ScalarReplAggregates.cpp revision 6399b7c51076846942a0c05b4823ca9a8f55b5fc
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 they 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/Operator.h" 33#include "llvm/Pass.h" 34#include "llvm/Analysis/DebugInfo.h" 35#include "llvm/Analysis/DIBuilder.h" 36#include "llvm/Analysis/Dominators.h" 37#include "llvm/Analysis/Loads.h" 38#include "llvm/Analysis/ValueTracking.h" 39#include "llvm/Target/TargetData.h" 40#include "llvm/Transforms/Utils/PromoteMemToReg.h" 41#include "llvm/Transforms/Utils/Local.h" 42#include "llvm/Transforms/Utils/SSAUpdater.h" 43#include "llvm/Support/CallSite.h" 44#include "llvm/Support/Debug.h" 45#include "llvm/Support/ErrorHandling.h" 46#include "llvm/Support/GetElementPtrTypeIterator.h" 47#include "llvm/Support/IRBuilder.h" 48#include "llvm/Support/MathExtras.h" 49#include "llvm/Support/raw_ostream.h" 50#include "llvm/ADT/SetVector.h" 51#include "llvm/ADT/SmallVector.h" 52#include "llvm/ADT/Statistic.h" 53using namespace llvm; 54 55STATISTIC(NumReplaced, "Number of allocas broken up"); 56STATISTIC(NumPromoted, "Number of allocas promoted"); 57STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); 58STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 59STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 60 61namespace { 62 struct SROA : public FunctionPass { 63 SROA(int T, bool hasDT, char &ID) 64 : FunctionPass(ID), HasDomTree(hasDT) { 65 if (T == -1) 66 SRThreshold = 128; 67 else 68 SRThreshold = T; 69 } 70 71 bool runOnFunction(Function &F); 72 73 bool performScalarRepl(Function &F); 74 bool performPromotion(Function &F); 75 76 private: 77 bool HasDomTree; 78 TargetData *TD; 79 80 /// DeadInsts - Keep track of instructions we have made dead, so that 81 /// we can remove them after we are done working. 82 SmallVector<Value*, 32> DeadInsts; 83 84 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 85 /// information about the uses. All these fields are initialized to false 86 /// and set to true when something is learned. 87 struct AllocaInfo { 88 /// The alloca to promote. 89 AllocaInst *AI; 90 91 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite 92 /// looping and avoid redundant work. 93 SmallPtrSet<PHINode*, 8> CheckedPHIs; 94 95 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 96 bool isUnsafe : 1; 97 98 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 99 bool isMemCpySrc : 1; 100 101 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 102 bool isMemCpyDst : 1; 103 104 /// hasSubelementAccess - This is true if a subelement of the alloca is 105 /// ever accessed, or false if the alloca is only accessed with mem 106 /// intrinsics or load/store that only access the entire alloca at once. 107 bool hasSubelementAccess : 1; 108 109 /// hasALoadOrStore - This is true if there are any loads or stores to it. 110 /// The alloca may just be accessed with memcpy, for example, which would 111 /// not set this. 112 bool hasALoadOrStore : 1; 113 114 explicit AllocaInfo(AllocaInst *ai) 115 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 116 hasSubelementAccess(false), hasALoadOrStore(false) {} 117 }; 118 119 unsigned SRThreshold; 120 121 void MarkUnsafe(AllocaInfo &I, Instruction *User) { 122 I.isUnsafe = true; 123 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); 124 } 125 126 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 127 128 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info); 129 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset, 130 AllocaInfo &Info); 131 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info); 132 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 133 Type *MemOpType, bool isStore, AllocaInfo &Info, 134 Instruction *TheAccess, bool AllowWholeAccess); 135 bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size); 136 uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset, 137 Type *&IdxTy); 138 139 void DoScalarReplacement(AllocaInst *AI, 140 std::vector<AllocaInst*> &WorkList); 141 void DeleteDeadInstructions(); 142 143 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 144 SmallVector<AllocaInst*, 32> &NewElts); 145 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 146 SmallVector<AllocaInst*, 32> &NewElts); 147 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 148 SmallVector<AllocaInst*, 32> &NewElts); 149 void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 150 uint64_t Offset, 151 SmallVector<AllocaInst*, 32> &NewElts); 152 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 153 AllocaInst *AI, 154 SmallVector<AllocaInst*, 32> &NewElts); 155 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 156 SmallVector<AllocaInst*, 32> &NewElts); 157 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 158 SmallVector<AllocaInst*, 32> &NewElts); 159 160 static MemTransferInst *isOnlyCopiedFromConstantGlobal( 161 AllocaInst *AI, SmallVector<Instruction*, 4> &ToDelete); 162 }; 163 164 // SROA_DT - SROA that uses DominatorTree. 165 struct SROA_DT : public SROA { 166 static char ID; 167 public: 168 SROA_DT(int T = -1) : SROA(T, true, ID) { 169 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 170 } 171 172 // getAnalysisUsage - This pass does not require any passes, but we know it 173 // will not alter the CFG, so say so. 174 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 175 AU.addRequired<DominatorTree>(); 176 AU.setPreservesCFG(); 177 } 178 }; 179 180 // SROA_SSAUp - SROA that uses SSAUpdater. 181 struct SROA_SSAUp : public SROA { 182 static char ID; 183 public: 184 SROA_SSAUp(int T = -1) : SROA(T, false, ID) { 185 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 186 } 187 188 // getAnalysisUsage - This pass does not require any passes, but we know it 189 // will not alter the CFG, so say so. 190 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 191 AU.setPreservesCFG(); 192 } 193 }; 194 195} 196 197char SROA_DT::ID = 0; 198char SROA_SSAUp::ID = 0; 199 200INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 201 "Scalar Replacement of Aggregates (DT)", false, false) 202INITIALIZE_PASS_DEPENDENCY(DominatorTree) 203INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 204 "Scalar Replacement of Aggregates (DT)", false, false) 205 206INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 207 "Scalar Replacement of Aggregates (SSAUp)", false, false) 208INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 209 "Scalar Replacement of Aggregates (SSAUp)", false, false) 210 211// Public interface to the ScalarReplAggregates pass 212FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 213 bool UseDomTree) { 214 if (UseDomTree) 215 return new SROA_DT(Threshold); 216 return new SROA_SSAUp(Threshold); 217} 218 219 220//===----------------------------------------------------------------------===// 221// Convert To Scalar Optimization. 222//===----------------------------------------------------------------------===// 223 224namespace { 225/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 226/// optimization, which scans the uses of an alloca and determines if it can 227/// rewrite it in terms of a single new alloca that can be mem2reg'd. 228class ConvertToScalarInfo { 229 /// AllocaSize - The size of the alloca being considered in bytes. 230 unsigned AllocaSize; 231 const TargetData &TD; 232 233 /// IsNotTrivial - This is set to true if there is some access to the object 234 /// which means that mem2reg can't promote it. 235 bool IsNotTrivial; 236 237 /// ScalarKind - Tracks the kind of alloca being considered for promotion, 238 /// computed based on the uses of the alloca rather than the LLVM type system. 239 enum { 240 Unknown, 241 242 // Accesses via GEPs that are consistent with element access of a vector 243 // type. This will not be converted into a vector unless there is a later 244 // access using an actual vector type. 245 ImplicitVector, 246 247 // Accesses via vector operations and GEPs that are consistent with the 248 // layout of a vector type. 249 Vector, 250 251 // An integer bag-of-bits with bitwise operations for insertion and 252 // extraction. Any combination of types can be converted into this kind 253 // of scalar. 254 Integer 255 } ScalarKind; 256 257 /// VectorTy - This tracks the type that we should promote the vector to if 258 /// it is possible to turn it into a vector. This starts out null, and if it 259 /// isn't possible to turn into a vector type, it gets set to VoidTy. 260 VectorType *VectorTy; 261 262 /// HadNonMemTransferAccess - True if there is at least one access to the 263 /// alloca that is not a MemTransferInst. We don't want to turn structs into 264 /// large integers unless there is some potential for optimization. 265 bool HadNonMemTransferAccess; 266 267public: 268 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 269 : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown), 270 VectorTy(0), HadNonMemTransferAccess(false) { } 271 272 AllocaInst *TryConvert(AllocaInst *AI); 273 274private: 275 bool CanConvertToScalar(Value *V, uint64_t Offset); 276 void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset); 277 bool MergeInVectorType(VectorType *VInTy, uint64_t Offset); 278 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 279 280 Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType, 281 uint64_t Offset, IRBuilder<> &Builder); 282 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 283 uint64_t Offset, IRBuilder<> &Builder); 284}; 285} // end anonymous namespace. 286 287 288/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 289/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 290/// alloca if possible or null if not. 291AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 292 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 293 // out. 294 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 295 return 0; 296 297 // If an alloca has only memset / memcpy uses, it may still have an Unknown 298 // ScalarKind. Treat it as an Integer below. 299 if (ScalarKind == Unknown) 300 ScalarKind = Integer; 301 302 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8) 303 ScalarKind = Integer; 304 305 // If we were able to find a vector type that can handle this with 306 // insert/extract elements, and if there was at least one use that had 307 // a vector type, promote this to a vector. We don't want to promote 308 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 309 // we just get a lot of insert/extracts. If at least one vector is 310 // involved, then we probably really do have a union of vector/array. 311 Type *NewTy; 312 if (ScalarKind == Vector) { 313 assert(VectorTy && "Missing type for vector scalar."); 314 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 315 << *VectorTy << '\n'); 316 NewTy = VectorTy; // Use the vector type. 317 } else { 318 unsigned BitWidth = AllocaSize * 8; 319 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) && 320 !