue; 1376 } 1377 1378 // If this alloca is impossible for us to promote, reject it early. 1379 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1380 continue; 1381 1382 // Check to see if this allocation is only modified by a memcpy/memmove from 1383 // a constant global. If this is the case, we can change all users to use 1384 // the constant global instead. This is commonly produced by the CFE by 1385 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1386 // is only subsequently read. 1387 SmallVector<Instruction *, 4> ToDelete; 1388 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) { 1389 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1390 DEBUG(dbgs() << " memcpy = " << *Copy << '\n'); 1391 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i) 1392 ToDelete[i]->eraseFromParent(); 1393 Constant *TheSrc = cast<Constant>(Copy->getSource()); 1394 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1395 Copy->eraseFromParent(); // Don't mutate the global. 1396 AI->eraseFromParent(); 1397 ++NumGlobals; 1398 Changed = true; 1399 continue; 1400 } 1401 1402 // Check to see if we can perform the core SROA transformation. We cannot 1403 // transform the allocation instruction if it is an array allocation 1404 // (allocations OF arrays are ok though), and an allocation of a scalar 1405 // value cannot be decomposed at all. 1406 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1407 1408 // Do not promote [0 x %struct]. 1409 if (AllocaSize == 0) continue; 1410 1411 // Do not promote any struct whose size is too big. 1412 if (AllocaSize > SRThreshold) continue; 1413 1414 // If the alloca looks like a good candidate for scalar replacement, and if 1415 // all its users can be transformed, then split up the aggregate into its 1416 // separate elements. 1417 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1418 DoScalarReplacement(AI, WorkList); 1419 Changed = true; 1420 continue; 1421 } 1422 1423 // If we can turn this aggregate value (potentially with casts) into a 1424 // simple scalar value that can be mem2reg'd into a register value. 1425 // IsNotTrivial tracks whether this is something that mem2reg could have 1426 // promoted itself. If so, we don't want to transform it needlessly. Note 1427 // that we can't just check based on the type: the alloca may be of an i32 1428 // but that has pointer arithmetic to set byte 3 of it or something. 1429 if (AllocaInst *NewAI = 1430 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1431 NewAI->takeName(AI); 1432 AI->eraseFromParent(); 1433 ++NumConverted; 1434 Changed = true; 1435 continue; 1436 } 1437 1438 // Otherwise, couldn't process this alloca. 1439 } 1440 1441 return Changed; 1442} 1443 1444/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1445/// predicate, do SROA now. 1446void SROA::DoScalarReplacement(AllocaInst *AI, 1447 std::vector<AllocaInst*> &WorkList) { 1448 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1449 SmallVector<AllocaInst*, 32> ElementAllocas; 1450 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1451 ElementAllocas.reserve(ST->getNumContainedTypes()); 1452 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1453 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1454 AI->getAlignment(), 1455 AI->getName() + "." + Twine(i), AI); 1456 ElementAllocas.push_back(NA); 1457 WorkList.push_back(NA); // Add to worklist for recursive processing 1458 } 1459 } else { 1460 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1461 ElementAllocas.reserve(AT->getNumElements()); 1462 Type *ElTy = AT->getElementType(); 1463 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1464 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1465 AI->getName() + "." + Twine(i), AI); 1466 ElementAllocas.push_back(NA); 1467 WorkList.push_back(NA); // Add to worklist for recursive processing 1468 } 1469 } 1470 1471 // Now that we have created the new alloca instructions, rewrite all the 1472 // uses of the old alloca. 1473 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1474 1475 // Now erase any instructions that were made dead while rewriting the alloca. 1476 DeleteDeadInstructions(); 1477 AI->eraseFromParent(); 1478 1479 ++NumReplaced; 1480} 1481 1482/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1483/// recursively including all their operands that become trivially dead. 1484void SROA::DeleteDeadInstructions() { 1485 while (!DeadInsts.empty()) { 1486 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1487 1488 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1489 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1490 // Zero out the operand and see if it becomes trivially dead. 1491 // (But, don't add allocas to the dead instruction list -- they are 1492 // already on the worklist and will be deleted separately.) 1493 *OI = 0; 1494 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1495 DeadInsts.push_back(U); 1496 } 1497 1498 I->eraseFromParent(); 1499 } 1500} 1501 1502/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1503/// performing scalar replacement of alloca AI. The results are flagged in 1504/// the Info parameter. Offset indicates the position within AI that is 1505/// referenced by this instruction. 