ScalarReplAggregates.cpp revision 125ef76934f37e3fdc8ce4a2cc238850d06c5912
1//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This transformation implements the well known scalar replacement of 11// aggregates transformation. This xform breaks up alloca instructions of 12// aggregate type (structure or array) into individual alloca instructions for 13// each member (if possible). Then, if possible, it transforms the individual 14// alloca instructions into nice clean scalar SSA form. 15// 16// This combines a simple SRoA algorithm with the Mem2Reg algorithm because 17// often interact, especially for C++ programs. As such, iterating between 18// SRoA, then Mem2Reg until we run out of things to promote works well. 19// 20//===----------------------------------------------------------------------===// 21 22#define DEBUG_TYPE "scalarrepl" 23#include "llvm/Transforms/Scalar.h" 24#include "llvm/Constants.h" 25#include "llvm/DerivedTypes.h" 26#include "llvm/Function.h" 27#include "llvm/GlobalVariable.h" 28#include "llvm/Instructions.h" 29#include "llvm/IntrinsicInst.h" 30#include "llvm/LLVMContext.h" 31#include "llvm/Module.h" 32#include "llvm/Pass.h" 33#include "llvm/Analysis/DebugInfo.h" 34#include "llvm/Analysis/DIBuilder.h" 35#include "llvm/Analysis/Dominators.h" 36#include "llvm/Analysis/Loads.h" 37#include "llvm/Analysis/ValueTracking.h" 38#include "llvm/Target/TargetData.h" 39#include "llvm/Transforms/Utils/PromoteMemToReg.h" 40#include "llvm/Transforms/Utils/Local.h" 41#include "llvm/Transforms/Utils/SSAUpdater.h" 42#include "llvm/Support/CallSite.h" 43#include "llvm/Support/Debug.h" 44#include "llvm/Support/ErrorHandling.h" 45#include "llvm/Support/GetElementPtrTypeIterator.h" 46#include "llvm/Support/IRBuilder.h" 47#include "llvm/Support/MathExtras.h" 48#include "llvm/Support/raw_ostream.h" 49#include "llvm/ADT/SetVector.h" 50#include "llvm/ADT/SmallVector.h" 51#include "llvm/ADT/Statistic.h" 52using namespace llvm; 53 54STATISTIC(NumReplaced, "Number of allocas broken up"); 55STATISTIC(NumPromoted, "Number of allocas promoted"); 56STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); 57STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 58STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 59 60namespace { 61 struct SROA : public FunctionPass { 62 SROA(int T, bool hasDT, char &ID) 63 : FunctionPass(ID), HasDomTree(hasDT) { 64 if (T == -1) 65 SRThreshold = 128; 66 else 67 SRThreshold = T; 68 } 69 70 bool runOnFunction(Function &F); 71 72 bool performScalarRepl(Function &F); 73 bool performPromotion(Function &F); 74 75 private: 76 bool HasDomTree; 77 TargetData *TD; 78 79 /// DeadInsts - Keep track of instructions we have made dead, so that 80 /// we can remove them after we are done working. 81 SmallVector<Value*, 32> DeadInsts; 82 83 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 84 /// information about the uses. All these fields are initialized to false 85 /// and set to true when something is learned. 86 struct AllocaInfo { 87 /// The alloca to promote. 88 AllocaInst *AI; 89 90 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite 91 /// looping and avoid redundant work. 92 SmallPtrSet<PHINode*, 8> CheckedPHIs; 93 94 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 95 bool isUnsafe : 1; 96 97 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 98 bool isMemCpySrc : 1; 99 100 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 101 bool isMemCpyDst : 1; 102 103 /// hasSubelementAccess - This is true if a subelement of the alloca is 104 /// ever accessed, or false if the alloca is only accessed with mem 105 /// intrinsics or load/store that only access the entire alloca at once. 106 bool hasSubelementAccess : 1; 107 108 /// hasALoadOrStore - This is true if there are any loads or stores to it. 109 /// The alloca may just be accessed with memcpy, for example, which would 110 /// not set this. 111 bool hasALoadOrStore : 1; 112 113 explicit AllocaInfo(AllocaInst *ai) 114 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 115 hasSubelementAccess(false), hasALoadOrStore(false) {} 116 }; 117 118 unsigned SRThreshold; 119 120 void MarkUnsafe(AllocaInfo &I, Instruction *User) { 121 I.isUnsafe = true; 122 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); 123 } 124 125 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 126 127 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info); 128 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset, 129 AllocaInfo &Info); 130 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info); 131 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 132 Type *MemOpType, bool isStore, AllocaInfo &Info, 133 Instruction *TheAccess, bool AllowWholeAccess); 134 bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size); 135 uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset, 136 Type *&IdxTy); 137 138 void DoScalarReplacement(AllocaInst *AI, 139 std::vector<AllocaInst*> &WorkList); 140 void DeleteDeadInstructions(); 141 142 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 143 SmallVector<AllocaInst*, 32> &NewElts); 144 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 145 SmallVector<AllocaInst*, 32> &NewElts); 146 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 147 SmallVector<AllocaInst*, 32> &NewElts); 148 void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 149 uint64_t Offset, 150 SmallVector<AllocaInst*, 32> &NewElts); 151 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 152 AllocaInst *AI, 153 SmallVector<AllocaInst*, 32> &NewElts); 154 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 155 SmallVector<AllocaInst*, 32> &NewElts); 156 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 157 SmallVector<AllocaInst*, 32> &NewElts); 158 159 static MemTransferInst *isOnlyCopiedFromConstantGlobal( 160 AllocaInst *AI, SmallVector<Instruction*, 4> &ToDelete); 161 }; 162 163 // SROA_DT - SROA that uses DominatorTree. 164 struct SROA_DT : public SROA { 165 static char ID; 166 public: 167 SROA_DT(int T = -1) : SROA(T, true, ID) { 168 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 169 } 170 171 // getAnalysisUsage - This pass does not require any passes, but we know it 172 // will not alter the CFG, so say so. 173 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 174 AU.addRequired<DominatorTree>(); 175 AU.setPreservesCFG(); 176 } 177 }; 178 179 // SROA_SSAUp - SROA that uses SSAUpdater. 180 struct SROA_SSAUp : public SROA { 181 static char ID; 182 public: 183 SROA_SSAUp(int T = -1) : SROA(T, false, ID) { 184 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 185 } 186 187 // getAnalysisUsage - This pass does not require any passes, but we know it 188 // will not alter the CFG, so say so. 189 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 190 AU.setPreservesCFG(); 191 } 192 }; 193 194} 195 196char SROA_DT::ID = 0; 197char SROA_SSAUp::ID = 0; 198 199INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 200 "Scalar Replacement of Aggregates (DT)", false, false) 201INITIALIZE_PASS_DEPENDENCY(DominatorTree) 202INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 203 "Scalar Replacement of Aggregates (DT)", false, false) 204 205INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 206 "Scalar Replacement of Aggregates (SSAUp)", false, false) 207INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 208 "Scalar Replacement of Aggregates (SSAUp)", false, false) 209 210// Public interface to the ScalarReplAggregates pass 211FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 212 bool UseDomTree) { 213 if (UseDomTree) 214 return new SROA_DT(Threshold); 215 return new SROA_SSAUp(Threshold); 216} 217 218 219//===----------------------------------------------------------------------===// 220// Convert To Scalar Optimization. 221//===----------------------------------------------------------------------===// 222 223namespace { 224/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 225/// optimization, which scans the uses of an alloca and determines if it can 226/// rewrite it in terms of a single new alloca that can be mem2reg'd. 227class ConvertToScalarInfo { 228 /// AllocaSize - The size of the alloca being considered in bytes. 229 unsigned AllocaSize; 230 const TargetData &TD; 231 232 /// IsNotTrivial - This is set to true if there is some access to the object 233 /// which means that mem2reg can't promote it. 234 bool IsNotTrivial; 235 236 /// ScalarKind - Tracks the kind of alloca being considered for promotion, 237 /// computed based on the uses of the alloca rather than the LLVM type system. 238 enum { 239 Unknown, 240 241 // Accesses via GEPs that are consistent with element access of a vector 242 // type. This will not be converted into a vector unless there is a later 243 // access using an actual vector type. 244 ImplicitVector, 245 246 // Accesses via vector operations and GEPs that are consistent with the 247 // layout of a vector type. 248 Vector, 249 250 // An integer bag-of-bits with bitwise operations for insertion and 251 // extraction. Any combination of types can be converted into this kind 252 // of scalar. 253 Integer 254 } ScalarKind; 255 256 /// VectorTy - This tracks the type that we should promote the vector to if 257 /// it is possible to turn it into a vector. This starts out null, and if it 258 /// isn't possible to turn into a vector type, it gets set to VoidTy. 259 VectorType *VectorTy; 260 261 /// HadNonMemTransferAccess - True if there is at least one access to the 262 /// alloca that is not a MemTransferInst. We don't want to turn structs into 263 /// large integers unless there is some potential for optimization. 264 bool HadNonMemTransferAccess; 265 266public: 267 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 268 : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown), 269 VectorTy(0), HadNonMemTransferAccess(false) { } 270 271 AllocaInst *TryConvert(AllocaInst *AI); 272 273private: 274 bool CanConvertToScalar(Value *V, uint64_t Offset); 275 void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset); 276 bool MergeInVectorType(VectorType *VInTy, uint64_t Offset); 277 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 278 279 Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType, 280 uint64_t Offset, IRBuilder<> &Builder); 281 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 282 uint64_t Offset, IRBuilder<> &Builder); 283}; 284} // end anonymous namespace. 