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth)) 321 return 0; 322 323 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 324 // Create and insert the integer alloca. 325 NewTy = IntegerType::get(AI->getContext(), BitWidth); 326 } 327 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 328 ConvertUsesToScalar(AI, NewAI, 0); 329 return NewAI; 330} 331 332/// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type 333/// (VectorTy) so far at the offset specified by Offset (which is specified in 334/// bytes). 335/// 336/// There are two cases we handle here: 337/// 1) A union of vector types of the same size and potentially its elements. 338/// Here we turn element accesses into insert/extract element operations. 339/// This promotes a <4 x float> with a store of float to the third element 340/// into a <4 x float> that uses insert element. 341/// 2) A fully general blob of memory, which we turn into some (potentially 342/// large) integer type with extract and insert operations where the loads 343/// and stores would mutate the memory. We mark this by setting VectorTy 344/// to VoidTy. 345void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In, 346 uint64_t Offset) { 347 // If we already decided to turn this into a blob of integer memory, there is 348 // nothing to be done. 349 if (ScalarKind == Integer) 350 return; 351 352 // If this could be contributing to a vector, analyze it. 353 354 // If the In type is a vector that is the same size as the alloca, see if it 355 // matches the existing VecTy. 356 if (VectorType *VInTy = dyn_cast<VectorType>(In)) { 357 if (MergeInVectorType(VInTy, Offset)) 358 return; 359 } else if (In->isFloatTy() || In->isDoubleTy() || 360 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 361 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 362 // Full width accesses can be ignored, because they can always be turned 363 // into bitcasts. 364 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 365 if (EltSize == AllocaSize) 366 return; 367 368 // If we're accessing something that could be an element of a vector, see 369 // if the implied vector agrees with what we already have and if Offset is 370 // compatible with it. 371 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 372 (!VectorTy || EltSize == VectorTy->getElementType() 373 ->getPrimitiveSizeInBits()/8)) { 374 if (!VectorTy) { 375 ScalarKind = ImplicitVector; 376 VectorTy = VectorType::get(In, AllocaSize/EltSize); 377 } 378 return; 379 } 380 } 381 382 // Otherwise, we have a case that we can't handle with an optimized vector 383 // form. We can still turn this into a large integer. 384 ScalarKind = Integer; 385} 386 387/// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore, 388/// returning true if the type was successfully merged and false otherwise. 389bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy, 390 uint64_t Offset) { 391 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 392 // If we're storing/loading a vector of the right size, allow it as a 393 // vector. If this the first vector we see, remember the type so that 394 // we know the element size. If this is a subsequent access, ignore it 395 // even if it is a differing type but the same size. Worst case we can 396 // bitcast the resultant vectors. 397 if (!VectorTy) 398 VectorTy = VInTy; 399 ScalarKind = Vector; 400 return true; 401 } 402 403 return false; 404} 405 406/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 407/// its accesses to a single vector type, return true and set VecTy to 408/// the new type. If we could convert the alloca into a single promotable 409/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 410/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 411/// is the current offset from the base of the alloca being analyzed. 412/// 413/// If we see at least one access to the value that is as a vector type, set the 414/// SawVec flag. 415bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 416 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 417 Instruction *User = cast<Instruction>(*UI); 418 419 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 420 // Don't break volatile loads. 421 if (!LI->isSimple()) 422 return false; 423 // Don't touch MMX operations. 424 if (LI->getType()->isX86_MMXTy()) 425 return false; 426 HadNonMemTransferAccess = true; 427 MergeInTypeForLoadOrStore(LI->getType(), Offset); 428 continue; 429 } 430 431 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 432 // Storing the pointer, not into the value? 433 if (SI->getOperand(0) == V || !SI->isSimple()) return false; 434 // Don't touch MMX operations. 435 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 436 return false; 437 HadNonMemTransferAccess = true; 438 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset); 439 continue; 440 } 441 442 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 443 if (!onlyUsedByLifetimeMarkers(BCI)) 444 IsNotTrivial = true; // Can't be mem2reg'd. 445 if (!CanConvertToScalar(BCI, Offset)) 446 return false; 447 continue; 448 } 449 450 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 451 // If this is a GEP with a variable indices, we can't handle it. 452 if (!GEP->hasAllConstantIndices()) 453 return false; 454 455 // Compute the offset that this GEP adds to the pointer. 456 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 457 if (!GEP->getPointerOperandType()->isPointerTy()) 458 return false; 459 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 460 Indices); 461 // See if all uses can be converted. 462 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 463 return false; 464 IsNotTrivial = true; // Can't be mem2reg'd. 465 HadNonMemTransferAccess = true; 466 continue; 467 } 468 469 // If this is a constant sized memset of a constant value (e.g. 0) we can 470 // handle it. 471 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 472 // Store of constant value. 473 if (!isa<ConstantInt>(MSI->getValue())) 474 return false; 475 476 // Store of constant size. 477 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength()); 478 if (!Len) 479 return false; 480 481 // If the size differs from the alloca, we can only convert the alloca to 482 // an integer bag-of-bits. 483 // FIXME: This should handle all of the cases that are currently accepted 484 // as vector element insertions. 485 if (Len->getZExtValue() != AllocaSize || Offset != 0) 486 ScalarKind = Integer; 487 488 IsNotTrivial = true; // Can't be mem2reg'd. 489 HadNonMemTransferAccess = true; 490 continue; 491 } 492 493 // If this is a memcpy or memmove into or out of the whole allocation, we 494 // can handle it like a load or store of the scalar type. 495 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 496 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 497 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 498 return false; 499 500 IsNotTrivial = true; // Can't be mem2reg'd. 501 continue; 502 } 503 504 // If this is a lifetime intrinsic, we can handle it. 505 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 506 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 507 II->getIntrinsicID() == Intrinsic::lifetime_end) { 508 continue; 509 } 510 } 511 512 // Otherwise, we cannot handle this! 513 return false; 514 } 515 516 return true; 517} 518 519/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 520/// directly. This happens when we are converting an "integer union" to a 521/// single integer scalar, or when we are converting a "vector union" to a 522/// vector with insert/extractelement instructions. 523/// 524/// Offset is an offset from the original alloca, in bits that need to be 525/// shifted to the right. By the end of this, there should be no uses of Ptr. 526void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 527 uint64_t Offset) { 528 while (!Ptr->use_empty()) { 529 Instruction *User = cast<Instruction>(Ptr->use_back()); 530 531 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 532 ConvertUsesToScalar(CI, NewAI, Offset); 533 CI->eraseFromParent(); 534 continue; 535 } 536 537 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 538 // Compute the offset that this GEP adds to the pointer. 539 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 540 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 541 Indices); 542 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 543 GEP->eraseFromParent(); 544 continue; 545 } 546 547 IRBuilder<> Builder(User); 548 549 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 550 // The load is a bit extract from NewAI shifted right by Offset bits. 551 Value *LoadedVal = Builder.CreateLoad(NewAI); 552 Value *NewLoadVal 553 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 554 LI->replaceAllUsesWith(NewLoadVal); 555 LI->eraseFromParent(); 556 continue; 557 } 558 559 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 560 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 561 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 562 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 563 Builder); 564 Builder.CreateStore(New, NewAI); 565 SI->eraseFromParent(); 566 567 // If the load we just inserted is now dead, then the inserted store 568 // overwrote the entire thing. 569 if (Old->use_empty()) 570 Old->eraseFromParent(); 571 continue; 572 } 573 574 // If this is a constant sized memset of a constant value (e.g. 0) we can 575 // transform it into a store of the expanded constant value. 576 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 577 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 578 int64_t SNumBytes = cast<ConstantInt>(MSI->getLength())->getSExtValue(); 579 if (SNumBytes > 0 && (SNumBytes >> 32) == 0) { 580 unsigned NumBytes = static_cast<unsigned>(SNumBytes); 581 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 582 583 // Compute the value replicated the right number of times. 584 APInt APVal(NumBytes*8, Val); 585 586 // Splat the value if non-zero. 587 if (Val) 588 for (unsigned i = 1; i != NumBytes; ++i) 589 APVal |= APVal << 8; 590 591 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 592 Value *New = ConvertScalar_InsertValue( 593 ConstantInt::get(User->getContext(), APVal), 594 Old, Offset, Builder); 595 Builder.CreateStore(New, NewAI); 596 597 // If the load we just inserted is now dead, then the memset overwrote 598 // the entire thing. 599 if (Old->use_empty()) 600 Old->eraseFromParent(); 601 } 602 MSI->eraseFromParent(); 603 continue; 604 } 605 606 // If this is a memcpy or memmove into or out of the whole allocation, we 607 // can handle it like a load or store of the scalar type. 608 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 609 assert(Offset == 0 && "must be store to start of alloca"); 610 611 // If the source and destination are both to the same alloca, then this is 612 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 613 // as appropriate. 