1506void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1507 AllocaInfo &Info) { 1508 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1509 Instruction *User = cast<Instruction>(*UI); 1510 1511 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1512 isSafeForScalarRepl(BC, Offset, Info); 1513 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1514 uint64_t GEPOffset = Offset; 1515 isSafeGEP(GEPI, GEPOffset, Info); 1516 if (!Info.isUnsafe) 1517 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1518 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1519 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1520 if (Length == 0) 1521 return MarkUnsafe(Info, User); 1522 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1523 UI.getOperandNo() == 0, Info, MI, 1524 true /*AllowWholeAccess*/); 1525 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1526 if (!LI->isSimple()) 1527 return MarkUnsafe(Info, User); 1528 Type *LIType = LI->getType(); 1529 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1530 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1531 Info.hasALoadOrStore = true; 1532 1533 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1534 // Store is ok if storing INTO the pointer, not storing the pointer 1535 if (!SI->isSimple() || SI->getOperand(0) == I) 1536 return MarkUnsafe(Info, User); 1537 1538 Type *SIType = SI->getOperand(0)->getType(); 1539 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1540 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1541 Info.hasALoadOrStore = true; 1542 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1543 if (II->getIntrinsicID() != Intrinsic::lifetime_start && 1544 II->getIntrinsicID() != Intrinsic::lifetime_end) 1545 return MarkUnsafe(Info, User); 1546 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1547 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1548 } else { 1549 return MarkUnsafe(Info, User); 1550 } 1551 if (Info.isUnsafe) return; 1552 } 1553} 1554 1555 1556/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1557/// derived from the alloca, we can often still split the alloca into elements. 1558/// This is useful if we have a large alloca where one element is phi'd 1559/// together somewhere: we can SRoA and promote all the other elements even if 1560/// we end up not being able to promote this one. 1561/// 1562/// All we require is that the uses of the PHI do not index into other parts of 1563/// the alloca. The most important use case for this is single load and stores 1564/// that are PHI'd together, which can happen due to code sinking. 1565void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1566 AllocaInfo &Info) { 1567 // If we've already checked this PHI, don't do it again. 1568 if (PHINode *PN = dyn_cast<PHINode>(I)) 1569 if (!Info.CheckedPHIs.insert(PN)) 1570 return; 1571 1572 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1573 Instruction *User = cast<Instruction>(*UI); 1574 1575 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1576 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1577 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1578 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1579 // but would have to prove that we're staying inside of an element being 1580 // promoted. 1581 if (!GEPI->hasAllZeroIndices()) 1582 return MarkUnsafe(Info, User); 1583 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1584 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1585 if (!LI->isSimple()) 1586 return MarkUnsafe(Info, User); 1587 Type *LIType = LI->getType(); 1588 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1589 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1590 Info.hasALoadOrStore = true; 1591 1592 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1593 // Store is ok if storing INTO the pointer, not storing the pointer 1594 if (!SI->isSimple() || SI->getOperand(0) == I) 1595 return MarkUnsafe(Info, User); 1596 1597 Type *SIType = SI->getOperand(0)->getType(); 1598 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1599 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1600 Info.hasALoadOrStore = true; 1601 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1602 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1603 } else { 1604 return MarkUnsafe(Info, User); 1605 } 1606 if (Info.isUnsafe) return; 1607 } 1608} 1609 1610/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1611/// replacement. It is safe when all the indices are constant, in-bounds 1612/// references, and when the resulting offset corresponds to an element within 1613/// the alloca type. The results are flagged in the Info parameter. Upon 1614/// return, Offset is adjusted as specified by the GEP indices. 1615void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1616 uint64_t &Offset, AllocaInfo &Info) { 1617 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1618 if (GEPIt == E) 1619 return; 1620 1621 // Walk through the GEP type indices, checking the types that this indexes 1622 // into. 1623 for (; GEPIt != E; ++GEPIt) { 1624 // Ignore struct elements, no extra checking needed for these. 1625 if ((*GEPIt)->isStructTy()) 1626 continue; 1627 1628 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1629 if (!IdxVal) 1630 return MarkUnsafe(Info, GEPI); 1631 } 1632 1633 // Compute the offset due to this GEP and check if the alloca has a 1634 // component element at that offset. 1635 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1636 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1637 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1638 MarkUnsafe(Info, GEPI); 1639} 1640 1641/// isHomogeneousAggregate - Check if type T is a struct or array containing 1642/// elements of the same type (which is always true for arrays). If so, 1643/// return true with NumElts and EltTy set to the number of elements and the 1644/// element type, respectively. 1645static bool isHomogeneousAggregate(Type *T, unsigned &NumElts, 1646 Type *&EltTy) { 1647 if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1648 NumElts = AT->getNumElements(); 1649 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1650 return true; 1651 } 1652 if (StructType *ST = dyn_cast<StructType>(T)) { 1653 NumElts = ST->getNumContainedTypes(); 1654 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1655 for (unsigned n = 1; n < NumElts; ++n) { 1656 if (ST->getContainedType(n) != EltTy) 1657 return false; 1658 } 1659 return true; 1660 } 1661 return false; 1662} 1663 1664/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1665/// "homogeneous" aggregates with the same element type and number of elements. 