285 286 287/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 288/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 289/// alloca if possible or null if not. 290AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 291 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 292 // out. 293 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 294 return 0; 295 296 // If an alloca has only memset / memcpy uses, it may still have an Unknown 297 // ScalarKind. Treat it as an Integer below. 298 if (ScalarKind == Unknown) 299 ScalarKind = Integer; 300 301 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8) 302 ScalarKind = Integer; 303 304 // If we were able to find a vector type that can handle this with 305 // insert/extract elements, and if there was at least one use that had 306 // a vector type, promote this to a vector. We don't want to promote 307 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 308 // we just get a lot of insert/extracts. If at least one vector is 309 // involved, then we probably really do have a union of vector/array. 310 Type *NewTy; 311 if (ScalarKind == Vector) { 312 assert(VectorTy && "Missing type for vector scalar."); 313 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 314 << *VectorTy << '\n'); 315 NewTy = VectorTy; // Use the vector type. 316 } else { 317 unsigned BitWidth = AllocaSize * 8; 318 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) && 319 !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth)) 320 return 0; 321 322 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 323 // Create and insert the integer alloca. 324 NewTy = IntegerType::get(AI->getContext(), BitWidth); 325 } 326 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 327 ConvertUsesToScalar(AI, NewAI, 0); 328 return NewAI; 329} 330 331/// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type 332/// (VectorTy) so far at the offset specified by Offset (which is specified in 333/// bytes). 334/// 335/// There are two cases we handle here: 336/// 1) A union of vector types of the same size and potentially its elements. 337/// Here we turn element accesses into insert/extract element operations. 338/// This promotes a <4 x float> with a store of float to the third element 339/// into a <4 x float> that uses insert element. 340/// 2) A fully general blob of memory, which we turn into some (potentially 341/// large) integer type with extract and insert operations where the loads 342/// and stores would mutate the memory. We mark this by setting VectorTy 343/// to VoidTy. 344void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In, 345 uint64_t Offset) { 346 // If we already decided to turn this into a blob of integer memory, there is 347 // nothing to be done. 348 if (ScalarKind == Integer) 349 return; 350 351 // If this could be contributing to a vector, analyze it. 352 353 // If the In type is a vector that is the same size as the alloca, see if it 354 // matches the existing VecTy. 355 if (VectorType *VInTy = dyn_cast<VectorType>(In)) { 356 if (MergeInVectorType(VInTy, Offset)) 357 return; 358 } else if (In->isFloatTy() || In->isDoubleTy() || 359 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 360 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 361 // Full width accesses can be ignored, because they can always be turned 362 // into bitcasts. 363 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 364 if (EltSize == AllocaSize) 365 return; 366 367 // If we're accessing something that could be an element of a vector, see 368 // if the implied vector agrees with what we already have and if Offset is 369 // compatible with it. 370 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 371 (!VectorTy || EltSize == VectorTy->getElementType() 372 ->getPrimitiveSizeInBits()/8)) { 373 if (!VectorTy) { 374 ScalarKind = ImplicitVector; 375 VectorTy = VectorType::get(In, AllocaSize/EltSize); 376 } 377 return; 378 } 379 } 380 381 // Otherwise, we have a case that we can't handle with an optimized vector 382 // form. We can still turn this into a large integer. 383 ScalarKind = Integer; 384} 385 386/// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore, 387/// returning true if the type was successfully merged and false otherwise. 388bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy, 389 uint64_t Offset) { 390 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 391 // If we're storing/loading a vector of the right size, allow it as a 392 // vector. If this the first vector we see, remember the type so that 393 // we know the element size. If this is a subsequent access, ignore it 394 // even if it is a differing type but the same size. Worst case we can 395 // bitcast the resultant vectors. 396 if (!VectorTy) 397 VectorTy = VInTy; 398 ScalarKind = Vector; 399 return true; 400 } 401 402 return false; 403} 404 405/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 406/// its accesses to a single vector type, return true and set VecTy to 407/// the new type. If we could convert the alloca into a single promotable 408/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 409/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 410/// is the current offset from the base of the alloca being analyzed. 411/// 412/// If we see at least one access to the value that is as a vector type, set the 413/// SawVec flag. 414bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 415 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 416 Instruction *User = cast<Instruction>(*UI); 417 418 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 419 // Don't break volatile loads. 420 if (!LI->isSimple()) 421 return false; 422 // Don't touch MMX operations. 423 if (LI->getType()->isX86_MMXTy()) 424 return false; 425 HadNonMemTransferAccess = true; 426 MergeInTypeForLoadOrStore(LI->getType(), Offset); 427 continue; 428 } 429 430 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 431 // Storing the pointer, not into the value? 432 if (SI->getOperand(0) == V || !SI->isSimple()) return false; 433 // Don't touch MMX operations. 434 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 435 return false; 436 HadNonMemTransferAccess = true; 437 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset); 438 continue; 439 } 440 441 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 442 if (!onlyUsedByLifetimeMarkers(BCI)) 443 IsNotTrivial = true; // Can't be mem2reg'd. 444 if (!CanConvertToScalar(BCI, Offset)) 445 return false; 446 continue; 447 } 448 449 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 450 // If this is a GEP with a variable indices, we can't handle it. 451 if (!GEP->hasAllConstantIndices()) 452 return false; 453 454 // Compute the offset that this GEP adds to the pointer. 455 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 456 if (!GEP->getPointerOperandType()->isPointerTy()) 457 return false; 458 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 459 Indices); 460 // See if all uses can be converted. 461 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 462 return false; 463 IsNotTrivial = true; // Can't be mem2reg'd. 464 HadNonMemTransferAccess = true; 465 continue; 466 } 467 468 // If this is a constant sized memset of a constant value (e.g. 0) we can 469 // handle it. 470 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 471 // Store of constant value. 472 if (!isa<ConstantInt>(MSI->getValue())) 473 return false; 474 475 // Store of constant size. 476 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength()); 477 if (!Len) 478 return false; 479 480 // If the size differs from the alloca, we can only convert the alloca to 481 // an integer bag-of-bits. 482 // FIXME: This should handle all of the cases that are currently accepted 483 // as vector element insertions. 484 if (Len->getZExtValue() != AllocaSize || Offset != 0) 485 ScalarKind = Integer; 486 487 IsNotTrivial = true; // Can't be mem2reg'd. 488 HadNonMemTransferAccess = true; 489 continue; 490 } 491 492 // If this is a memcpy or memmove into or out of the whole allocation, we 493 // can handle it like a load or store of the scalar type. 494 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 495 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 496 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 497 return false; 498 499 IsNotTrivial = true; // Can't be mem2reg'd. 500 continue; 501 } 502 503 // If this is a lifetime intrinsic, we can handle it. 504 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 505 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 506 II->getIntrinsicID() == Intrinsic::lifetime_end) { 507 continue; 508 } 509 } 510 511 // Otherwise, we cannot handle this! 512 return false; 513 } 514 515 return true; 516} 517 518/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 519/// directly. This happens when we are converting an "integer union" to a 520/// single integer scalar, or when we are converting a "vector union" to a 521/// vector with insert/extractelement instructions. 522/// 523/// Offset is an offset from the original alloca, in bits that need to be 524/// shifted to the right. By the end of this, there should be no uses of Ptr. 525void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 526 uint64_t Offset) { 527 while (!Ptr->use_empty()) { 528 Instruction *User = cast<Instruction>(Ptr->use_back()); 529 530 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 531 ConvertUsesToScalar(CI, NewAI, Offset); 532 CI->eraseFromParent(); 533 continue; 534 } 535 536 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 537 // Compute the offset that this GEP adds to the pointer. 538 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 539 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 540 Indices); 541 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 542 GEP->eraseFromParent(); 543 continue; 544 } 545 546 IRBuilder<> Builder(User); 547 548 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 549 // The load is a bit extract from NewAI shifted right by Offset bits. 550 Value *LoadedVal = Builder.CreateLoad(NewAI); 551 Value *NewLoadVal 552 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 553 LI->replaceAllUsesWith(NewLoadVal); 554 LI->eraseFromParent(); 555 continue; 556 } 557 558 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 559 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 560 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 561 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 562 Builder); 563 Builder.CreateStore(New, NewAI); 564 SI->eraseFromParent(); 565 566 // If the load we just inserted is now dead, then the inserted store 567 // overwrote the entire thing. 