614 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 615 616 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 617 // Dest must be OrigAI, change this to be a load from the original 618 // pointer (bitcasted), then a store to our new alloca. 619 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 620 Value *SrcPtr = MTI->getSource(); 621 PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 622 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 623 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 624 AIPTy = PointerType::get(AIPTy->getElementType(), 625 SPTy->getAddressSpace()); 626 } 627 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 628 629 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 630 SrcVal->setAlignment(MTI->getAlignment()); 631 Builder.CreateStore(SrcVal, NewAI); 632 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 633 // Src must be OrigAI, change this to be a load from NewAI then a store 634 // through the original dest pointer (bitcasted). 635 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 636 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 637 638 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 639 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 640 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 641 AIPTy = PointerType::get(AIPTy->getElementType(), 642 DPTy->getAddressSpace()); 643 } 644 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 645 646 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 647 NewStore->setAlignment(MTI->getAlignment()); 648 } else { 649 // Noop transfer. Src == Dst 650 } 651 652 MTI->eraseFromParent(); 653 continue; 654 } 655 656 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 657 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 658 II->getIntrinsicID() == Intrinsic::lifetime_end) { 659 // There's no need to preserve these, as the resulting alloca will be 660 // converted to a register anyways. 661 II->eraseFromParent(); 662 continue; 663 } 664 } 665 666 llvm_unreachable("Unsupported operation!"); 667 } 668} 669 670/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 671/// or vector value FromVal, extracting the bits from the offset specified by 672/// Offset. This returns the value, which is of type ToType. 673/// 674/// This happens when we are converting an "integer union" to a single 675/// integer scalar, or when we are converting a "vector union" to a vector with 676/// insert/extractelement instructions. 677/// 678/// Offset is an offset from the original alloca, in bits that need to be 679/// shifted to the right. 680Value *ConvertToScalarInfo:: 681ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, 682 uint64_t Offset, IRBuilder<> &Builder) { 683 // If the load is of the whole new alloca, no conversion is needed. 684 Type *FromType = FromVal->getType(); 685 if (FromType == ToType && Offset == 0) 686 return FromVal; 687 688 // If the result alloca is a vector type, this is either an element 689 // access or a bitcast to another vector type of the same size. 690 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) { 691 unsigned FromTypeSize = TD.getTypeAllocSize(FromType); 692 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 693 if (FromTypeSize == ToTypeSize) 694 return Builder.CreateBitCast(FromVal, ToType); 695 696 // Otherwise it must be an element access. 697 unsigned Elt = 0; 698 if (Offset) { 699 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 700 Elt = Offset/EltSize; 701 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 702 } 703 // Return the element extracted out of it. 704 Value *V = Builder.CreateExtractElement(FromVal, Builder.getInt32(Elt)); 705 if (V->getType() != ToType) 706 V = Builder.CreateBitCast(V, ToType); 707 return V; 708 } 709 710 // If ToType is a first class aggregate, extract out each of the pieces and 711 // use insertvalue's to form the FCA. 712 if (StructType *ST = dyn_cast<StructType>(ToType)) { 713 const StructLayout &Layout = *TD.getStructLayout(ST); 714 Value *Res = UndefValue::get(ST); 715 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 716 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 717 Offset+Layout.getElementOffsetInBits(i), 718 Builder); 719 Res = Builder.CreateInsertValue(Res, Elt, i); 720 } 721 return Res; 722 } 723 724 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 725 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 726 Value *Res = UndefValue::get(AT); 727 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 728 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 729 Offset+i*EltSize, Builder); 730 Res = Builder.CreateInsertValue(Res, Elt, i); 731 } 732 return Res; 733 } 734 735 // Otherwise, this must be a union that was converted to an integer value. 736 IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 737 738 // If this is a big-endian system and the load is narrower than the 739 // full alloca type, we need to do a shift to get the right bits. 740 int ShAmt = 0; 741 if (TD.isBigEndian()) { 742 // On big-endian machines, the lowest bit is stored at the bit offset 743 // from the pointer given by getTypeStoreSizeInBits. This matters for 744 // integers with a bitwidth that is not a multiple of 8. 745 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 746 TD.getTypeStoreSizeInBits(ToType) - Offset; 747 } else { 748 ShAmt = Offset; 749 } 750 751 // Note: we support negative bitwidths (with shl) which are not defined. 752 // We do this to support (f.e.) loads off the end of a structure where 753 // only some bits are used. 754 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 755 FromVal = Builder.CreateLShr(FromVal, 756 ConstantInt::get(FromVal->getType(), ShAmt)); 757 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 758 FromVal = Builder.CreateShl(FromVal, 759 ConstantInt::get(FromVal->getType(), -ShAmt)); 760 761 // Finally, unconditionally truncate the integer to the right width. 762 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 763 if (LIBitWidth < NTy->getBitWidth()) 764 FromVal = 765 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 766 LIBitWidth)); 767 else if (LIBitWidth > NTy->getBitWidth()) 768 FromVal = 769 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 770 LIBitWidth)); 771 772 // If the result is an integer, this is a trunc or bitcast. 773 if (ToType->isIntegerTy()) { 774 // Should be done. 775 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 776 // Just do a bitcast, we know the sizes match up. 777 FromVal = Builder.CreateBitCast(FromVal, ToType); 778 } else { 779 // Otherwise must be a pointer. 780 FromVal = Builder.CreateIntToPtr(FromVal, ToType); 781 } 782 assert(FromVal->getType() == ToType && "Didn't convert right?"); 783 return FromVal; 784} 785 786/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 787/// or vector value "Old" at the offset specified by Offset. 788/// 789/// This happens when we are converting an "integer union" to a 790/// single integer scalar, or when we are converting a "vector union" to a 791/// vector with insert/extractelement instructions. 792/// 793/// Offset is an offset from the original alloca, in bits that need to be 794/// shifted to the right. 795Value *ConvertToScalarInfo:: 796ConvertScalar_InsertValue(Value *SV, Value *Old, 797 uint64_t Offset, IRBuilder<> &Builder) { 798 // Convert the stored type to the actual type, shift it left to insert 799 // then 'or' into place. 800 Type *AllocaType = Old->getType(); 801 LLVMContext &Context = Old->getContext(); 802 803 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 804 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 805 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 806 807 // Changing the whole vector with memset or with an access of a different 808 // vector type? 809 if (ValSize == VecSize) 810 return Builder.CreateBitCast(SV, AllocaType); 811 812 // Must be an element insertion. 813 Type *EltTy = VTy->getElementType(); 814 if (SV->getType() != EltTy) 815 SV = Builder.CreateBitCast(SV, EltTy); 816 uint64_t EltSize = TD.getTypeAllocSizeInBits(EltTy); 817 unsigned Elt = Offset/EltSize; 818 return Builder.CreateInsertElement(Old, SV, Builder.getInt32(Elt)); 819 } 820 821 // If SV is a first-class aggregate value, insert each value recursively. 822 if (StructType *ST = dyn_cast<StructType>(SV->getType())) { 823 const StructLayout &Layout = *TD.getStructLayout(ST); 824 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 825 Value *Elt = Builder.CreateExtractValue(SV, i); 826 Old = ConvertScalar_InsertValue(Elt, Old, 827 Offset+Layout.getElementOffsetInBits(i), 828 Builder); 829 } 830 return Old; 831 } 832 833 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 834 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 835 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 836 Value *Elt = Builder.CreateExtractValue(SV, i); 837 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 838 } 839 return Old; 840 } 841 842 // If SV is a float, convert it to the appropriate integer type. 843 // If it is a pointer, do the same. 844 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 845 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 846 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 847 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 848 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 849 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth)); 850 else if (SV->getType()->isPointerTy()) 851 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext())); 852 853 // Zero extend or truncate the value if needed. 854 if (SV->getType() != AllocaType) { 855 if (SV->getType()->getPrimitiveSizeInBits() < 856 AllocaType->getPrimitiveSizeInBits()) 857 SV = Builder.CreateZExt(SV, AllocaType); 858 else { 859 // Truncation may be needed if storing more than the alloca can hold 860 // (undefined behavior). 861 SV = Builder.CreateTrunc(SV, AllocaType); 862 SrcWidth = DestWidth; 863 SrcStoreWidth = DestStoreWidth; 864 } 865 } 866 867 // If this is a big-endian system and the store is narrower than the 868 // full alloca type, we need to do a shift to get the right bits. 869 int ShAmt = 0; 870 if (TD.isBigEndian()) { 871 // On big-endian machines, the lowest bit is stored at the bit offset 872 // from the pointer given by getTypeStoreSizeInBits. This matters for 873 // integers with a bitwidth that is not a multiple of 8. 874 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 875 } else { 876 ShAmt = Offset; 877 } 878 879 // Note: we support negative bitwidths (with shr) which are not defined. 