1666static bool isCompatibleAggregate(Type *T1, Type *T2) { 1667 if (T1 == T2) 1668 return true; 1669 1670 unsigned NumElts1, NumElts2; 1671 Type *EltTy1, *EltTy2; 1672 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1673 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1674 NumElts1 == NumElts2 && 1675 EltTy1 == EltTy2) 1676 return true; 1677 1678 return false; 1679} 1680 1681/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1682/// alloca or has an offset and size that corresponds to a component element 1683/// within it. The offset checked here may have been formed from a GEP with a 1684/// pointer bitcasted to a different type. 1685/// 1686/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1687/// unit. If false, it only allows accesses known to be in a single element. 1688void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1689 Type *MemOpType, bool isStore, 1690 AllocaInfo &Info, Instruction *TheAccess, 1691 bool AllowWholeAccess) { 1692 // Check if this is a load/store of the entire alloca. 1693 if (Offset == 0 && AllowWholeAccess && 1694 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1695 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1696 // loads/stores (which are essentially the same as the MemIntrinsics with 1697 // regard to copying padding between elements). But, if an alloca is 1698 // flagged as both a source and destination of such operations, we'll need 1699 // to check later for padding between elements. 1700 if (!MemOpType || MemOpType->isIntegerTy()) { 1701 if (isStore) 1702 Info.isMemCpyDst = true; 1703 else 1704 Info.isMemCpySrc = true; 1705 return; 1706 } 1707 // This is also safe for references using a type that is compatible with 1708 // the type of the alloca, so that loads/stores can be rewritten using 1709 // insertvalue/extractvalue. 1710 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1711 Info.hasSubelementAccess = true; 1712 return; 1713 } 1714 } 1715 // Check if the offset/size correspond to a component within the alloca type. 1716 Type *T = Info.AI->getAllocatedType(); 1717 if (TypeHasComponent(T, Offset, MemSize)) { 1718 Info.hasSubelementAccess = true; 1719 return; 1720 } 1721 1722 return MarkUnsafe(Info, TheAccess); 1723} 1724 1725/// TypeHasComponent - Return true if T has a component type with the 1726/// specified offset and size. If Size is zero, do not check the size. 1727bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { 1728 Type *EltTy; 1729 uint64_t EltSize; 1730 if (StructType *ST = dyn_cast<StructType>(T)) { 1731 const StructLayout *Layout = TD->getStructLayout(ST); 1732 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1733 EltTy = ST->getContainedType(EltIdx); 1734 EltSize = TD->getTypeAllocSize(EltTy); 1735 Offset -= Layout->getElementOffset(EltIdx); 1736 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1737 EltTy = AT->getElementType(); 1738 EltSize = TD->getTypeAllocSize(EltTy); 1739 if (Offset >= AT->getNumElements() * EltSize) 1740 return false; 1741 Offset %= EltSize; 1742 } else { 1743 return false; 1744 } 1745 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1746 return true; 1747 // Check if the component spans multiple elements. 1748 if (Offset + Size > EltSize) 1749 return false; 1750 return TypeHasComponent(EltTy, Offset, Size); 1751} 1752 1753/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1754/// the instruction I, which references it, to use the separate elements. 1755/// Offset indicates the position within AI that is referenced by this 1756/// instruction. 1757void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1758 SmallVector<AllocaInst*, 32> &NewElts) { 1759 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1760 Use &TheUse = UI.getUse(); 1761 Instruction *User = cast<Instruction>(*UI++); 1762 1763 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1764 RewriteBitCast(BC, AI, Offset, NewElts); 1765 continue; 1766 } 1767 1768 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1769 RewriteGEP(GEPI, AI, Offset, NewElts); 1770 continue; 1771 } 1772 1773 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1774 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1775 uint64_t MemSize = Length->getZExtValue(); 1776 if (Offset == 0 && 1777 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1778 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1779 // Otherwise the intrinsic can only touch a single element and the 1780 // address operand will be updated, so nothing else needs to be done. 1781 continue; 1782 } 1783 1784 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1785 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 1786 II->getIntrinsicID() == Intrinsic::lifetime_end) { 1787 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts); 1788 } 1789 continue; 1790 } 1791 1792 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1793 Type *LIType = LI->getType(); 1794 1795 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1796 // Replace: 1797 // %res = load { i32, i32 }* %alloc 1798 // with: 1799 // %load.0 = load i32* %alloc.0 1800 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1801 // %load.1 = load i32* %alloc.1 1802 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1803 // (Also works for arrays instead of structs) 1804 Value *Insert = UndefValue::get(LIType); 1805 IRBuilder<> Builder(LI); 1806 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1807 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1808 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1809 } 1810 LI->replaceAllUsesWith(Insert); 1811 DeadInsts.push_back(LI); 1812 } else if (LIType->isIntegerTy() && 1813 TD->getTypeAllocSize(LIType) == 1814 