568 if (Old->use_empty()) 569 Old->eraseFromParent(); 570 continue; 571 } 572 573 // If this is a constant sized memset of a constant value (e.g. 0) we can 574 // transform it into a store of the expanded constant value. 575 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 576 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 577 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 578 if (NumBytes != 0) { 579 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 580 581 // Compute the value replicated the right number of times. 582 APInt APVal(NumBytes*8, Val); 583 584 // Splat the value if non-zero. 585 if (Val) 586 for (unsigned i = 1; i != NumBytes; ++i) 587 APVal |= APVal << 8; 588 589 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 590 Value *New = ConvertScalar_InsertValue( 591 ConstantInt::get(User->getContext(), APVal), 592 Old, Offset, Builder); 593 Builder.CreateStore(New, NewAI); 594 595 // If the load we just inserted is now dead, then the memset overwrote 596 // the entire thing. 597 if (Old->use_empty()) 598 Old->eraseFromParent(); 599 } 600 MSI->eraseFromParent(); 601 continue; 602 } 603 604 // If this is a memcpy or memmove into or out of the whole allocation, we 605 // can handle it like a load or store of the scalar type. 606 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 607 assert(Offset == 0 && "must be store to start of alloca"); 608 609 // If the source and destination are both to the same alloca, then this is 610 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 611 // as appropriate. 612 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 613 614 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 615 // Dest must be OrigAI, change this to be a load from the original 616 // pointer (bitcasted), then a store to our new alloca. 617 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 618 Value *SrcPtr = MTI->getSource(); 619 PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 620 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 621 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 622 AIPTy = PointerType::get(AIPTy->getElementType(), 623 SPTy->getAddressSpace()); 624 } 625 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 626 627 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 628 SrcVal->setAlignment(MTI->getAlignment()); 629 Builder.CreateStore(SrcVal, NewAI); 630 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 631 // Src must be OrigAI, change this to be a load from NewAI then a store 632 // through the original dest pointer (bitcasted). 633 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 634 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 635 636 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 637 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 638 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 639 AIPTy = PointerType::get(AIPTy->getElementType(), 640 DPTy->getAddressSpace()); 641 } 642 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 643 644 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 645 NewStore->setAlignment(MTI->getAlignment()); 646 } else { 647 // Noop transfer. Src == Dst 648 } 649 650 MTI->eraseFromParent(); 651 continue; 652 } 653 654 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 655 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 656 II->getIntrinsicID() == Intrinsic::lifetime_end) { 657 // There's no need to preserve these, as the resulting alloca will be 658 // converted to a register anyways. 659 II->eraseFromParent(); 660 continue; 661 } 662 } 663 664 llvm_unreachable("Unsupported operation!"); 665 } 666} 667 668/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 669/// or vector value FromVal, extracting the bits from the offset specified by 670/// Offset. This returns the value, which is of type ToType. 671/// 672/// This happens when we are converting an "integer union" to a single 673/// integer scalar, or when we are converting a "vector union" to a vector with 674/// insert/extractelement instructions. 675/// 676/// Offset is an offset from the original alloca, in bits that need to be 677/// shifted to the right. 678Value *ConvertToScalarInfo:: 679ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, 680 uint64_t Offset, IRBuilder<> &Builder) { 681 // If the load is of the whole new alloca, no conversion is needed. 682 Type *FromType = FromVal->getType(); 683 if (FromType == ToType && Offset == 0) 684 return FromVal; 685 686 // If the result alloca is a vector type, this is either an element 687 // access or a bitcast to another vector type of the same size. 688 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) { 689 unsigned FromTypeSize = TD.getTypeAllocSize(FromType); 690 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 691 if (FromTypeSize == ToTypeSize) 692 return Builder.CreateBitCast(FromVal, ToType); 693 694 // Otherwise it must be an element access. 695 unsigned Elt = 0; 696 if (Offset) { 697 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 698 Elt = Offset/EltSize; 699 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 700 } 701 // Return the element extracted out of it. 702 Value *V = Builder.CreateExtractElement(FromVal, Builder.getInt32(Elt)); 703 if (V->getType() != ToType) 704 V = Builder.CreateBitCast(V, ToType); 705 return V; 706 } 707 708 // If ToType is a first class aggregate, extract out each of the pieces and 709 // use insertvalue's to form the FCA. 710 if (StructType *ST = dyn_cast<StructType>(ToType)) { 711 const StructLayout &Layout = *TD.getStructLayout(ST); 712 Value *Res = UndefValue::get(ST); 713 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 714 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 715 Offset+Layout.getElementOffsetInBits(i), 716 Builder); 717 Res = Builder.CreateInsertValue(Res, Elt, i); 718 } 719 return Res; 720 } 721 722 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 723 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 724 Value *Res = UndefValue::get(AT); 725 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 726 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 727 Offset+i*EltSize, Builder); 728 Res = Builder.CreateInsertValue(Res, Elt, i); 729 } 730 return Res; 731 } 732 733 // Otherwise, this must be a union that was converted to an integer value. 734 IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 735 736 // If this is a big-endian system and the load is narrower than the 737 // full alloca type, we need to do a shift to get the right bits. 738 int ShAmt = 0; 739 if (TD.isBigEndian()) { 740 // On big-endian machines, the lowest bit is stored at the bit offset 741 // from the pointer given by getTypeStoreSizeInBits. This matters for 742 // integers with a bitwidth that is not a multiple of 8. 743 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 744 TD.getTypeStoreSizeInBits(ToType) - Offset; 745 } else { 746 ShAmt = Offset; 747 } 748 749 // Note: we support negative bitwidths (with shl) which are not defined. 750 // We do this to support (f.e.) loads off the end of a structure where 751 // only some bits are used. 752 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 753 FromVal = Builder.CreateLShr(FromVal, 754 ConstantInt::get(FromVal->getType(), ShAmt)); 755 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 756 FromVal = Builder.CreateShl(FromVal, 757 ConstantInt::get(FromVal->getType(), -ShAmt)); 758 759 // Finally, unconditionally truncate the integer to the right width. 760 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 761 if (LIBitWidth < NTy->getBitWidth()) 762 FromVal = 763 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 764 LIBitWidth)); 765 else if (LIBitWidth > NTy->getBitWidth()) 766 FromVal = 767 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 768 LIBitWidth)); 769 770 // If the result is an integer, this is a trunc or bitcast. 771 if (ToType->isIntegerTy()) { 772 // Should be done. 773 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 774 // Just do a bitcast, we know the sizes match up. 775 FromVal = Builder.CreateBitCast(FromVal, ToType); 776 } else { 777 // Otherwise must be a pointer. 778 FromVal = Builder.CreateIntToPtr(FromVal, ToType); 779 } 780 assert(FromVal->getType() == ToType && "Didn't convert right?"); 781 return FromVal; 782} 783 784/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 785/// or vector value "Old" at the offset specified by Offset. 786/// 787/// This happens when we are converting an "integer union" to a 788/// single integer scalar, or when we are converting a "vector union" to a 789/// vector with insert/extractelement instructions. 790/// 791/// Offset is an offset from the original alloca, in bits that need to be 792/// shifted to the right. 793Value *ConvertToScalarInfo:: 794ConvertScalar_InsertValue(Value *SV, Value *Old, 795 uint64_t Offset, IRBuilder<> &Builder) { 796 // Convert the stored type to the actual type, shift it left to insert 797 // then 'or' into place. 798 Type *AllocaType = Old->getType(); 799 LLVMContext &Context = Old->getContext(); 800 801 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 802 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 803 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 804 805 // Changing the whole vector with memset or with an access of a different 806 // vector type? 807 if (ValSize == VecSize) 808 return Builder.CreateBitCast(SV, AllocaType); 809 810 // Must be an element insertion. 811 Type *EltTy = VTy->getElementType(); 812 if (SV->getType() != EltTy) 813 SV = Builder.CreateBitCast(SV, EltTy); 814 uint64_t EltSize = TD.getTypeAllocSizeInBits(EltTy); 815 unsigned Elt = Offset/EltSize; 816 return Builder.CreateInsertElement(Old, SV, Builder.getInt32(Elt)); 817 } 818 819 // If SV is a first-class aggregate value, insert each value recursively. 