880 // We do this to support (f.e.) stores off the end of a structure where 881 // only some bits in the structure are set. 882 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 883 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 884 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt)); 885 Mask <<= ShAmt; 886 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 887 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt)); 888 Mask = Mask.lshr(-ShAmt); 889 } 890 891 // Mask out the bits we are about to insert from the old value, and or 892 // in the new bits. 893 if (SrcWidth != DestWidth) { 894 assert(DestWidth > SrcWidth); 895 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 896 SV = Builder.CreateOr(Old, SV, "ins"); 897 } 898 return SV; 899} 900 901 902//===----------------------------------------------------------------------===// 903// SRoA Driver 904//===----------------------------------------------------------------------===// 905 906 907bool SROA::runOnFunction(Function &F) { 908 TD = getAnalysisIfAvailable<TargetData>(); 909 910 bool Changed = performPromotion(F); 911 912 // FIXME: ScalarRepl currently depends on TargetData more than it 913 // theoretically needs to. It should be refactored in order to support 914 // target-independent IR. Until this is done, just skip the actual 915 // scalar-replacement portion of this pass. 916 if (!TD) return Changed; 917 918 while (1) { 919 bool LocalChange = performScalarRepl(F); 920 if (!LocalChange) break; // No need to repromote if no scalarrepl 921 Changed = true; 922 LocalChange = performPromotion(F); 923 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 924 } 925 926 return Changed; 927} 928 929namespace { 930class AllocaPromoter : public LoadAndStorePromoter { 931 AllocaInst *AI; 932 DIBuilder *DIB; 933 SmallVector<DbgDeclareInst *, 4> DDIs; 934 SmallVector<DbgValueInst *, 4> DVIs; 935public: 936 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 937 DIBuilder *DB) 938 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {} 939 940 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 941 // Remember which alloca we're promoting (for isInstInList). 942 this->AI = AI; 943 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) { 944 for (Value::use_iterator UI = DebugNode->use_begin(), 945 E = DebugNode->use_end(); UI != E; ++UI) 946 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI)) 947 DDIs.push_back(DDI); 948 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI)) 949 DVIs.push_back(DVI); 950 } 951 952 LoadAndStorePromoter::run(Insts); 953 AI->eraseFromParent(); 954 for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(), 955 E = DDIs.end(); I != E; ++I) { 956 DbgDeclareInst *DDI = *I; 957 DDI->eraseFromParent(); 958 } 959 for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(), 960 E = DVIs.end(); I != E; ++I) { 961 DbgValueInst *DVI = *I; 962 DVI->eraseFromParent(); 963 } 964 } 965 966 virtual bool isInstInList(Instruction *I, 967 const SmallVectorImpl<Instruction*> &Insts) const { 968 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 969 return LI->getOperand(0) == AI; 970 return cast<StoreInst>(I)->getPointerOperand() == AI; 971 } 972 973 virtual void updateDebugInfo(Instruction *Inst) const { 974 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(), 975 E = DDIs.end(); I != E; ++I) { 976 DbgDeclareInst *DDI = *I; 977 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 978 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB); 979 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 980 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB); 981 } 982 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(), 983 E = DVIs.end(); I != E; ++I) { 984 DbgValueInst *DVI = *I; 985 Value *Arg = NULL; 986 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 987 // If an argument is zero extended then use argument directly. The ZExt 988 // may be zapped by an optimization pass in future. 989 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 990 Arg = dyn_cast<Argument>(ZExt->getOperand(0)); 991 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 992 Arg = dyn_cast<Argument>(SExt->getOperand(0)); 993 if (!Arg) 994 Arg = SI->getOperand(0); 995 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 996 Arg = LI->getOperand(0); 997 } else { 998 continue; 999 } 1000 Instruction *DbgVal = 1001 DIB->insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()), 1002 Inst); 1003 DbgVal->setDebugLoc(DVI->getDebugLoc()); 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/// getPointeeAlignment - Compute the minimum alignment of the value pointed 1351/// to by the given pointer. 1352static unsigned getPointeeAlignment(Value *V, const TargetData &TD) { 1353 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1354 if (CE->getOpcode() == Instruction::BitCast || 1355 (CE->getOpcode() == Instruction::GetElementPtr && 1356 cast<GEPOperator>(CE)->hasAllZeroIndices())) 1357 return getPointeeAlignment(CE->getOperand(0), TD); 1358 1359 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1360 if (!GV->isDeclaration()) 1361 return TD.getPreferredAlignment(GV); 1362 1363 if (PointerType *PT = dyn_cast<PointerType>(V->getType())) 1364 return TD.getABITypeAlignment(PT->getElementType()); 1365 1366 return 0; 1367} 1368 1369 1370// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1371// which runs on all of the alloca instructions in the function, removing them 1372// if they are only used by getelementptr instructions. 1373// 1374bool SROA::performScalarRepl(Function &F) { 1375 std::vector<AllocaInst*> WorkList; 1376 1377 // Scan the entry basic block, adding allocas to the worklist. 1378 BasicBlock &BB = F.getEntryBlock(); 1379 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1380 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1381 WorkList.push_back(A); 1382 1383 // Process the worklist 1384 bool Changed = false; 1385 while (!WorkList.empty()) { 1386 AllocaInst *AI = WorkList.back(); 1387 WorkList.pop_back(); 1388 1389 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1390 // with unused elements. 1391 if (AI->use_empty()) { 1392 AI->eraseFromParent(); 1393 Changed = true; 1394 continue; 1395 } 1396 1397 // If this alloca is impossible for us to promote, reject it early. 1398 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1399 continue; 1400 1401 // Check to see if this allocation is only modified by a memcpy/memmove from 1402 // a constant global whose alignment is equal to or exceeds that of the 1403 // allocation. If this is the case, we can change all users to use 1404 // the constant global instead. This is commonly produced by the CFE by 1405 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1406 // is only subsequently read. 1407 SmallVector<Instruction *, 4> ToDelete; 1408 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) { 1409 if (AI->getAlignment() <= getPointeeAlignment(Copy->getSource(), *TD)) { 1410 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1411 DEBUG(dbgs() << " memcpy = " << *Copy << '\n'); 1412 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i) 1413 ToDelete[i]->eraseFromParent(); 1414 Constant *TheSrc = cast<Constant>(Copy->getSource()); 1415 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1416 Copy->eraseFromParent(); // Don't mutate the global. 1417 AI->eraseFromParent(); 1418 ++NumGlobals; 1419 Changed = true; 1420 continue; 1421 } 1422 } 1423 1424 // Check to see if we can perform the core SROA transformation. We cannot 1425 // transform the allocation instruction if it is an array allocation 1426 // (allocations OF arrays are ok though), and an allocation of a scalar 1427 // value cannot be decomposed at all. 1428 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1429 1430 // Do not promote [0 x %struct]. 1431 if (AllocaSize == 0) continue; 1432 1433 // Do not promote any struct whose size is too big. 1434 if (AllocaSize > SRThreshold) continue; 1435 1436 // If the alloca looks like a good candidate for scalar replacement, and if 1437 // all its users can be transformed, then split up the aggregate into its 1438 // separate elements. 1439 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1440 DoScalarReplacement(AI, WorkList); 1441 Changed = true; 1442 continue; 1443 } 1444 1445 // If we can turn this aggregate value (potentially with casts) into a 1446 // simple scalar value that can be mem2reg'd into a register value. 1447 // IsNotTrivial tracks whether this is something that mem2reg could have 1448 // promoted itself. If so, we don't want to transform it needlessly. Note 1449 // that we can't just check based on the type: the alloca may be of an i32 1450 // but that has pointer arithmetic to set byte 3 of it or something. 1451 if (AllocaInst *NewAI = 1452 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1453 NewAI->takeName(AI); 1454 AI->eraseFromParent(); 1455 ++NumConverted; 1456 Changed = true; 1457 continue; 1458 } 1459 1460 // Otherwise, couldn't process this alloca. 1461 } 1462 1463 return Changed; 1464} 1465 1466/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1467/// predicate, do SROA now. 1468void SROA::DoScalarReplacement(AllocaInst *AI, 1469 std::vector<AllocaInst*> &WorkList) { 1470 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1471 SmallVector<AllocaInst*, 32> ElementAllocas; 1472 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1473 ElementAllocas.reserve(ST->getNumContainedTypes()); 1474 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1475 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1476 AI->getAlignment(), 1477 AI->getName() + "." + Twine(i), AI); 1478 ElementAllocas.push_back(NA); 1479 WorkList.push_back(NA); // Add to worklist for recursive processing 1480 } 1481 } else { 1482 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1483 ElementAllocas.reserve(AT->getNumElements()); 1484 Type *ElTy = AT->getElementType(); 1485 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1486 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1487 AI->getName() + "." + Twine(i), AI); 1488 ElementAllocas.push_back(NA); 1489 WorkList.push_back(NA); // Add to worklist for recursive processing 1490 } 1491 } 1492 1493 // Now that we have created the new alloca instructions, rewrite all the 1494 // uses of the old alloca. 1495 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1496 1497 // Now erase any instructions that were made dead while rewriting the alloca. 