TD->getTypeAllocSize(AI->getAllocatedType())) { 1815 // If this is a load of the entire alloca to an integer, rewrite it. 1816 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1817 } 1818 continue; 1819 } 1820 1821 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1822 Value *Val = SI->getOperand(0); 1823 Type *SIType = Val->getType(); 1824 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1825 // Replace: 1826 // store { i32, i32 } %val, { i32, i32 }* %alloc 1827 // with: 1828 // %val.0 = extractvalue { i32, i32 } %val, 0 1829 // store i32 %val.0, i32* %alloc.0 1830 // %val.1 = extractvalue { i32, i32 } %val, 1 1831 // store i32 %val.1, i32* %alloc.1 1832 // (Also works for arrays instead of structs) 1833 IRBuilder<> Builder(SI); 1834 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1835 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1836 Builder.CreateStore(Extract, NewElts[i]); 1837 } 1838 DeadInsts.push_back(SI); 1839 } else if (SIType->isIntegerTy() && 1840 TD->getTypeAllocSize(SIType) == 1841 TD->getTypeAllocSize(AI->getAllocatedType())) { 1842 // If this is a store of the entire alloca from an integer, rewrite it. 1843 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1844 } 1845 continue; 1846 } 1847 1848 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1849 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1850 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1851 // the new pointer. 1852 if (!isa<AllocaInst>(I)) continue; 1853 1854 assert(Offset == 0 && NewElts[0] && 1855 "Direct alloca use should have a zero offset"); 1856 1857 // If we have a use of the alloca, we know the derived uses will be 1858 // utilizing just the first element of the scalarized result. Insert a 1859 // bitcast of the first alloca before the user as required. 1860 AllocaInst *NewAI = NewElts[0]; 1861 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1862 NewAI->moveBefore(BCI); 1863 TheUse = BCI; 1864 continue; 1865 } 1866 } 1867} 1868 1869/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1870/// and recursively continue updating all of its uses. 1871void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1872 SmallVector<AllocaInst*, 32> &NewElts) { 1873 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1874 if (BC->getOperand(0) != AI) 1875 return; 1876 1877 // The bitcast references the original alloca. Replace its uses with 1878 // references to the first new element alloca. 1879 Instruction *Val = NewElts[0]; 1880 if (Val->getType() != BC->getDestTy()) { 1881 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1882 Val->takeName(BC); 1883 } 1884 BC->replaceAllUsesWith(Val); 1885 DeadInsts.push_back(BC); 1886} 1887 1888/// FindElementAndOffset - Return the index of the element containing Offset 1889/// within the specified type, which must be either a struct or an array. 1890/// Sets T to the type of the element and Offset to the offset within that 1891/// element. IdxTy is set to the type of the index result to be used in a 1892/// GEP instruction. 1893uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, 1894 Type *&IdxTy) { 1895 uint64_t Idx = 0; 1896 if (StructType *ST = dyn_cast<StructType>(T)) { 1897 const StructLayout *Layout = TD->getStructLayout(ST); 1898 Idx = Layout->getElementContainingOffset(Offset); 1899 T = ST->getContainedType(Idx); 1900 Offset -= Layout->getElementOffset(Idx); 1901 IdxTy = Type::getInt32Ty(T->getContext()); 1902 return Idx; 1903 } 1904 ArrayType *AT = cast<ArrayType>(T); 1905 T = AT->getElementType(); 1906 uint64_t EltSize = TD->getTypeAllocSize(T); 1907 Idx = Offset / EltSize; 1908 Offset -= Idx * EltSize; 1909 IdxTy = Type::getInt64Ty(T->getContext()); 1910 return Idx; 1911} 1912 1913/// RewriteGEP - Check if this GEP instruction moves the pointer across 1914/// elements of the alloca that are being split apart, and if so, rewrite 1915/// the GEP to be relative to the new element. 1916void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1917 SmallVector<AllocaInst*, 32> &NewElts) { 1918 uint64_t OldOffset = Offset; 1919 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1920 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1921 1922 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1923 1924 Type *T = AI->getAllocatedType(); 1925 Type *IdxTy; 1926 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1927 if (GEPI->getOperand(0) == AI) 1928 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1929 1930 T = AI->getAllocatedType(); 1931 uint64_t EltOffset = Offset; 1932 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1933 1934 // If this GEP does not move the pointer across elements of the alloca 1935 // being split, then it does not needs to be rewritten. 1936 if (Idx == OldIdx) 1937 return; 1938 1939 Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1940 SmallVector<Value*, 8> NewArgs; 1941 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1942 while (EltOffset != 0) { 1943 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1944 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1945 } 1946 Instruction *Val = NewElts[Idx]; 1947 if (NewArgs.size() > 1) { 1948 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI); 1949 Val->takeName(GEPI); 1950 } 1951 if (Val->getType() != GEPI->getType()) 1952 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1953 GEPI->replaceAllUsesWith(Val); 1954 DeadInsts.push_back(GEPI); 1955} 1956 1957/// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it 1958/// to mark the lifetime of the scalarized memory. 1959void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 1960 uint64_t Offset, 1961 SmallVector<AllocaInst*, 32> &NewElts) { 1962 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0)); 1963 // Put matching lifetime markers on everything from Offset up to 1964 // Offset+OldSize. 