820 if (StructType *ST = dyn_cast<StructType>(SV->getType())) { 821 const StructLayout &Layout = *TD.getStructLayout(ST); 822 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 823 Value *Elt = Builder.CreateExtractValue(SV, i); 824 Old = ConvertScalar_InsertValue(Elt, Old, 825 Offset+Layout.getElementOffsetInBits(i), 826 Builder); 827 } 828 return Old; 829 } 830 831 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 832 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 833 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 834 Value *Elt = Builder.CreateExtractValue(SV, i); 835 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 836 } 837 return Old; 838 } 839 840 // If SV is a float, convert it to the appropriate integer type. 841 // If it is a pointer, do the same. 842 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 843 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 844 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 845 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 846 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 847 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth)); 848 else if (SV->getType()->isPointerTy()) 849 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext())); 850 851 // Zero extend or truncate the value if needed. 852 if (SV->getType() != AllocaType) { 853 if (SV->getType()->getPrimitiveSizeInBits() < 854 AllocaType->getPrimitiveSizeInBits()) 855 SV = Builder.CreateZExt(SV, AllocaType); 856 else { 857 // Truncation may be needed if storing more than the alloca can hold 858 // (undefined behavior). 859 SV = Builder.CreateTrunc(SV, AllocaType); 860 SrcWidth = DestWidth; 861 SrcStoreWidth = DestStoreWidth; 862 } 863 } 864 865 // If this is a big-endian system and the store is narrower than the 866 // full alloca type, we need to do a shift to get the right bits. 867 int ShAmt = 0; 868 if (TD.isBigEndian()) { 869 // On big-endian machines, the lowest bit is stored at the bit offset 870 // from the pointer given by getTypeStoreSizeInBits. This matters for 871 // integers with a bitwidth that is not a multiple of 8. 872 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 873 } else { 874 ShAmt = Offset; 875 } 876 877 // Note: we support negative bitwidths (with shr) which are not defined. 878 // We do this to support (f.e.) stores off the end of a structure where 879 // only some bits in the structure are set. 880 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 881 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 882 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt)); 883 Mask <<= ShAmt; 884 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 885 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt)); 886 Mask = Mask.lshr(-ShAmt); 887 } 888 889 // Mask out the bits we are about to insert from the old value, and or 890 // in the new bits. 891 if (SrcWidth != DestWidth) { 892 assert(DestWidth > SrcWidth); 893 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 894 SV = Builder.CreateOr(Old, SV, "ins"); 895 } 896 return SV; 897} 898 899 900//===----------------------------------------------------------------------===// 901// SRoA Driver 902//===----------------------------------------------------------------------===// 903 904 905bool SROA::runOnFunction(Function &F) { 906 TD = getAnalysisIfAvailable<TargetData>(); 907 908 bool Changed = performPromotion(F); 909 910 // FIXME: ScalarRepl currently depends on TargetData more than it 911 // theoretically needs to. It should be refactored in order to support 912 // target-independent IR. Until this is done, just skip the actual 913 // scalar-replacement portion of this pass. 914 if (!TD) return Changed; 915 916 while (1) { 917 bool LocalChange = performScalarRepl(F); 918 if (!LocalChange) break; // No need to repromote if no scalarrepl 919 Changed = true; 920 LocalChange = performPromotion(F); 921 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 922 } 923 924 return Changed; 925} 926 927namespace { 928class AllocaPromoter : public LoadAndStorePromoter { 929 AllocaInst *AI; 930 DIBuilder *DIB; 931 SmallVector<DbgDeclareInst *, 4> DDIs; 932 SmallVector<DbgValueInst *, 4> DVIs; 933public: 934 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 935 DIBuilder *DB) 936 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {} 937 938 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 939 // Remember which alloca we're promoting (for isInstInList). 940 this->AI = AI; 941 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) { 942 for (Value::use_iterator UI = DebugNode->use_begin(), 943 E = DebugNode->use_end(); UI != E; ++UI) 944 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI)) 945 DDIs.push_back(DDI); 946 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI)) 947 DVIs.push_back(DVI); 948 } 949 950 LoadAndStorePromoter::run(Insts); 951 AI->eraseFromParent(); 952 for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(), 953 E = DDIs.end(); I != E; ++I) { 954 DbgDeclareInst *DDI = *I; 955 DDI->eraseFromParent(); 956 } 957 for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(), 958 E = DVIs.end(); I != E; ++I) { 959 DbgValueInst *DVI = *I; 960 DVI->eraseFromParent(); 961 } 962 } 963 964 virtual bool isInstInList(Instruction *I, 965 const SmallVectorImpl<Instruction*> &Insts) const { 966 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 967 return LI->getOperand(0) == AI; 968 return cast<StoreInst>(I)->getPointerOperand() == AI; 969 } 970 971 virtual void updateDebugInfo(Instruction *Inst) const { 972 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(), 973 E = DDIs.end(); I != E; ++I) { 974 DbgDeclareInst *DDI = *I; 975 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 976 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB); 977 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 978 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB); 979 } 980 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(), 981 E = DVIs.end(); I != E; ++I) { 982 DbgValueInst *DVI = *I; 983 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 984 Instruction *DbgVal = NULL; 985 // If an argument is zero extended then use argument directly. The ZExt 986 // may be zapped by an optimization pass in future. 987 Argument *ExtendedArg = NULL; 988 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 989 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0)); 990 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 991 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0)); 992 if (ExtendedArg) 993 DbgVal = DIB->insertDbgValueIntrinsic(ExtendedArg, 0, 994 DIVariable(DVI->getVariable()), 995 SI); 996 else 997 DbgVal = DIB->insertDbgValueIntrinsic(SI->getOperand(0), 0, 998 DIVariable(DVI->getVariable()), 999 SI); 1000 DbgVal->setDebugLoc(DVI->getDebugLoc()); 1001 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 1002 Instruction *DbgVal = 1003 DIB->insertDbgValueIntrinsic(LI->getOperand(0), 0, 1004 DIVariable(DVI->getVariable()), LI); 1005 DbgVal->setDebugLoc(DVI->getDebugLoc()); 1006 } 1007 } 1008 } 1009}; 1010} // end anon namespace 1011 1012/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1013/// subsequently loaded can be rewritten to load both input pointers and then 1014/// select between the result, allowing the load of the alloca to be promoted. 1015/// From this: 1016/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1017/// %V = load i32* %P2 1018/// to: 1019/// %V1 = load i32* %Alloca -> will be mem2reg'd 1020/// %V2 = load i32* %Other 1021/// %V = select i1 %cond, i32 %V1, i32 %V2 1022/// 1023/// We can do this to a select if its only uses are loads and if the operand to 1024/// the select can be loaded unconditionally. 1025static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1026 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1027 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1028 1029 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1030 UI != UE; ++UI) { 1031 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1032 if (LI == 0 || !LI->isSimple()) return false; 1033 1034 // Both operands to the select need to be dereferencable, either absolutely 1035 // (e.g. allocas) or at this point because we can see other accesses to it. 1036 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1037 LI->getAlignment(), TD)) 1038 return false; 1039 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1040 LI->getAlignment(), TD)) 1041 return false; 1042 } 1043 1044 return true; 1045} 1046 1047/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1048/// subsequently loaded can be rewritten to load both input pointers in the pred 1049/// blocks and then PHI the results, allowing the load of the alloca to be 1050/// promoted. 1051/// From this: 1052/// %P2 = phi [i32* %Alloca, i32* %Other] 1053/// %V = load i32* %P2 1054/// to: 1055/// %V1 = load i32* %Alloca -> will be mem2reg'd 1056/// ... 1057/// %V2 = load i32* %Other 1058/// ... 1059/// %V = phi [i32 %V1, i32 %V2] 1060/// 1061/// We can do this to a select if its only uses are loads and if the operand to 1062/// the select can be loaded unconditionally. 1063static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1064 // For now, we can only do this promotion if the load is in the same block as 1065 // the PHI, and if there are no stores between the phi and load. 1066 // TODO: Allow recursive phi users. 1067 // TODO: Allow stores. 1068 BasicBlock *BB = PN->getParent(); 1069 unsigned MaxAlign = 0; 1070 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1071 UI != UE; ++UI) { 1072 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1073 if (LI == 0 || !LI->isSimple()) return false; 1074 1075 // For now we only allow loads in the same block as the PHI. This is a 1076 // common case that happens when instcombine merges two loads through a PHI. 1077 if (LI->getParent() != BB) return false; 1078 1079 // Ensure that there are no instructions between the PHI and the load that 1080 // could store. 1081 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1082 if (BBI->mayWriteToMemory()) 1083 return false; 1084 1085 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1086 } 1087 1088 // Okay, we know that we have one or more loads in the same block as the PHI. 1089 // We can transform this if it is safe to push the loads into the predecessor 1090 // blocks. The only thing to watch out for is that we can't put a possibly 1091 // trapping load in the predecessor if it is a critical edge. 1092 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1093 BasicBlock *Pred = PN->getIncomingBlock(i); 1094 Value *InVal = PN->getIncomingValue(i); 1095 1096 // If the terminator of the predecessor has side-effects (an invoke), 1097 // there is no safe place to put a load in the predecessor. 1098 if (Pred->getTerminator()->mayHaveSideEffects()) 1099 return false; 1100 1101 // If the value is produced by the terminator of the predecessor 1102 // (an invoke), there is no valid place to put a load in the predecessor. 1103 if (Pred->getTerminator() == InVal) 1104 return false; 1105 1106 // If the predecessor has a single successor, then the edge isn't critical. 