1498 DeleteDeadInstructions(); 1499 AI->eraseFromParent(); 1500 1501 ++NumReplaced; 1502} 1503 1504/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1505/// recursively including all their operands that become trivially dead. 1506void SROA::DeleteDeadInstructions() { 1507 while (!DeadInsts.empty()) { 1508 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1509 1510 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1511 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1512 // Zero out the operand and see if it becomes trivially dead. 1513 // (But, don't add allocas to the dead instruction list -- they are 1514 // already on the worklist and will be deleted separately.) 1515 *OI = 0; 1516 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1517 DeadInsts.push_back(U); 1518 } 1519 1520 I->eraseFromParent(); 1521 } 1522} 1523 1524/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1525/// performing scalar replacement of alloca AI. The results are flagged in 1526/// the Info parameter. Offset indicates the position within AI that is 1527/// referenced by this instruction. 1528void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1529 AllocaInfo &Info) { 1530 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1531 Instruction *User = cast<Instruction>(*UI); 1532 1533 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1534 isSafeForScalarRepl(BC, Offset, Info); 1535 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1536 uint64_t GEPOffset = Offset; 1537 isSafeGEP(GEPI, GEPOffset, Info); 1538 if (!Info.isUnsafe) 1539 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1540 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1541 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1542 if (Length == 0) 1543 return MarkUnsafe(Info, User); 1544 if (Length->isNegative()) 1545 return MarkUnsafe(Info, User); 1546 1547 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1548 UI.getOperandNo() == 0, Info, MI, 1549 true /*AllowWholeAccess*/); 1550 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1551 if (!LI->isSimple()) 1552 return MarkUnsafe(Info, User); 1553 Type *LIType = LI->getType(); 1554 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1555 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1556 Info.hasALoadOrStore = true; 1557 1558 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1559 // Store is ok if storing INTO the pointer, not storing the pointer 1560 if (!SI->isSimple() || SI->getOperand(0) == I) 1561 return MarkUnsafe(Info, User); 1562 1563 Type *SIType = SI->getOperand(0)->getType(); 1564 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1565 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1566 Info.hasALoadOrStore = true; 1567 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1568 if (II->getIntrinsicID() != Intrinsic::lifetime_start && 1569 II->getIntrinsicID() != Intrinsic::lifetime_end) 1570 return MarkUnsafe(Info, User); 1571 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1572 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1573 } else { 1574 return MarkUnsafe(Info, User); 1575 } 1576 if (Info.isUnsafe) return; 1577 } 1578} 1579 1580 1581/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1582/// derived from the alloca, we can often still split the alloca into elements. 1583/// This is useful if we have a large alloca where one element is phi'd 1584/// together somewhere: we can SRoA and promote all the other elements even if 1585/// we end up not being able to promote this one. 1586/// 1587/// All we require is that the uses of the PHI do not index into other parts of 1588/// the alloca. The most important use case for this is single load and stores 1589/// that are PHI'd together, which can happen due to code sinking. 1590void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1591 AllocaInfo &Info) { 1592 // If we've already checked this PHI, don't do it again. 1593 if (PHINode *PN = dyn_cast<PHINode>(I)) 1594 if (!Info.CheckedPHIs.insert(PN)) 1595 return; 1596 1597 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1598 Instruction *User = cast<Instruction>(*UI); 1599 1600 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1601 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1602 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1603 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1604 // but would have to prove that we're staying inside of an element being 1605 // promoted. 1606 if (!GEPI->hasAllZeroIndices()) 1607 return MarkUnsafe(Info, User); 1608 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1609 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1610 if (!LI->isSimple()) 1611 return MarkUnsafe(Info, User); 1612 Type *LIType = LI->getType(); 1613 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1614 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1615 Info.hasALoadOrStore = true; 1616 1617 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1618 // Store is ok if storing INTO the pointer, not storing the pointer 1619 if (!SI->isSimple() || SI->getOperand(0) == I) 1620 return MarkUnsafe(Info, User); 1621 1622 Type *SIType = SI->getOperand(0)->getType(); 1623 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1624 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1625 Info.hasALoadOrStore = true; 1626 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1627 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1628 } else { 1629 return MarkUnsafe(Info, User); 1630 } 1631 if (Info.isUnsafe) return; 1632 } 1633} 1634 1635/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1636/// replacement. It is safe when all the indices are constant, in-bounds 1637/// references, and when the resulting offset corresponds to an element within 1638/// the alloca type. The results are flagged in the Info parameter. Upon 1639/// return, Offset is adjusted as specified by the GEP indices. 1640void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1641 uint64_t &Offset, AllocaInfo &Info) { 1642 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1643 if (GEPIt == E) 1644 return; 1645 1646 // Walk through the GEP type indices, checking the types that this indexes 1647 // into. 1648 for (; GEPIt != E; ++GEPIt) { 1649 // Ignore struct elements, no extra checking needed for these. 1650 if ((*GEPIt)->isStructTy()) 1651 continue; 1652 1653 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1654 if (!IdxVal) 1655 return MarkUnsafe(Info, GEPI); 1656 } 1657 1658 // Compute the offset due to this GEP and check if the alloca has a 1659 // component element at that offset. 1660 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1661 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1662 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1663 MarkUnsafe(Info, GEPI); 1664} 1665 1666/// isHomogeneousAggregate - Check if type T is a struct or array containing 1667/// elements of the same type (which is always true for arrays). If so, 1668/// return true with NumElts and EltTy set to the number of elements and the 1669/// element type, respectively. 1670static bool isHomogeneousAggregate(Type *T, unsigned &NumElts, 1671 Type *&EltTy) { 1672 if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1673 NumElts = AT->getNumElements(); 1674 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1675 return true; 1676 } 1677 if (StructType *ST = dyn_cast<StructType>(T)) { 1678 NumElts = ST->getNumContainedTypes(); 1679 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1680 for (unsigned n = 1; n < NumElts; ++n) { 1681 if (ST->getContainedType(n) != EltTy) 1682 return false; 1683 } 1684 return true; 1685 } 1686 return false; 1687} 1688 1689/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1690/// "homogeneous" aggregates with the same element type and number of elements. 1691static bool isCompatibleAggregate(Type *T1, Type *T2) { 1692 if (T1 == T2) 1693 return true; 1694 1695 unsigned NumElts1, NumElts2; 1696 Type *EltTy1, *EltTy2; 1697 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1698 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1699 NumElts1 == NumElts2 && 1700 EltTy1 == EltTy2) 1701 return true; 1702 1703 return false; 1704} 1705 1706/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1707/// alloca or has an offset and size that corresponds to a component element 1708/// within it. The offset checked here may have been formed from a GEP with a 1709/// pointer bitcasted to a different type. 1710/// 1711/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1712/// unit. If false, it only allows accesses known to be in a single element. 1713void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1714 Type *MemOpType, bool isStore, 1715 AllocaInfo &Info, Instruction *TheAccess, 1716 bool AllowWholeAccess) { 1717 // Check if this is a load/store of the entire alloca. 1718 if (Offset == 0 && AllowWholeAccess && 1719 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1720 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1721 // loads/stores (which are essentially the same as the MemIntrinsics with 1722 // regard to copying padding between elements). But, if an alloca is 1723 // flagged as both a source and destination of such operations, we'll need 1724 // to check later for padding between elements. 1725 if (!MemOpType || MemOpType->isIntegerTy()) { 1726 if (isStore) 1727 Info.isMemCpyDst = true; 1728 else 1729 Info.isMemCpySrc = true; 1730 return; 1731 } 1732 // This is also safe for references using a type that is compatible with 1733 // the type of the alloca, so that loads/stores can be rewritten using 1734 // insertvalue/extractvalue. 1735 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1736 Info.hasSubelementAccess = true; 1737 return; 1738 } 1739 } 1740 // Check if the offset/size correspond to a component within the alloca type. 1741 Type *T = Info.AI->getAllocatedType(); 1742 if (TypeHasComponent(T, Offset, MemSize)) { 1743 Info.hasSubelementAccess = true; 1744 return; 1745 } 1746 1747 return MarkUnsafe(Info, TheAccess); 1748} 1749 1750/// TypeHasComponent - Return true if T has a component type with the 1751/// specified offset and size. If Size is zero, do not check the size. 1752bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { 1753 Type *EltTy; 1754 uint64_t EltSize; 1755 if (StructType *ST = dyn_cast<StructType>(T)) { 1756 const StructLayout *Layout = TD->getStructLayout(ST); 1757 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1758 EltTy = ST->getContainedType(EltIdx); 1759 EltSize = TD->getTypeAllocSize(EltTy); 1760 Offset -= Layout->getElementOffset(EltIdx); 1761 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1762 EltTy = AT->getElementType(); 1763 EltSize = TD->getTypeAllocSize(EltTy); 1764 if (Offset >= AT->getNumElements() * EltSize) 1765 return false; 1766 Offset %= EltSize; 1767 } else if (VectorType *VT = dyn_cast<VectorType>(T)) { 1768 EltTy = VT->getElementType(); 1769 EltSize = TD->getTypeAllocSize(EltTy); 1770 if (Offset >= VT->getNumElements() * EltSize) 1771 return false; 1772 Offset %= EltSize; 1773 } else { 1774 return false; 1775 } 1776 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1777 return true; 1778 // Check if the component spans multiple elements. 