1965 Type *AIType = AI->getAllocatedType(); 1966 uint64_t NewOffset = Offset; 1967 Type *IdxTy; 1968 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy); 1969 1970 IRBuilder<> Builder(II); 1971 uint64_t Size = OldSize->getLimitedValue(); 1972 1973 if (NewOffset) { 1974 // Splice the first element and index 'NewOffset' bytes in. SROA will 1975 // split the alloca again later. 1976 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy()); 1977 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset)); 1978 1979 IdxTy = NewElts[Idx]->getAllocatedType(); 1980 uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset; 1981 if (EltSize > Size) { 1982 EltSize = Size; 1983 Size = 0; 1984 } else { 1985 Size -= EltSize; 1986 } 1987 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 1988 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize)); 1989 else 1990 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize)); 1991 ++Idx; 1992 } 1993 1994 for (; Idx != NewElts.size() && Size; ++Idx) { 1995 IdxTy = NewElts[Idx]->getAllocatedType(); 1996 uint64_t EltSize = TD->getTypeAllocSize(IdxTy); 1997 if (EltSize > Size) { 1998 EltSize = Size; 1999 Size = 0; 2000 } else { 2001 Size -= EltSize; 2002 } 2003 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2004 Builder.CreateLifetimeStart(NewElts[Idx], 2005 Builder.getInt64(EltSize)); 2006 else 2007 Builder.CreateLifetimeEnd(NewElts[Idx], 2008 Builder.getInt64(EltSize)); 2009 } 2010 DeadInsts.push_back(II); 2011} 2012 2013/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2014/// Rewrite it to copy or set the elements of the scalarized memory. 2015void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2016 AllocaInst *AI, 2017 SmallVector<AllocaInst*, 32> &NewElts) { 2018 // If this is a memcpy/memmove, construct the other pointer as the 2019 // appropriate type. The "Other" pointer is the pointer that goes to memory 2020 // that doesn't have anything to do with the alloca that we are promoting. For 2021 // memset, this Value* stays null. 2022 Value *OtherPtr = 0; 2023 unsigned MemAlignment = MI->getAlignment(); 2024 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2025 if (Inst == MTI->getRawDest()) 2026 OtherPtr = MTI->getRawSource(); 2027 else { 2028 assert(Inst == MTI->getRawSource()); 2029 OtherPtr = MTI->getRawDest(); 2030 } 2031 } 2032 2033 // If there is an other pointer, we want to convert it to the same pointer 2034 // type as AI has, so we can GEP through it safely. 2035 if (OtherPtr) { 2036 unsigned AddrSpace = 2037 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2038 2039 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2040 // optimization, but it's also required to detect the corner case where 2041 // both pointer operands are referencing the same memory, and where 2042 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2043 // function is only called for mem intrinsics that access the whole 2044 // aggregate, so non-zero GEPs are not an issue here.) 2045 OtherPtr = OtherPtr->stripPointerCasts(); 2046 2047 // Copying the alloca to itself is a no-op: just delete it. 2048 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2049 // This code will run twice for a no-op memcpy -- once for each operand. 2050 // Put only one reference to MI on the DeadInsts list. 2051 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2052 E = DeadInsts.end(); I != E; ++I) 2053 if (*I == MI) return; 2054 DeadInsts.push_back(MI); 2055 return; 2056 } 2057 2058 // If the pointer is not the right type, insert a bitcast to the right 2059 // type. 2060 Type *NewTy = 2061 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2062 2063 if (OtherPtr->getType() != NewTy) 2064 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2065 } 2066 2067 // Process each element of the aggregate. 2068 bool SROADest = MI->getRawDest() == Inst; 2069 2070 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2071 2072 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2073 // If this is a memcpy/memmove, emit a GEP of the other element address. 2074 Value *OtherElt = 0; 2075 unsigned OtherEltAlign = MemAlignment; 2076 2077 if (OtherPtr) { 2078 Value *Idx[2] = { Zero, 2079 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2080 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, 2081 OtherPtr->getName()+"."+Twine(i), 2082 MI); 2083 uint64_t EltOffset; 2084 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2085 Type *OtherTy = OtherPtrTy->getElementType(); 2086 if (StructType *ST = dyn_cast<StructType>(OtherTy)) { 2087 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2088 } else { 2089 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2090 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2091 } 2092 2093 // The alignment of the other pointer is the guaranteed alignment of the 2094 // element, which is affected by both the known alignment of the whole 2095 // mem intrinsic and the alignment of the element. If the alignment of 2096 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2097 // known alignment is just 4 bytes. 2098 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2099 } 2100 2101 Value *EltPtr = NewElts[i]; 2102 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2103 2104 // If we got down to a scalar, insert a load or store as appropriate. 2105 if (EltTy->isSingleValueType()) { 2106 if (isa<MemTransferInst>(MI)) { 2107 if (SROADest) { 2108 // From Other to Alloca. 2109 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2110 new StoreInst(Elt, EltPtr, MI); 2111 } else { 2112 // From Alloca to Other. 2113 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2114 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2115 } 2116 continue; 2117 } 2118 assert(isa<MemSetInst>(MI)); 2119 2120 // If the stored element is zero (common case), just store a null 2121 // constant. 