1107 if (Pred->getTerminator()->getNumSuccessors() == 1) 1108 continue; 1109 1110 // If this pointer is always safe to load, or if we can prove that there is 1111 // already a load in the block, then we can move the load to the pred block. 1112 if (InVal->isDereferenceablePointer() || 1113 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1114 continue; 1115 1116 return false; 1117 } 1118 1119 return true; 1120} 1121 1122 1123/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1124/// direct (non-volatile) loads and stores to it. If the alloca is close but 1125/// not quite there, this will transform the code to allow promotion. As such, 1126/// it is a non-pure predicate. 1127static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1128 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1129 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1130 1131 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1132 UI != UE; ++UI) { 1133 User *U = *UI; 1134 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1135 if (!LI->isSimple()) 1136 return false; 1137 continue; 1138 } 1139 1140 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1141 if (SI->getOperand(0) == AI || !SI->isSimple()) 1142 return false; // Don't allow a store OF the AI, only INTO the AI. 1143 continue; 1144 } 1145 1146 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1147 // If the condition being selected on is a constant, fold the select, yes 1148 // this does (rarely) happen early on. 1149 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1150 Value *Result = SI->getOperand(1+CI->isZero()); 1151 SI->replaceAllUsesWith(Result); 1152 SI->eraseFromParent(); 1153 1154 // This is very rare and we just scrambled the use list of AI, start 1155 // over completely. 1156 return tryToMakeAllocaBePromotable(AI, TD); 1157 } 1158 1159 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1160 // loads, then we can transform this by rewriting the select. 1161 if (!isSafeSelectToSpeculate(SI, TD)) 1162 return false; 1163 1164 InstsToRewrite.insert(SI); 1165 continue; 1166 } 1167 1168 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1169 if (PN->use_empty()) { // Dead PHIs can be stripped. 1170 InstsToRewrite.insert(PN); 1171 continue; 1172 } 1173 1174 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1175 // in the pred blocks, then we can transform this by rewriting the PHI. 1176 if (!isSafePHIToSpeculate(PN, TD)) 1177 return false; 1178 1179 InstsToRewrite.insert(PN); 1180 continue; 1181 } 1182 1183 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1184 if (onlyUsedByLifetimeMarkers(BCI)) { 1185 InstsToRewrite.insert(BCI); 1186 continue; 1187 } 1188 } 1189 1190 return false; 1191 } 1192 1193 // If there are no instructions to rewrite, then all uses are load/stores and 1194 // we're done! 1195 if (InstsToRewrite.empty()) 1196 return true; 1197 1198 // If we have instructions that need to be rewritten for this to be promotable 1199 // take care of it now. 1200 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1201 if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) { 1202 // This could only be a bitcast used by nothing but lifetime intrinsics. 1203 for (BitCastInst::use_iterator I = BCI->use_begin(), E = BCI->use_end(); 1204 I != E;) { 1205 Use &U = I.getUse(); 1206 ++I; 1207 cast<Instruction>(U.getUser())->eraseFromParent(); 1208 } 1209 BCI->eraseFromParent(); 1210 continue; 1211 } 1212 1213 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1214 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1215 // loads with a new select. 1216 while (!SI->use_empty()) { 1217 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1218 1219 IRBuilder<> Builder(LI); 1220 LoadInst *TrueLoad = 1221 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1222 LoadInst *FalseLoad = 1223 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f"); 1224 1225 // Transfer alignment and TBAA info if present. 1226 TrueLoad->setAlignment(LI->getAlignment()); 1227 FalseLoad->setAlignment(LI->getAlignment()); 1228 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1229 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1230 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1231 } 1232 1233 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1234 V->takeName(LI); 1235 LI->replaceAllUsesWith(V); 1236 LI->eraseFromParent(); 1237 } 1238 1239 // Now that all the loads are gone, the select is gone too. 1240 SI->eraseFromParent(); 1241 continue; 1242 } 1243 1244 // Otherwise, we have a PHI node which allows us to push the loads into the 1245 // predecessors. 1246 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1247 if (PN->use_empty()) { 1248 PN->eraseFromParent(); 1249 continue; 1250 } 1251 1252 Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1253 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1254 PN->getName()+".ld", PN); 1255 1256 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1257 // matter which one we get and if any differ, it doesn't matter. 1258 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1259 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1260 unsigned Align = SomeLoad->getAlignment(); 1261 1262 // Rewrite all loads of the PN to use the new PHI. 1263 while (!PN->use_empty()) { 1264 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1265 LI->replaceAllUsesWith(NewPN); 1266 LI->eraseFromParent(); 1267 } 1268 1269 // Inject loads into all of the pred blocks. Keep track of which blocks we 1270 // insert them into in case we have multiple edges from the same block. 1271 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1272 1273 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1274 BasicBlock *Pred = PN->getIncomingBlock(i); 1275 LoadInst *&Load = InsertedLoads[Pred]; 1276 if (Load == 0) { 1277 Load = new LoadInst(PN->getIncomingValue(i), 1278 PN->getName() + "." + Pred->getName(), 1279 Pred->getTerminator()); 1280 Load->setAlignment(Align); 1281 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1282 } 1283 1284 NewPN->addIncoming(Load, Pred); 1285 } 1286 1287 PN->eraseFromParent(); 1288 } 1289 1290 ++NumAdjusted; 1291 return true; 1292} 1293 1294bool SROA::performPromotion(Function &F) { 1295 std::vector<AllocaInst*> Allocas; 1296 DominatorTree *DT = 0; 1297 if (HasDomTree) 1298 DT = &getAnalysis<DominatorTree>(); 1299 1300 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1301 DIBuilder DIB(*F.getParent()); 1302 bool Changed = false; 1303 SmallVector<Instruction*, 64> Insts; 1304 while (1) { 1305 Allocas.clear(); 1306 1307 // Find allocas that are safe to promote, by looking at all instructions in 1308 // the entry node 1309 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1310 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1311 if (tryToMakeAllocaBePromotable(AI, TD)) 1312 Allocas.push_back(AI); 1313 1314 if (Allocas.empty()) break; 1315 1316 if (HasDomTree) 1317 PromoteMemToReg(Allocas, *DT); 1318 else { 1319 SSAUpdater SSA; 1320 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1321 AllocaInst *AI = Allocas[i]; 1322 1323 // Build list of instructions to promote. 1324 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1325 UI != E; ++UI) 1326 Insts.push_back(cast<Instruction>(*UI)); 1327 AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts); 1328 Insts.clear(); 1329 } 1330 } 1331 NumPromoted += Allocas.size(); 1332 Changed = true; 1333 } 1334 1335 return Changed; 1336} 1337 1338 1339/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1340/// SROA. It must be a struct or array type with a small number of elements. 1341static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1342 Type *T = AI->getAllocatedType(); 1343 // Do not promote any struct into more than 32 separate vars. 1344 if (StructType *ST = dyn_cast<StructType>(T)) 1345 return ST->getNumElements() <= 32; 1346 // Arrays are much less likely to be safe for SROA; only consider 1347 // them if they are very small. 1348 if (ArrayType *AT = dyn_cast<ArrayType>(T)) 1349 return AT->getNumElements() <= 8; 1350 return false; 1351} 1352 1353 1354// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1355// which runs on all of the alloca instructions in the function, removing them 1356// if they are only used by getelementptr instructions. 1357// 1358bool SROA::performScalarRepl(Function &F) { 1359 std::vector<AllocaInst*> WorkList; 1360 1361 // Scan the entry basic block, adding allocas to the worklist. 1362 BasicBlock &BB = F.getEntryBlock(); 1363 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1364 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1365 WorkList.push_back(A); 1366 1367 // Process the worklist 1368 bool Changed = false; 1369 while (!WorkList.empty()) { 1370 AllocaInst *AI = WorkList.back(); 1371 WorkList.pop_back(); 1372 1373 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1374 // with unused elements. 1375 if (AI->use_empty()) { 1376 AI->eraseFromParent(); 1377 Changed = true; 1378 continue; 1379 } 1380 1381 // If this alloca is impossible for us to promote, reject it early. 1382 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1383 continue; 1384 1385 // Check to see if this allocation is only modified by a memcpy/memmove from 1386 // a constant global. If this is the case, we can change all users to use 1387 // the constant global instead. This is commonly produced by the CFE by 1388 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1389 // is only subsequently read. 1390 SmallVector<Instruction *, 4> ToDelete; 1391 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) { 1392 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1393 DEBUG(dbgs() << " memcpy = " << *Copy << '\n'); 1394 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i) 1395 ToDelete[i]->eraseFromParent(); 1396 Constant *TheSrc = cast<Constant>(Copy->getSource()); 1397 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1398 Copy->eraseFromParent(); // Don't mutate the global. 1399 AI->eraseFromParent(); 1400 ++NumGlobals; 1401 Changed = true; 1402 continue; 1403 } 1404 1405 // Check to see if we can perform the core SROA transformation. We cannot 1406 // transform the allocation instruction if it is an array allocation 1407 // (allocations OF arrays are ok though), and an allocation of a scalar 1408 // value cannot be decomposed at all. 1409 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1410 1411 // Do not promote [0 x %struct]. 