1779 if (Offset + Size > EltSize) 1780 return false; 1781 return TypeHasComponent(EltTy, Offset, Size); 1782} 1783 1784/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1785/// the instruction I, which references it, to use the separate elements. 1786/// Offset indicates the position within AI that is referenced by this 1787/// instruction. 1788void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1789 SmallVector<AllocaInst*, 32> &NewElts) { 1790 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1791 Use &TheUse = UI.getUse(); 1792 Instruction *User = cast<Instruction>(*UI++); 1793 1794 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1795 RewriteBitCast(BC, AI, Offset, NewElts); 1796 continue; 1797 } 1798 1799 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1800 RewriteGEP(GEPI, AI, Offset, NewElts); 1801 continue; 1802 } 1803 1804 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1805 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1806 uint64_t MemSize = Length->getZExtValue(); 1807 if (Offset == 0 && 1808 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1809 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1810 // Otherwise the intrinsic can only touch a single element and the 1811 // address operand will be updated, so nothing else needs to be done. 1812 continue; 1813 } 1814 1815 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1816 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 1817 II->getIntrinsicID() == Intrinsic::lifetime_end) { 1818 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts); 1819 } 1820 continue; 1821 } 1822 1823 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1824 Type *LIType = LI->getType(); 1825 1826 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1827 // Replace: 1828 // %res = load { i32, i32 }* %alloc 1829 // with: 1830 // %load.0 = load i32* %alloc.0 1831 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1832 // %load.1 = load i32* %alloc.1 1833 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1834 // (Also works for arrays instead of structs) 1835 Value *Insert = UndefValue::get(LIType); 1836 IRBuilder<> Builder(LI); 1837 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1838 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1839 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1840 } 1841 LI->replaceAllUsesWith(Insert); 1842 DeadInsts.push_back(LI); 1843 } else if (LIType->isIntegerTy() && 1844 TD->getTypeAllocSize(LIType) == 1845 TD->getTypeAllocSize(AI->getAllocatedType())) { 1846 // If this is a load of the entire alloca to an integer, rewrite it. 1847 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1848 } 1849 continue; 1850 } 1851 1852 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1853 Value *Val = SI->getOperand(0); 1854 Type *SIType = Val->getType(); 1855 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1856 // Replace: 1857 // store { i32, i32 } %val, { i32, i32 }* %alloc 1858 // with: 1859 // %val.0 = extractvalue { i32, i32 } %val, 0 1860 // store i32 %val.0, i32* %alloc.0 1861 // %val.1 = extractvalue { i32, i32 } %val, 1 1862 // store i32 %val.1, i32* %alloc.1 1863 // (Also works for arrays instead of structs) 1864 IRBuilder<> Builder(SI); 1865 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1866 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1867 Builder.CreateStore(Extract, NewElts[i]); 1868 } 1869 DeadInsts.push_back(SI); 1870 } else if (SIType->isIntegerTy() && 1871 TD->getTypeAllocSize(SIType) == 1872 TD->getTypeAllocSize(AI->getAllocatedType())) { 1873 // If this is a store of the entire alloca from an integer, rewrite it. 1874 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1875 } 1876 continue; 1877 } 1878 1879 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1880 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1881 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1882 // the new pointer. 1883 if (!isa<AllocaInst>(I)) continue; 1884 1885 assert(Offset == 0 && NewElts[0] && 1886 "Direct alloca use should have a zero offset"); 1887 1888 // If we have a use of the alloca, we know the derived uses will be 1889 // utilizing just the first element of the scalarized result. Insert a 1890 // bitcast of the first alloca before the user as required. 1891 AllocaInst *NewAI = NewElts[0]; 1892 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1893 NewAI->moveBefore(BCI); 1894 TheUse = BCI; 1895 continue; 1896 } 1897 } 1898} 1899 1900/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1901/// and recursively continue updating all of its uses. 1902void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1903 SmallVector<AllocaInst*, 32> &NewElts) { 1904 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1905 if (BC->getOperand(0) != AI) 1906 return; 1907 1908 // The bitcast references the original alloca. Replace its uses with 1909 // references to the alloca containing offset zero (which is normally at 1910 // index zero, but might not be in cases involving structs with elements 1911 // of size zero). 1912 Type *T = AI->getAllocatedType(); 1913 uint64_t EltOffset = 0; 1914 Type *IdxTy; 1915 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1916 Instruction *Val = NewElts[Idx]; 1917 if (Val->getType() != BC->getDestTy()) { 1918 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1919 Val->takeName(BC); 1920 } 1921 BC->replaceAllUsesWith(Val); 1922 DeadInsts.push_back(BC); 1923} 1924 1925/// FindElementAndOffset - Return the index of the element containing Offset 1926/// within the specified type, which must be either a struct or an array. 1927/// Sets T to the type of the element and Offset to the offset within that 1928/// element. IdxTy is set to the type of the index result to be used in a 1929/// GEP instruction. 1930uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, 1931 Type *&IdxTy) { 1932 uint64_t Idx = 0; 1933 if (StructType *ST = dyn_cast<StructType>(T)) { 1934 const StructLayout *Layout = TD->getStructLayout(ST); 1935 Idx = Layout->getElementContainingOffset(Offset); 1936 T = ST->getContainedType(Idx); 1937 Offset -= Layout->getElementOffset(Idx); 1938 IdxTy = Type::getInt32Ty(T->getContext()); 1939 return Idx; 1940 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1941 T = AT->getElementType(); 1942 uint64_t EltSize = TD->getTypeAllocSize(T); 1943 Idx = Offset / EltSize; 1944 Offset -= Idx * EltSize; 1945 IdxTy = Type::getInt64Ty(T->getContext()); 1946 return Idx; 1947 } 1948 VectorType *VT = cast<VectorType>(T); 1949 T = VT->getElementType(); 1950 uint64_t EltSize = TD->getTypeAllocSize(T); 1951 Idx = Offset / EltSize; 1952 Offset -= Idx * EltSize; 1953 IdxTy = Type::getInt64Ty(T->getContext()); 1954 return Idx; 1955} 1956 1957/// RewriteGEP - Check if this GEP instruction moves the pointer across 1958/// elements of the alloca that are being split apart, and if so, rewrite 1959/// the GEP to be relative to the new element. 1960void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1961 SmallVector<AllocaInst*, 32> &NewElts) { 1962 uint64_t OldOffset = Offset; 1963 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1964 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1965 1966 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1967 1968 Type *T = AI->getAllocatedType(); 1969 Type *IdxTy; 1970 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1971 if (GEPI->getOperand(0) == AI) 1972 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1973 1974 T = AI->getAllocatedType(); 1975 uint64_t EltOffset = Offset; 1976 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1977 1978 // If this GEP does not move the pointer across elements of the alloca 1979 // being split, then it does not needs to be rewritten. 1980 if (Idx == OldIdx) 1981 return; 1982 1983 Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1984 SmallVector<Value*, 8> NewArgs; 1985 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1986 while (EltOffset != 0) { 1987 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1988 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1989 } 1990 Instruction *Val = NewElts[Idx]; 1991 if (NewArgs.size() > 1) { 1992 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI); 1993 Val->takeName(GEPI); 1994 } 1995 if (Val->getType() != GEPI->getType()) 1996 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1997 GEPI->replaceAllUsesWith(Val); 1998 DeadInsts.push_back(GEPI); 1999} 2000 2001/// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it 2002/// to mark the lifetime of the scalarized memory. 2003void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 2004 uint64_t Offset, 2005 SmallVector<AllocaInst*, 32> &NewElts) { 2006 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0)); 2007 // Put matching lifetime markers on everything from Offset up to 2008 // Offset+OldSize. 2009 Type *AIType = AI->getAllocatedType(); 2010 uint64_t NewOffset = Offset; 2011 Type *IdxTy; 2012 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy); 2013 2014 IRBuilder<> Builder(II); 2015 uint64_t Size = OldSize->getLimitedValue(); 2016 2017 if (NewOffset) { 2018 // Splice the first element and index 'NewOffset' bytes in. SROA will 2019 // split the alloca again later. 2020 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy()); 2021 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset)); 2022 2023 IdxTy = NewElts[Idx]->getAllocatedType(); 2024 uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset; 2025 if (EltSize > Size) { 2026 EltSize = Size; 2027 Size = 0; 2028 } else { 2029 Size -= EltSize; 2030 } 2031 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2032 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize)); 2033 else 2034 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize)); 2035 ++Idx; 2036 } 2037 2038 for (; Idx != NewElts.size() && Size; ++Idx) { 2039 IdxTy = NewElts[Idx]->getAllocatedType(); 2040 uint64_t EltSize = TD->getTypeAllocSize(IdxTy); 2041 if (EltSize > Size) { 2042 EltSize = Size; 2043 Size = 0; 2044 } else { 2045 Size -= EltSize; 2046 } 2047 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2048 Builder.CreateLifetimeStart(NewElts[Idx], 2049 Builder.getInt64(EltSize)); 2050 else 2051 Builder.CreateLifetimeEnd(NewElts[Idx], 2052 Builder.getInt64(EltSize)); 2053 } 2054 DeadInsts.push_back(II); 2055} 2056 2057/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2058/// Rewrite it to copy or set the elements of the scalarized memory. 