2122 Constant *StoreVal; 2123 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2124 if (CI->isZero()) { 2125 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2126 } else { 2127 // If EltTy is a vector type, get the element type. 2128 Type *ValTy = EltTy->getScalarType(); 2129 2130 // Construct an integer with the right value. 2131 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2132 APInt OneVal(EltSize, CI->getZExtValue()); 2133 APInt TotalVal(OneVal); 2134 // Set each byte. 2135 for (unsigned i = 0; 8*i < EltSize; ++i) { 2136 TotalVal = TotalVal.shl(8); 2137 TotalVal |= OneVal; 2138 } 2139 2140 // Convert the integer value to the appropriate type. 2141 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2142 if (ValTy->isPointerTy()) 2143 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2144 else if (ValTy->isFloatingPointTy()) 2145 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2146 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2147 2148 // If the requested value was a vector constant, create it. 2149 if (EltTy->isVectorTy()) { 2150 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements(); 2151 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2152 StoreVal = ConstantVector::get(Elts); 2153 } 2154 } 2155 new StoreInst(StoreVal, EltPtr, MI); 2156 continue; 2157 } 2158 // Otherwise, if we're storing a byte variable, use a memset call for 2159 // this element. 2160 } 2161 2162 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2163 2164 IRBuilder<> Builder(MI); 2165 2166 // Finally, insert the meminst for this element. 2167 if (isa<MemSetInst>(MI)) { 2168 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2169 MI->isVolatile()); 2170 } else { 2171 assert(isa<MemTransferInst>(MI)); 2172 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2173 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2174 2175 if (isa<MemCpyInst>(MI)) 2176 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2177 else 2178 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2179 } 2180 } 2181 DeadInsts.push_back(MI); 2182} 2183 2184/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2185/// overwrites the entire allocation. Extract out the pieces of the stored 2186/// integer and store them individually. 2187void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2188 SmallVector<AllocaInst*, 32> &NewElts){ 2189 // Extract each element out of the integer according to its structure offset 2190 // and store the element value to the individual alloca. 2191 Value *SrcVal = SI->getOperand(0); 2192 Type *AllocaEltTy = AI->getAllocatedType(); 2193 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2194 2195 IRBuilder<> Builder(SI); 2196 2197 // Handle tail padding by extending the operand 2198 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2199 SrcVal = Builder.CreateZExt(SrcVal, 2200 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2201 2202 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2203 << '\n'); 2204 2205 // There are two forms here: AI could be an array or struct. Both cases 2206 // have different ways to compute the element offset. 2207 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2208 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2209 2210 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2211 // Get the number of bits to shift SrcVal to get the value. 2212 Type *FieldTy = EltSTy->getElementType(i); 2213 uint64_t Shift = Layout->getElementOffsetInBits(i); 2214 2215 if (TD->isBigEndian()) 2216 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2217 2218 Value *EltVal = SrcVal; 2219 if (Shift) { 2220 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2221 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2222 } 2223 2224 // Truncate down to an integer of the right size. 2225 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2226 2227 // Ignore zero sized fields like {}, they obviously contain no data. 2228 if (FieldSizeBits == 0) continue; 2229 2230 if (FieldSizeBits != AllocaSizeBits) 2231 EltVal = Builder.CreateTrunc(EltVal, 2232 IntegerType::get(SI->getContext(), FieldSizeBits)); 2233 Value *DestField = NewElts[i]; 2234 if (EltVal->getType() == FieldTy) { 2235 // Storing to an integer field of this size, just do it. 2236 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2237 // Bitcast to the right element type (for fp/vector values). 2238 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2239 } else { 2240 // Otherwise, bitcast the dest pointer (for aggregates). 2241 DestField = Builder.CreateBitCast(DestField, 2242 PointerType::getUnqual(EltVal->getType())); 2243 } 2244 new StoreInst(EltVal, DestField, SI); 2245 } 2246 2247 } else { 2248 ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2249 Type *ArrayEltTy = ATy->getElementType(); 2250 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2251 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2252 2253 uint64_t Shift; 2254 2255 if (TD->isBigEndian()) 2256 Shift = AllocaSizeBits-ElementOffset; 2257 else 2258 Shift = 0; 2259 2260 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2261 // Ignore zero sized fields like {}, they obviously contain no data. 2262 if (ElementSizeBits == 0) continue; 2263 2264 Value *EltVal = SrcVal; 2265 if (Shift) { 2266 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2267 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2268 } 2269 2270 // Truncate down to an integer of the right size. 2271 if (ElementSizeBits != AllocaSizeBits) 2272 EltVal = Builder.CreateTrunc(EltVal, 2273 IntegerType::get(SI->getContext(), 2274 ElementSizeBits)); 2275 Value *DestField = NewElts[i]; 2276 if (EltVal->getType() == ArrayEltTy) { 2277 // Storing to an integer field of this size, just do it. 2278 } else if (ArrayEltTy->isFloatingPointTy() || 2279 ArrayEltTy->isVectorTy()) { 2280 // Bitcast to the right element type (for fp/vector values). 