1412 if (AllocaSize == 0) continue; 1413 1414 // Do not promote any struct whose size is too big. 1415 if (AllocaSize > SRThreshold) continue; 1416 1417 // If the alloca looks like a good candidate for scalar replacement, and if 1418 // all its users can be transformed, then split up the aggregate into its 1419 // separate elements. 1420 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1421 DoScalarReplacement(AI, WorkList); 1422 Changed = true; 1423 continue; 1424 } 1425 1426 // If we can turn this aggregate value (potentially with casts) into a 1427 // simple scalar value that can be mem2reg'd into a register value. 1428 // IsNotTrivial tracks whether this is something that mem2reg could have 1429 // promoted itself. If so, we don't want to transform it needlessly. Note 1430 // that we can't just check based on the type: the alloca may be of an i32 1431 // but that has pointer arithmetic to set byte 3 of it or something. 1432 if (AllocaInst *NewAI = 1433 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1434 NewAI->takeName(AI); 1435 AI->eraseFromParent(); 1436 ++NumConverted; 1437 Changed = true; 1438 continue; 1439 } 1440 1441 // Otherwise, couldn't process this alloca. 1442 } 1443 1444 return Changed; 1445} 1446 1447/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1448/// predicate, do SROA now. 1449void SROA::DoScalarReplacement(AllocaInst *AI, 1450 std::vector<AllocaInst*> &WorkList) { 1451 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1452 SmallVector<AllocaInst*, 32> ElementAllocas; 1453 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1454 ElementAllocas.reserve(ST->getNumContainedTypes()); 1455 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1456 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1457 AI->getAlignment(), 1458 AI->getName() + "." + Twine(i), AI); 1459 ElementAllocas.push_back(NA); 1460 WorkList.push_back(NA); // Add to worklist for recursive processing 1461 } 1462 } else { 1463 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1464 ElementAllocas.reserve(AT->getNumElements()); 1465 Type *ElTy = AT->getElementType(); 1466 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1467 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1468 AI->getName() + "." + Twine(i), AI); 1469 ElementAllocas.push_back(NA); 1470 WorkList.push_back(NA); // Add to worklist for recursive processing 1471 } 1472 } 1473 1474 // Now that we have created the new alloca instructions, rewrite all the 1475 // uses of the old alloca. 1476 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1477 1478 // Now erase any instructions that were made dead while rewriting the alloca. 1479 DeleteDeadInstructions(); 1480 AI->eraseFromParent(); 1481 1482 ++NumReplaced; 1483} 1484 1485/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1486/// recursively including all their operands that become trivially dead. 1487void SROA::DeleteDeadInstructions() { 1488 while (!DeadInsts.empty()) { 1489 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1490 1491 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1492 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1493 // Zero out the operand and see if it becomes trivially dead. 1494 // (But, don't add allocas to the dead instruction list -- they are 1495 // already on the worklist and will be deleted separately.) 1496 *OI = 0; 1497 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1498 DeadInsts.push_back(U); 1499 } 1500 1501 I->eraseFromParent(); 1502 } 1503} 1504 1505/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1506/// performing scalar replacement of alloca AI. The results are flagged in 1507/// the Info parameter. Offset indicates the position within AI that is 1508/// referenced by this instruction. 1509void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1510 AllocaInfo &Info) { 1511 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1512 Instruction *User = cast<Instruction>(*UI); 1513 1514 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1515 isSafeForScalarRepl(BC, Offset, Info); 1516 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1517 uint64_t GEPOffset = Offset; 1518 isSafeGEP(GEPI, GEPOffset, Info); 1519 if (!Info.isUnsafe) 1520 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1521 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1522 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1523 if (Length == 0) 1524 return MarkUnsafe(Info, User); 1525 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1526 UI.getOperandNo() == 0, Info, MI, 1527 true /*AllowWholeAccess*/); 1528 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1529 if (!LI->isSimple()) 1530 return MarkUnsafe(Info, User); 1531 Type *LIType = LI->getType(); 1532 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1533 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1534 Info.hasALoadOrStore = true; 1535 1536 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1537 // Store is ok if storing INTO the pointer, not storing the pointer 1538 if (!SI->isSimple() || SI->getOperand(0) == I) 1539 return MarkUnsafe(Info, User); 1540 1541 Type *SIType = SI->getOperand(0)->getType(); 1542 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1543 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1544 Info.hasALoadOrStore = true; 1545 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1546 if (II->getIntrinsicID() != Intrinsic::lifetime_start && 1547 II->getIntrinsicID() != Intrinsic::lifetime_end) 1548 return MarkUnsafe(Info, User); 1549 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1550 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1551 } else { 1552 return MarkUnsafe(Info, User); 1553 } 1554 if (Info.isUnsafe) return; 1555 } 1556} 1557 1558 1559/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1560/// derived from the alloca, we can often still split the alloca into elements. 1561/// This is useful if we have a large alloca where one element is phi'd 1562/// together somewhere: we can SRoA and promote all the other elements even if 1563/// we end up not being able to promote this one. 1564/// 1565/// All we require is that the uses of the PHI do not index into other parts of 1566/// the alloca. The most important use case for this is single load and stores 1567/// that are PHI'd together, which can happen due to code sinking. 1568void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1569 AllocaInfo &Info) { 1570 // If we've already checked this PHI, don't do it again. 1571 if (PHINode *PN = dyn_cast<PHINode>(I)) 1572 if (!Info.CheckedPHIs.insert(PN)) 1573 return; 1574 1575 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1576 Instruction *User = cast<Instruction>(*UI); 1577 1578 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1579 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1580 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1581 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1582 // but would have to prove that we're staying inside of an element being 1583 // promoted. 1584 if (!GEPI->hasAllZeroIndices()) 1585 return MarkUnsafe(Info, User); 1586 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1587 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1588 if (!LI->isSimple()) 1589 return MarkUnsafe(Info, User); 1590 Type *LIType = LI->getType(); 1591 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1592 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1593 Info.hasALoadOrStore = true; 1594 1595 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1596 // Store is ok if storing INTO the pointer, not storing the pointer 1597 if (!SI->isSimple() || SI->getOperand(0) == I) 1598 return MarkUnsafe(Info, User); 1599 1600 Type *SIType = SI->getOperand(0)->getType(); 1601 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1602 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1603 Info.hasALoadOrStore = true; 1604 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1605 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1606 } else { 1607 return MarkUnsafe(Info, User); 1608 } 1609 if (Info.isUnsafe) return; 1610 } 1611} 1612 1613/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1614/// replacement. It is safe when all the indices are constant, in-bounds 1615/// references, and when the resulting offset corresponds to an element within 1616/// the alloca type. The results are flagged in the Info parameter. Upon 1617/// return, Offset is adjusted as specified by the GEP indices. 1618void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1619 uint64_t &Offset, AllocaInfo &Info) { 1620 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1621 if (GEPIt == E) 1622 return; 1623 1624 // Walk through the GEP type indices, checking the types that this indexes 1625 // into. 1626 for (; GEPIt != E; ++GEPIt) { 1627 // Ignore struct elements, no extra checking needed for these. 1628 if ((*GEPIt)->isStructTy()) 1629 continue; 1630 1631 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1632 if (!IdxVal) 1633 return MarkUnsafe(Info, GEPI); 1634 } 1635 1636 // Compute the offset due to this GEP and check if the alloca has a 1637 // component element at that offset. 1638 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1639 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1640 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1641 MarkUnsafe(Info, GEPI); 1642} 1643 1644/// isHomogeneousAggregate - Check if type T is a struct or array containing 1645/// elements of the same type (which is always true for arrays). If so, 1646/// return true with NumElts and EltTy set to the number of elements and the 1647/// element type, respectively. 1648static bool isHomogeneousAggregate(Type *T, unsigned &NumElts, 1649 Type *&EltTy) { 1650 if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1651 NumElts = AT->getNumElements(); 1652 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1653 return true; 1654 } 1655 if (StructType *ST = dyn_cast<StructType>(T)) { 1656 NumElts = ST->getNumContainedTypes(); 1657 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1658 for (unsigned n = 1; n < NumElts; ++n) { 1659 if (ST->getContainedType(n) != EltTy) 1660 return false; 1661 } 1662 return true; 1663 } 1664 return false; 1665} 1666 1667/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1668/// "homogeneous" aggregates with the same element type and number of elements. 