2059void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2060 AllocaInst *AI, 2061 SmallVector<AllocaInst*, 32> &NewElts) { 2062 // If this is a memcpy/memmove, construct the other pointer as the 2063 // appropriate type. The "Other" pointer is the pointer that goes to memory 2064 // that doesn't have anything to do with the alloca that we are promoting. For 2065 // memset, this Value* stays null. 2066 Value *OtherPtr = 0; 2067 unsigned MemAlignment = MI->getAlignment(); 2068 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2069 if (Inst == MTI->getRawDest()) 2070 OtherPtr = MTI->getRawSource(); 2071 else { 2072 assert(Inst == MTI->getRawSource()); 2073 OtherPtr = MTI->getRawDest(); 2074 } 2075 } 2076 2077 // If there is an other pointer, we want to convert it to the same pointer 2078 // type as AI has, so we can GEP through it safely. 2079 if (OtherPtr) { 2080 unsigned AddrSpace = 2081 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2082 2083 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2084 // optimization, but it's also required to detect the corner case where 2085 // both pointer operands are referencing the same memory, and where 2086 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2087 // function is only called for mem intrinsics that access the whole 2088 // aggregate, so non-zero GEPs are not an issue here.) 2089 OtherPtr = OtherPtr->stripPointerCasts(); 2090 2091 // Copying the alloca to itself is a no-op: just delete it. 2092 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2093 // This code will run twice for a no-op memcpy -- once for each operand. 2094 // Put only one reference to MI on the DeadInsts list. 2095 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2096 E = DeadInsts.end(); I != E; ++I) 2097 if (*I == MI) return; 2098 DeadInsts.push_back(MI); 2099 return; 2100 } 2101 2102 // If the pointer is not the right type, insert a bitcast to the right 2103 // type. 2104 Type *NewTy = 2105 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2106 2107 if (OtherPtr->getType() != NewTy) 2108 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2109 } 2110 2111 // Process each element of the aggregate. 2112 bool SROADest = MI->getRawDest() == Inst; 2113 2114 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2115 2116 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2117 // If this is a memcpy/memmove, emit a GEP of the other element address. 2118 Value *OtherElt = 0; 2119 unsigned OtherEltAlign = MemAlignment; 2120 2121 if (OtherPtr) { 2122 Value *Idx[2] = { Zero, 2123 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2124 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, 2125 OtherPtr->getName()+"."+Twine(i), 2126 MI); 2127 uint64_t EltOffset; 2128 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2129 Type *OtherTy = OtherPtrTy->getElementType(); 2130 if (StructType *ST = dyn_cast<StructType>(OtherTy)) { 2131 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2132 } else { 2133 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2134 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2135 } 2136 2137 // The alignment of the other pointer is the guaranteed alignment of the 2138 // element, which is affected by both the known alignment of the whole 2139 // mem intrinsic and the alignment of the element. If the alignment of 2140 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2141 // known alignment is just 4 bytes. 2142 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2143 } 2144 2145 Value *EltPtr = NewElts[i]; 2146 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2147 2148 // If we got down to a scalar, insert a load or store as appropriate. 2149 if (EltTy->isSingleValueType()) { 2150 if (isa<MemTransferInst>(MI)) { 2151 if (SROADest) { 2152 // From Other to Alloca. 2153 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2154 new StoreInst(Elt, EltPtr, MI); 2155 } else { 2156 // From Alloca to Other. 2157 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2158 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2159 } 2160 continue; 2161 } 2162 assert(isa<MemSetInst>(MI)); 2163 2164 // If the stored element is zero (common case), just store a null 2165 // constant. 2166 Constant *StoreVal; 2167 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2168 if (CI->isZero()) { 2169 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2170 } else { 2171 // If EltTy is a vector type, get the element type. 2172 Type *ValTy = EltTy->getScalarType(); 2173 2174 // Construct an integer with the right value. 2175 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2176 APInt OneVal(EltSize, CI->getZExtValue()); 2177 APInt TotalVal(OneVal); 2178 // Set each byte. 2179 for (unsigned i = 0; 8*i < EltSize; ++i) { 2180 TotalVal = TotalVal.shl(8); 2181 TotalVal |= OneVal; 2182 } 2183 2184 // Convert the integer value to the appropriate type. 2185 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2186 if (ValTy->isPointerTy()) 2187 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2188 else if (ValTy->isFloatingPointTy()) 2189 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2190 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2191 2192 // If the requested value was a vector constant, create it. 2193 if (EltTy->isVectorTy()) { 2194 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements(); 2195 StoreVal = ConstantVector::getSplat(NumElts, StoreVal); 2196 } 2197 } 2198 new StoreInst(StoreVal, EltPtr, MI); 2199 continue; 2200 } 2201 // Otherwise, if we're storing a byte variable, use a memset call for 2202 // this element. 2203 } 2204 2205 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2206 if (!EltSize) 2207 continue; 2208 2209 IRBuilder<> Builder(MI); 2210 2211 // Finally, insert the meminst for this element. 2212 if (isa<MemSetInst>(MI)) { 2213 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2214 MI->isVolatile()); 2215 } else { 2216 assert(isa<MemTransferInst>(MI)); 2217 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2218 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2219 2220 if (isa<MemCpyInst>(MI)) 2221 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2222 else 2223 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2224 } 2225 } 2226 DeadInsts.push_back(MI); 2227} 2228 2229/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2230/// overwrites the entire allocation. Extract out the pieces of the stored 2231/// integer and store them individually. 2232void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2233 SmallVector<AllocaInst*, 32> &NewElts){ 2234 // Extract each element out of the integer according to its structure offset 2235 // and store the element value to the individual alloca. 2236 Value *SrcVal = SI->getOperand(0); 2237 Type *AllocaEltTy = AI->getAllocatedType(); 2238 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2239 2240 IRBuilder<> Builder(SI); 2241 2242 // Handle tail padding by extending the operand 2243 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2244 SrcVal = Builder.CreateZExt(SrcVal, 2245 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2246 2247 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2248 << '\n'); 2249 2250 // There are two forms here: AI could be an array or struct. Both cases 2251 // have different ways to compute the element offset. 2252 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2253 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2254 2255 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2256 // Get the number of bits to shift SrcVal to get the value. 2257 Type *FieldTy = EltSTy->getElementType(i); 2258 uint64_t Shift = Layout->getElementOffsetInBits(i); 2259 2260 if (TD->isBigEndian()) 2261 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2262 2263 Value *EltVal = SrcVal; 2264 if (Shift) { 2265 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2266 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2267 } 2268 2269 // Truncate down to an integer of the right size. 2270 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2271 2272 // Ignore zero sized fields like {}, they obviously contain no data. 2273 if (FieldSizeBits == 0) continue; 2274 2275 if (FieldSizeBits != AllocaSizeBits) 2276 EltVal = Builder.CreateTrunc(EltVal, 2277 IntegerType::get(SI->getContext(), FieldSizeBits)); 2278 Value *DestField = NewElts[i]; 2279 if (EltVal->getType() == FieldTy) { 2280 // Storing to an integer field of this size, just do it. 2281 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2282 // Bitcast to the right element type (for fp/vector values). 2283 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2284 } else { 2285 // Otherwise, bitcast the dest pointer (for aggregates). 2286 DestField = Builder.CreateBitCast(DestField, 2287 PointerType::getUnqual(EltVal->getType())); 2288 } 2289 new StoreInst(EltVal, DestField, SI); 2290 } 2291 2292 } else { 2293 ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2294 Type *ArrayEltTy = ATy->getElementType(); 2295 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2296 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2297 2298 uint64_t Shift; 2299 2300 if (TD->isBigEndian()) 2301 Shift = AllocaSizeBits-ElementOffset; 2302 else 2303 Shift = 0; 2304 2305 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2306 // Ignore zero sized fields like {}, they obviously contain no data. 2307 if (ElementSizeBits == 0) continue; 2308 2309 Value *EltVal = SrcVal; 2310 if (Shift) { 2311 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2312 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2313 } 2314 2315 // Truncate down to an integer of the right size. 2316 if (ElementSizeBits != AllocaSizeBits) 2317 EltVal = Builder.CreateTrunc(EltVal, 2318 IntegerType::get(SI->getContext(), 2319 ElementSizeBits)); 2320 Value *DestField = NewElts[i]; 2321 if (EltVal->getType() == ArrayEltTy) { 2322 // Storing to an integer field of this size, just do it. 2323 } else if (ArrayEltTy->isFloatingPointTy() || 2324 ArrayEltTy->isVectorTy()) { 2325 // Bitcast to the right element type (for fp/vector values). 2326 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2327 } else { 2328 // Otherwise, bitcast the dest pointer (for aggregates). 2329 DestField = Builder.CreateBitCast(DestField, 2330 PointerType::getUnqual(EltVal->getType())); 2331 } 2332 new StoreInst(EltVal, DestField, SI); 2333 2334 if (TD->isBigEndian()) 2335 Shift -= ElementOffset; 2336 else 2337 Shift += ElementOffset; 2338 } 2339 } 2340 2341 DeadInsts.push_back(SI); 2342} 2343 2344/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2345/// an integer. Load the individual pieces to form the aggregate value. 2346void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2347 SmallVector<AllocaInst*, 32> &NewElts) { 2348 // Extract each element out of the NewElts according to its structure offset 2349 // and form the result value. 