2281 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2282 } else { 2283 // Otherwise, bitcast the dest pointer (for aggregates). 2284 DestField = Builder.CreateBitCast(DestField, 2285 PointerType::getUnqual(EltVal->getType())); 2286 } 2287 new StoreInst(EltVal, DestField, SI); 2288 2289 if (TD->isBigEndian()) 2290 Shift -= ElementOffset; 2291 else 2292 Shift += ElementOffset; 2293 } 2294 } 2295 2296 DeadInsts.push_back(SI); 2297} 2298 2299/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2300/// an integer. Load the individual pieces to form the aggregate value. 2301void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2302 SmallVector<AllocaInst*, 32> &NewElts) { 2303 // Extract each element out of the NewElts according to its structure offset 2304 // and form the result value. 2305 Type *AllocaEltTy = AI->getAllocatedType(); 2306 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2307 2308 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2309 << '\n'); 2310 2311 // There are two forms here: AI could be an array or struct. Both cases 2312 // have different ways to compute the element offset. 2313 const StructLayout *Layout = 0; 2314 uint64_t ArrayEltBitOffset = 0; 2315 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2316 Layout = TD->getStructLayout(EltSTy); 2317 } else { 2318 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2319 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2320 } 2321 2322 Value *ResultVal = 2323 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2324 2325 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2326 // Load the value from the alloca. If the NewElt is an aggregate, cast 2327 // the pointer to an integer of the same size before doing the load. 2328 Value *SrcField = NewElts[i]; 2329 Type *FieldTy = 2330 cast<PointerType>(SrcField->getType())->getElementType(); 2331 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2332 2333 // Ignore zero sized fields like {}, they obviously contain no data. 2334 if (FieldSizeBits == 0) continue; 2335 2336 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2337 FieldSizeBits); 2338 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2339 !FieldTy->isVectorTy()) 2340 SrcField = new BitCastInst(SrcField, 2341 PointerType::getUnqual(FieldIntTy), 2342 "", LI); 2343 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2344 2345 // If SrcField is a fp or vector of the right size but that isn't an 2346 // integer type, bitcast to an integer so we can shift it. 2347 if (SrcField->getType() != FieldIntTy) 2348 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2349 2350 // Zero extend the field to be the same size as the final alloca so that 2351 // we can shift and insert it. 2352 if (SrcField->getType() != ResultVal->getType()) 2353 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2354 2355 // Determine the number of bits to shift SrcField. 2356 uint64_t Shift; 2357 if (Layout) // Struct case. 2358 Shift = Layout->getElementOffsetInBits(i); 2359 else // Array case. 2360 Shift = i*ArrayEltBitOffset; 2361 2362 if (TD->isBigEndian()) 2363 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2364 2365 if (Shift) { 2366 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2367 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2368 } 2369 2370 // Don't create an 'or x, 0' on the first iteration. 2371 if (!isa<Constant>(ResultVal) || 2372 !cast<Constant>(ResultVal)->isNullValue()) 2373 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2374 else 2375 ResultVal = SrcField; 2376 } 2377 2378 // Handle tail padding by truncating the result 2379 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2380 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2381 2382 LI->replaceAllUsesWith(ResultVal); 2383 DeadInsts.push_back(LI); 2384} 2385 2386/// HasPadding - Return true if the specified type has any structure or 2387/// alignment padding in between the elements that would be split apart 2388/// by SROA; return false otherwise. 2389static bool HasPadding(Type *Ty, const TargetData &TD) { 2390 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2391 Ty = ATy->getElementType(); 2392 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2393 } 2394 2395 // SROA currently handles only Arrays and Structs. 2396 StructType *STy = cast<StructType>(Ty); 2397 const StructLayout *SL = TD.getStructLayout(STy); 2398 unsigned PrevFieldBitOffset = 0; 2399 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2400 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2401 2402 // Check to see if there is any padding between this element and the 2403 // previous one. 2404 if (i) { 2405 unsigned PrevFieldEnd = 2406 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2407 if (PrevFieldEnd < FieldBitOffset) 2408 return true; 2409 } 2410 PrevFieldBitOffset = FieldBitOffset; 2411 } 2412 // Check for tail padding. 2413 if (unsigned EltCount = STy->getNumElements()) { 2414 unsigned PrevFieldEnd = PrevFieldBitOffset + 2415 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2416 if (PrevFieldEnd < SL->getSizeInBits()) 2417 return true; 2418 } 2419 return false; 2420} 2421 2422/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2423/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2424/// or 1 if safe after canonicalization has been performed. 2425bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2426 // Loop over the use list of the alloca. We can only transform it if all of 2427 // the users are safe to transform. 2428 AllocaInfo Info(AI); 2429 2430 isSafeForScalarRepl(AI, 0, Info); 2431 if (Info.isUnsafe) { 2432 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2433 return false; 2434 } 2435 2436 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2437 // source and destination, we have to be careful. In particular, the memcpy 2438 // could be moving around elements that live in structure padding of the LLVM 2439 // types, but may actually be used. In these cases, we refuse to promote the 2440 // struct. 