1669static bool isCompatibleAggregate(Type *T1, Type *T2) { 1670 if (T1 == T2) 1671 return true; 1672 1673 unsigned NumElts1, NumElts2; 1674 Type *EltTy1, *EltTy2; 1675 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1676 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1677 NumElts1 == NumElts2 && 1678 EltTy1 == EltTy2) 1679 return true; 1680 1681 return false; 1682} 1683 1684/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1685/// alloca or has an offset and size that corresponds to a component element 1686/// within it. The offset checked here may have been formed from a GEP with a 1687/// pointer bitcasted to a different type. 1688/// 1689/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1690/// unit. If false, it only allows accesses known to be in a single element. 1691void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1692 Type *MemOpType, bool isStore, 1693 AllocaInfo &Info, Instruction *TheAccess, 1694 bool AllowWholeAccess) { 1695 // Check if this is a load/store of the entire alloca. 1696 if (Offset == 0 && AllowWholeAccess && 1697 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1698 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1699 // loads/stores (which are essentially the same as the MemIntrinsics with 1700 // regard to copying padding between elements). But, if an alloca is 1701 // flagged as both a source and destination of such operations, we'll need 1702 // to check later for padding between elements. 1703 if (!MemOpType || MemOpType->isIntegerTy()) { 1704 if (isStore) 1705 Info.isMemCpyDst = true; 1706 else 1707 Info.isMemCpySrc = true; 1708 return; 1709 } 1710 // This is also safe for references using a type that is compatible with 1711 // the type of the alloca, so that loads/stores can be rewritten using 1712 // insertvalue/extractvalue. 1713 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1714 Info.hasSubelementAccess = true; 1715 return; 1716 } 1717 } 1718 // Check if the offset/size correspond to a component within the alloca type. 1719 Type *T = Info.AI->getAllocatedType(); 1720 if (TypeHasComponent(T, Offset, MemSize)) { 1721 Info.hasSubelementAccess = true; 1722 return; 1723 } 1724 1725 return MarkUnsafe(Info, TheAccess); 1726} 1727 1728/// TypeHasComponent - Return true if T has a component type with the 1729/// specified offset and size. If Size is zero, do not check the size. 1730bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { 1731 Type *EltTy; 1732 uint64_t EltSize; 1733 if (StructType *ST = dyn_cast<StructType>(T)) { 1734 const StructLayout *Layout = TD->getStructLayout(ST); 1735 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1736 EltTy = ST->getContainedType(EltIdx); 1737 EltSize = TD->getTypeAllocSize(EltTy); 1738 Offset -= Layout->getElementOffset(EltIdx); 1739 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1740 EltTy = AT->getElementType(); 1741 EltSize = TD->getTypeAllocSize(EltTy); 1742 if (Offset >= AT->getNumElements() * EltSize) 1743 return false; 1744 Offset %= EltSize; 1745 } else { 1746 return false; 1747 } 1748 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1749 return true; 1750 // Check if the component spans multiple elements. 1751 if (Offset + Size > EltSize) 1752 return false; 1753 return TypeHasComponent(EltTy, Offset, Size); 1754} 1755 1756/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1757/// the instruction I, which references it, to use the separate elements. 1758/// Offset indicates the position within AI that is referenced by this 1759/// instruction. 1760void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1761 SmallVector<AllocaInst*, 32> &NewElts) { 1762 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1763 Use &TheUse = UI.getUse(); 1764 Instruction *User = cast<Instruction>(*UI++); 1765 1766 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1767 RewriteBitCast(BC, AI, Offset, NewElts); 1768 continue; 1769 } 1770 1771 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1772 RewriteGEP(GEPI, AI, Offset, NewElts); 1773 continue; 1774 } 1775 1776 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1777 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1778 uint64_t MemSize = Length->getZExtValue(); 1779 if (Offset == 0 && 1780 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1781 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1782 // Otherwise the intrinsic can only touch a single element and the 1783 // address operand will be updated, so nothing else needs to be done. 1784 continue; 1785 } 1786 1787 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1788 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 1789 II->getIntrinsicID() == Intrinsic::lifetime_end) { 1790 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts); 1791 } 1792 continue; 1793 } 1794 1795 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1796 Type *LIType = LI->getType(); 1797 1798 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1799 // Replace: 1800 // %res = load { i32, i32 }* %alloc 1801 // with: 1802 // %load.0 = load i32* %alloc.0 1803 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1804 // %load.1 = load i32* %alloc.1 1805 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1806 // (Also works for arrays instead of structs) 1807 Value *Insert = UndefValue::get(LIType); 1808 IRBuilder<> Builder(LI); 1809 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1810 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1811 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1812 } 1813 LI->replaceAllUsesWith(Insert); 1814 DeadInsts.push_back(LI); 1815 } else if (LIType->isIntegerTy() && 1816 TD->getTypeAllocSize(LIType) == 1817 TD->getTypeAllocSize(AI->getAllocatedType())) { 1818 // If this is a load of the entire alloca to an integer, rewrite it. 1819 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1820 } 1821 continue; 1822 } 1823 1824 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1825 Value *Val = SI->getOperand(0); 1826 Type *SIType = Val->getType(); 1827 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1828 // Replace: 1829 // store { i32, i32 } %val, { i32, i32 }* %alloc 1830 // with: 1831 // %val.0 = extractvalue { i32, i32 } %val, 0 1832 // store i32 %val.0, i32* %alloc.0 1833 // %val.1 = extractvalue { i32, i32 } %val, 1 1834 // store i32 %val.1, i32* %alloc.1 1835 // (Also works for arrays instead of structs) 1836 IRBuilder<> Builder(SI); 1837 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1838 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1839 Builder.CreateStore(Extract, NewElts[i]); 1840 } 1841 DeadInsts.push_back(SI); 1842 } else if (SIType->isIntegerTy() && 1843 TD->getTypeAllocSize(SIType) == 1844 TD->getTypeAllocSize(AI->getAllocatedType())) { 1845 // If this is a store of the entire alloca from an integer, rewrite it. 1846 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1847 } 1848 continue; 1849 } 1850 1851 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1852 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1853 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1854 // the new pointer. 1855 if (!isa<AllocaInst>(I)) continue; 1856 1857 assert(Offset == 0 && NewElts[0] && 1858 "Direct alloca use should have a zero offset"); 1859 1860 // If we have a use of the alloca, we know the derived uses will be 1861 // utilizing just the first element of the scalarized result. Insert a 1862 // bitcast of the first alloca before the user as required. 1863 AllocaInst *NewAI = NewElts[0]; 1864 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1865 NewAI->moveBefore(BCI); 1866 TheUse = BCI; 1867 continue; 1868 } 1869 } 1870} 1871 1872/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1873/// and recursively continue updating all of its uses. 1874void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1875 SmallVector<AllocaInst*, 32> &NewElts) { 1876 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1877 if (BC->getOperand(0) != AI) 1878 return; 1879 1880 // The bitcast references the original alloca. Replace its uses with 1881 // references to the alloca containing offset zero (which is normally at 1882 // index zero, but might not be in cases involving structs with elements 1883 // of size zero). 1884 Type *T = AI->getAllocatedType(); 1885 uint64_t EltOffset = 0; 1886 Type *IdxTy; 1887 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1888 Instruction *Val = NewElts[Idx]; 1889 if (Val->getType() != BC->getDestTy()) { 1890 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1891 Val->takeName(BC); 1892 } 1893 BC->replaceAllUsesWith(Val); 1894 DeadInsts.push_back(BC); 1895} 1896 1897/// FindElementAndOffset - Return the index of the element containing Offset 1898/// within the specified type, which must be either a struct or an array. 1899/// Sets T to the type of the element and Offset to the offset within that 1900/// element. IdxTy is set to the type of the index result to be used in a 1901/// GEP instruction. 1902uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, 1903 Type *&IdxTy) { 1904 uint64_t Idx = 0; 1905 if (StructType *ST = dyn_cast<StructType>(T)) { 1906 const StructLayout *Layout = TD->getStructLayout(ST); 1907 Idx = Layout->getElementContainingOffset(Offset); 1908 T = ST->getContainedType(Idx); 1909 Offset -= Layout->getElementOffset(Idx); 1910 IdxTy = Type::getInt32Ty(T->getContext()); 1911 return Idx; 1912 } 1913 ArrayType *AT = cast<ArrayType>(T); 1914 T = AT->getElementType(); 1915 uint64_t EltSize = TD->getTypeAllocSize(T); 1916 Idx = Offset / EltSize; 1917 Offset -= Idx * EltSize; 1918 IdxTy = Type::getInt64Ty(T->getContext()); 1919 return Idx; 1920} 1921 1922/// RewriteGEP - Check if this GEP instruction moves the pointer across 1923/// elements of the alloca that are being split apart, and if so, rewrite 1924/// the GEP to be relative to the new element. 1925void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1926 SmallVector<AllocaInst*, 32> &NewElts) { 1927 uint64_t OldOffset = Offset; 1928 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1929 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1930 1931 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1932 1933 Type *T = AI->getAllocatedType(); 1934 Type *IdxTy; 1935 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1936 if (GEPI->getOperand(0) == AI) 1937 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1938 1939 T = AI->getAllocatedType(); 1940 uint64_t EltOffset = Offset; 1941 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1942 1943 // If this GEP does not move the pointer across elements of the alloca 1944 // being split, then it does not needs to be rewritten. 