2350 Type *AllocaEltTy = AI->getAllocatedType(); 2351 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2352 2353 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2354 << '\n'); 2355 2356 // There are two forms here: AI could be an array or struct. Both cases 2357 // have different ways to compute the element offset. 2358 const StructLayout *Layout = 0; 2359 uint64_t ArrayEltBitOffset = 0; 2360 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2361 Layout = TD->getStructLayout(EltSTy); 2362 } else { 2363 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2364 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2365 } 2366 2367 Value *ResultVal = 2368 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2369 2370 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2371 // Load the value from the alloca. If the NewElt is an aggregate, cast 2372 // the pointer to an integer of the same size before doing the load. 2373 Value *SrcField = NewElts[i]; 2374 Type *FieldTy = 2375 cast<PointerType>(SrcField->getType())->getElementType(); 2376 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2377 2378 // Ignore zero sized fields like {}, they obviously contain no data. 2379 if (FieldSizeBits == 0) continue; 2380 2381 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2382 FieldSizeBits); 2383 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2384 !FieldTy->isVectorTy()) 2385 SrcField = new BitCastInst(SrcField, 2386 PointerType::getUnqual(FieldIntTy), 2387 "", LI); 2388 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2389 2390 // If SrcField is a fp or vector of the right size but that isn't an 2391 // integer type, bitcast to an integer so we can shift it. 2392 if (SrcField->getType() != FieldIntTy) 2393 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2394 2395 // Zero extend the field to be the same size as the final alloca so that 2396 // we can shift and insert it. 2397 if (SrcField->getType() != ResultVal->getType()) 2398 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2399 2400 // Determine the number of bits to shift SrcField. 2401 uint64_t Shift; 2402 if (Layout) // Struct case. 2403 Shift = Layout->getElementOffsetInBits(i); 2404 else // Array case. 2405 Shift = i*ArrayEltBitOffset; 2406 2407 if (TD->isBigEndian()) 2408 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2409 2410 if (Shift) { 2411 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2412 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2413 } 2414 2415 // Don't create an 'or x, 0' on the first iteration. 2416 if (!isa<Constant>(ResultVal) || 2417 !cast<Constant>(ResultVal)->isNullValue()) 2418 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2419 else 2420 ResultVal = SrcField; 2421 } 2422 2423 // Handle tail padding by truncating the result 2424 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2425 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2426 2427 LI->replaceAllUsesWith(ResultVal); 2428 DeadInsts.push_back(LI); 2429} 2430 2431/// HasPadding - Return true if the specified type has any structure or 2432/// alignment padding in between the elements that would be split apart 2433/// by SROA; return false otherwise. 2434static bool HasPadding(Type *Ty, const TargetData &TD) { 2435 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2436 Ty = ATy->getElementType(); 2437 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2438 } 2439 2440 // SROA currently handles only Arrays and Structs. 2441 StructType *STy = cast<StructType>(Ty); 2442 const StructLayout *SL = TD.getStructLayout(STy); 2443 unsigned PrevFieldBitOffset = 0; 2444 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2445 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2446 2447 // Check to see if there is any padding between this element and the 2448 // previous one. 2449 if (i) { 2450 unsigned PrevFieldEnd = 2451 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2452 if (PrevFieldEnd < FieldBitOffset) 2453 return true; 2454 } 2455 PrevFieldBitOffset = FieldBitOffset; 2456 } 2457 // Check for tail padding. 2458 if (unsigned EltCount = STy->getNumElements()) { 2459 unsigned PrevFieldEnd = PrevFieldBitOffset + 2460 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2461 if (PrevFieldEnd < SL->getSizeInBits()) 2462 return true; 2463 } 2464 return false; 2465} 2466 2467/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2468/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2469/// or 1 if safe after canonicalization has been performed. 2470bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2471 // Loop over the use list of the alloca. We can only transform it if all of 2472 // the users are safe to transform. 2473 AllocaInfo Info(AI); 2474 2475 isSafeForScalarRepl(AI, 0, Info); 2476 if (Info.isUnsafe) { 2477 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2478 return false; 2479 } 2480 2481 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2482 // source and destination, we have to be careful. In particular, the memcpy 2483 // could be moving around elements that live in structure padding of the LLVM 2484 // types, but may actually be used. In these cases, we refuse to promote the 2485 // struct. 2486 if (Info.isMemCpySrc && Info.isMemCpyDst && 2487 HasPadding(AI->getAllocatedType(), *TD)) 2488 return false; 2489 2490 // If the alloca never has an access to just *part* of it, but is accessed 2491 // via loads and stores, then we should use ConvertToScalarInfo to promote 2492 // the alloca instead of promoting each piece at a time and inserting fission 2493 // and fusion code. 2494 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2495 // If the struct/array just has one element, use basic SRoA. 2496 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2497 if (ST->getNumElements() > 1) return false; 2498 } else { 2499 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2500 return false; 2501 } 2502 } 2503 2504 return true; 2505} 2506 2507 2508 2509/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2510/// some part of a constant global variable. This intentionally only accepts 2511/// constant expressions because we don't can't rewrite arbitrary instructions. 2512static bool PointsToConstantGlobal(Value *V) { 2513 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2514 return GV->isConstant(); 2515 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2516 if (CE->getOpcode() == Instruction::BitCast || 2517 CE->getOpcode() == Instruction::GetElementPtr) 2518 return PointsToConstantGlobal(CE->getOperand(0)); 2519 return false; 2520} 2521 2522/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2523/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2524/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2525/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2526/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2527/// the alloca, and if the source pointer is a pointer to a constant global, we 2528/// can optimize this. 2529static bool 2530isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2531 bool isOffset, 2532 SmallVector<Instruction *, 4> &LifetimeMarkers) { 2533 // We track lifetime intrinsics as we encounter them. If we decide to go 2534 // ahead and replace the value with the global, this lets the caller quickly 2535 // eliminate the markers. 2536 2537 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2538 User *U = cast<Instruction>(*UI); 2539 2540 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2541 // Ignore non-volatile loads, they are always ok. 2542 if (!LI->isSimple()) return false; 2543 continue; 2544 } 2545 2546 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2547 // If uses of the bitcast are ok, we are ok. 2548 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset, 2549 LifetimeMarkers)) 2550 return false; 2551 continue; 2552 } 2553 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2554 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2555 // doesn't, it does. 2556 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2557 isOffset || !GEP->hasAllZeroIndices(), 2558 LifetimeMarkers)) 2559 return false; 2560 continue; 2561 } 2562 2563 if (CallSite CS = U) { 2564 // If this is the function being called then we treat it like a load and 2565 // ignore it. 2566 if (CS.isCallee(UI)) 2567 continue; 2568 2569 // If this is a readonly/readnone call site, then we know it is just a 2570 // load (but one that potentially returns the value itself), so we can 2571 // ignore it if we know that the value isn't captured. 2572 unsigned ArgNo = CS.getArgumentNo(UI); 2573 if (CS.onlyReadsMemory() && 2574 (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo))) 2575 continue; 2576 2577 // If this is being passed as a byval argument, the caller is making a 2578 // copy, so it is only a read of the alloca. 2579 if (CS.isByValArgument(ArgNo)) 2580 continue; 2581 } 2582 2583 // Lifetime intrinsics can be handled by the caller. 2584 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 2585 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 2586 II->getIntrinsicID() == Intrinsic::lifetime_end) { 2587 assert(II->use_empty() && "Lifetime markers have no result to use!"); 2588 LifetimeMarkers.push_back(II); 2589 continue; 2590 } 2591 } 2592 2593 // If this is isn't our memcpy/memmove, reject it as something we can't 2594 // handle. 2595 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2596 if (MI == 0) 2597 return false; 2598 2599 // If the transfer is using the alloca as a source of the transfer, then 2600 // ignore it since it is a load (unless the transfer is volatile). 2601 if (UI.getOperandNo() == 1) { 2602 if (MI->isVolatile()) return false; 2603 continue; 2604 } 2605 2606 // If we already have seen a copy, reject the second one. 2607 if (TheCopy) return false; 2608 2609 // If the pointer has been offset from the start of the alloca, we can't 2610 // safely handle this. 2611 if (isOffset) return false; 2612 2613 // If the memintrinsic isn't using the alloca as the dest, reject it. 2614 if (UI.getOperandNo() != 0) return false; 2615 2616 // If the source of the memcpy/move is not a constant global, reject it. 2617 if (!PointsToConstantGlobal(MI->getSource())) 2618 return false; 2619 2620 // Otherwise, the transform is safe. Remember the copy instruction. 2621 TheCopy = MI; 2622 } 2623 return true; 2624} 2625 2626/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2627/// modified by a copy from a constant global. If we can prove this, we can 2628/// replace any uses of the alloca with uses of the global directly. 2629MemTransferInst * 2630SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI, 2631 SmallVector<Instruction*, 4> &ToDelete) { 2632 MemTransferInst *TheCopy = 0; 2633 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete)) 2634 return TheCopy; 2635 return 0; 2636} 2637