2441 if (Info.isMemCpySrc && Info.isMemCpyDst && 2442 HasPadding(AI->getAllocatedType(), *TD)) 2443 return false; 2444 2445 // If the alloca never has an access to just *part* of it, but is accessed 2446 // via loads and stores, then we should use ConvertToScalarInfo to promote 2447 // the alloca instead of promoting each piece at a time and inserting fission 2448 // and fusion code. 2449 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2450 // If the struct/array just has one element, use basic SRoA. 2451 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2452 if (ST->getNumElements() > 1) return false; 2453 } else { 2454 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2455 return false; 2456 } 2457 } 2458 2459 return true; 2460} 2461 2462 2463 2464/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2465/// some part of a constant global variable. This intentionally only accepts 2466/// constant expressions because we don't can't rewrite arbitrary instructions. 2467static bool PointsToConstantGlobal(Value *V) { 2468 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2469 return GV->isConstant(); 2470 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2471 if (CE->getOpcode() == Instruction::BitCast || 2472 CE->getOpcode() == Instruction::GetElementPtr) 2473 return PointsToConstantGlobal(CE->getOperand(0)); 2474 return false; 2475} 2476 2477/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2478/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2479/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2480/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2481/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2482/// the alloca, and if the source pointer is a pointer to a constant global, we 2483/// can optimize this. 2484static bool 2485isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2486 bool isOffset, 2487 SmallVector<Instruction *, 4> &LifetimeMarkers) { 2488 // We track lifetime intrinsics as we encounter them. If we decide to go 2489 // ahead and replace the value with the global, this lets the caller quickly 2490 // eliminate the markers. 2491 2492 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2493 User *U = cast<Instruction>(*UI); 2494 2495 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2496 // Ignore non-volatile loads, they are always ok. 2497 if (!LI->isSimple()) return false; 2498 continue; 2499 } 2500 2501 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2502 // If uses of the bitcast are ok, we are ok. 2503 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset, 2504 LifetimeMarkers)) 2505 return false; 2506 continue; 2507 } 2508 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2509 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2510 // doesn't, it does. 2511 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2512 isOffset || !GEP->hasAllZeroIndices(), 2513 LifetimeMarkers)) 2514 return false; 2515 continue; 2516 } 2517 2518 if (CallSite CS = U) { 2519 // If this is the function being called then we treat it like a load and 2520 // ignore it. 2521 if (CS.isCallee(UI)) 2522 continue; 2523 2524 // If this is a readonly/readnone call site, then we know it is just a 2525 // load (but one that potentially returns the value itself), so we can 2526 // ignore it if we know that the value isn't captured. 2527 unsigned ArgNo = CS.getArgumentNo(UI); 2528 if (CS.onlyReadsMemory() && 2529 (CS.getInstruction()->use_empty() || 2530 CS.paramHasAttr(ArgNo+1, Attribute::NoCapture))) 2531 continue; 2532 2533 // If this is being passed as a byval argument, the caller is making a 2534 // copy, so it is only a read of the alloca. 2535 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2536 continue; 2537 } 2538 2539 // Lifetime intrinsics can be handled by the caller. 2540 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 2541 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 2542 II->getIntrinsicID() == Intrinsic::lifetime_end) { 2543 assert(II->use_empty() && "Lifetime markers have no result to use!"); 2544 LifetimeMarkers.push_back(II); 2545 continue; 2546 } 2547 } 2548 2549 // If this is isn't our memcpy/memmove, reject it as something we can't 2550 // handle. 2551 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2552 if (MI == 0) 2553 return false; 2554 2555 // If the transfer is using the alloca as a source of the transfer, then 2556 // ignore it since it is a load (unless the transfer is volatile). 2557 if (UI.getOperandNo() == 1) { 2558 if (MI->isVolatile()) return false; 2559 continue; 2560 } 2561 2562 // If we already have seen a copy, reject the second one. 2563 if (TheCopy) return false; 2564 2565 // If the pointer has been offset from the start of the alloca, we can't 2566 // safely handle this. 2567 if (isOffset) return false; 2568 2569 // If the memintrinsic isn't using the alloca as the dest, reject it. 2570 if (UI.getOperandNo() != 0) return false; 2571 2572 // If the source of the memcpy/move is not a constant global, reject it. 2573 if (!PointsToConstantGlobal(MI->getSource())) 2574 return false; 2575 2576 // Otherwise, the transform is safe. Remember the copy instruction. 2577 TheCopy = MI; 2578 } 2579 return true; 2580} 2581 2582/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2583/// modified by a copy from a constant global. If we can prove this, we can 2584/// replace any uses of the alloca with uses of the global directly. 2585MemTransferInst * 2586SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI, 2587 SmallVector<Instruction*, 4> &ToDelete) { 2588 MemTransferInst *TheCopy = 0; 2589 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete)) 2590 return TheCopy; 2591 return 0; 2592} 2593