1945 if (Idx == OldIdx) 1946 return; 1947 1948 Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1949 SmallVector<Value*, 8> NewArgs; 1950 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1951 while (EltOffset != 0) { 1952 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1953 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1954 } 1955 Instruction *Val = NewElts[Idx]; 1956 if (NewArgs.size() > 1) { 1957 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI); 1958 Val->takeName(GEPI); 1959 } 1960 if (Val->getType() != GEPI->getType()) 1961 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1962 GEPI->replaceAllUsesWith(Val); 1963 DeadInsts.push_back(GEPI); 1964} 1965 1966/// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it 1967/// to mark the lifetime of the scalarized memory. 1968void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 1969 uint64_t Offset, 1970 SmallVector<AllocaInst*, 32> &NewElts) { 1971 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0)); 1972 // Put matching lifetime markers on everything from Offset up to 1973 // Offset+OldSize. 1974 Type *AIType = AI->getAllocatedType(); 1975 uint64_t NewOffset = Offset; 1976 Type *IdxTy; 1977 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy); 1978 1979 IRBuilder<> Builder(II); 1980 uint64_t Size = OldSize->getLimitedValue(); 1981 1982 if (NewOffset) { 1983 // Splice the first element and index 'NewOffset' bytes in. SROA will 1984 // split the alloca again later. 1985 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy()); 1986 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset)); 1987 1988 IdxTy = NewElts[Idx]->getAllocatedType(); 1989 uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset; 1990 if (EltSize > Size) { 1991 EltSize = Size; 1992 Size = 0; 1993 } else { 1994 Size -= EltSize; 1995 } 1996 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 1997 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize)); 1998 else 1999 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize)); 2000 ++Idx; 2001 } 2002 2003 for (; Idx != NewElts.size() && Size; ++Idx) { 2004 IdxTy = NewElts[Idx]->getAllocatedType(); 2005 uint64_t EltSize = TD->getTypeAllocSize(IdxTy); 2006 if (EltSize > Size) { 2007 EltSize = Size; 2008 Size = 0; 2009 } else { 2010 Size -= EltSize; 2011 } 2012 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2013 Builder.CreateLifetimeStart(NewElts[Idx], 2014 Builder.getInt64(EltSize)); 2015 else 2016 Builder.CreateLifetimeEnd(NewElts[Idx], 2017 Builder.getInt64(EltSize)); 2018 } 2019 DeadInsts.push_back(II); 2020} 2021 2022/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2023/// Rewrite it to copy or set the elements of the scalarized memory. 2024void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2025 AllocaInst *AI, 2026 SmallVector<AllocaInst*, 32> &NewElts) { 2027 // If this is a memcpy/memmove, construct the other pointer as the 2028 // appropriate type. The "Other" pointer is the pointer that goes to memory 2029 // that doesn't have anything to do with the alloca that we are promoting. For 2030 // memset, this Value* stays null. 2031 Value *OtherPtr = 0; 2032 unsigned MemAlignment = MI->getAlignment(); 2033 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2034 if (Inst == MTI->getRawDest()) 2035 OtherPtr = MTI->getRawSource(); 2036 else { 2037 assert(Inst == MTI->getRawSource()); 2038 OtherPtr = MTI->getRawDest(); 2039 } 2040 } 2041 2042 // If there is an other pointer, we want to convert it to the same pointer 2043 // type as AI has, so we can GEP through it safely. 2044 if (OtherPtr) { 2045 unsigned AddrSpace = 2046 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2047 2048 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2049 // optimization, but it's also required to detect the corner case where 2050 // both pointer operands are referencing the same memory, and where 2051 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2052 // function is only called for mem intrinsics that access the whole 2053 // aggregate, so non-zero GEPs are not an issue here.) 2054 OtherPtr = OtherPtr->stripPointerCasts(); 2055 2056 // Copying the alloca to itself is a no-op: just delete it. 2057 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2058 // This code will run twice for a no-op memcpy -- once for each operand. 2059 // Put only one reference to MI on the DeadInsts list. 2060 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2061 E = DeadInsts.end(); I != E; ++I) 2062 if (*I == MI) return; 2063 DeadInsts.push_back(MI); 2064 return; 2065 } 2066 2067 // If the pointer is not the right type, insert a bitcast to the right 2068 // type. 2069 Type *NewTy = 2070 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2071 2072 if (OtherPtr->getType() != NewTy) 2073 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2074 } 2075 2076 // Process each element of the aggregate. 2077 bool SROADest = MI->getRawDest() == Inst; 2078 2079 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2080 2081 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2082 // If this is a memcpy/memmove, emit a GEP of the other element address. 2083 Value *OtherElt = 0; 2084 unsigned OtherEltAlign = MemAlignment; 2085 2086 if (OtherPtr) { 2087 Value *Idx[2] = { Zero, 2088 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2089 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, 2090 OtherPtr->getName()+"."+Twine(i), 2091 MI); 2092 uint64_t EltOffset; 2093 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2094 Type *OtherTy = OtherPtrTy->getElementType(); 2095 if (StructType *ST = dyn_cast<StructType>(OtherTy)) { 2096 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2097 } else { 2098 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2099 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2100 } 2101 2102 // The alignment of the other pointer is the guaranteed alignment of the 2103 // element, which is affected by both the known alignment of the whole 2104 // mem intrinsic and the alignment of the element. If the alignment of 2105 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2106 // known alignment is just 4 bytes. 2107 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2108 } 2109 2110 Value *EltPtr = NewElts[i]; 2111 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2112 2113 // If we got down to a scalar, insert a load or store as appropriate. 2114 if (EltTy->isSingleValueType()) { 2115 if (isa<MemTransferInst>(MI)) { 2116 if (SROADest) { 2117 // From Other to Alloca. 2118 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2119 new StoreInst(Elt, EltPtr, MI); 2120 } else { 2121 // From Alloca to Other. 2122 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2123 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2124 } 2125 continue; 2126 } 2127 assert(isa<MemSetInst>(MI)); 2128 2129 // If the stored element is zero (common case), just store a null 2130 // constant. 2131 Constant *StoreVal; 2132 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2133 if (CI->isZero()) { 2134 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2135 } else { 2136 // If EltTy is a vector type, get the element type. 2137 Type *ValTy = EltTy->getScalarType(); 2138 2139 // Construct an integer with the right value. 2140 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2141 APInt OneVal(EltSize, CI->getZExtValue()); 2142 APInt TotalVal(OneVal); 2143 // Set each byte. 2144 for (unsigned i = 0; 8*i < EltSize; ++i) { 2145 TotalVal = TotalVal.shl(8); 2146 TotalVal |= OneVal; 2147 } 2148 2149 // Convert the integer value to the appropriate type. 2150 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2151 if (ValTy->isPointerTy()) 2152 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2153 else if (ValTy->isFloatingPointTy()) 2154 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2155 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2156 2157 // If the requested value was a vector constant, create it. 2158 if (EltTy->isVectorTy()) { 2159 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements(); 2160 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2161 StoreVal = ConstantVector::get(Elts); 2162 } 2163 } 2164 new StoreInst(StoreVal, EltPtr, MI); 2165 continue; 2166 } 2167 // Otherwise, if we're storing a byte variable, use a memset call for 2168 // this element. 2169 } 2170 2171 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2172 if (!EltSize) 2173 continue; 2174 2175 IRBuilder<> Builder(MI); 2176 2177 // Finally, insert the meminst for this element. 2178 if (isa<MemSetInst>(MI)) { 2179 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2180 MI->isVolatile()); 2181 } else { 2182 assert(isa<MemTransferInst>(MI)); 2183 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2184 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2185 2186 if (isa<MemCpyInst>(MI)) 2187 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2188 else 2189 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2190 } 2191 } 2192 DeadInsts.push_back(MI); 2193} 2194 2195/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2196/// overwrites the entire allocation. Extract out the pieces of the stored 2197/// integer and store them individually. 2198void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2199 SmallVector<AllocaInst*, 32> &NewElts){ 2200 // Extract each element out of the integer according to its structure offset 2201 // and store the element value to the individual alloca. 2202 Value *SrcVal = SI->getOperand(0); 2203 Type *AllocaEltTy = AI->getAllocatedType(); 2204 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2205 2206 IRBuilder<> Builder(SI); 2207 2208 // Handle tail padding by extending the operand 2209 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2210 SrcVal = Builder.CreateZExt(SrcVal, 2211 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2212 2213 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2214 << '\n'); 2215 2216 // There are two forms here: AI could be an array or struct. Both cases 2217 // have different ways to compute the element offset. 2218 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2219 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2220 2221 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2222 // Get the number of bits to shift SrcVal to get the value. 2223 Type *FieldTy = EltSTy->getElementType(i); 2224 uint64_t Shift = Layout->getElementOffsetInBits(i); 2225 2226 if (TD->isBigEndian()) 2227 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2228 2229 Value *EltVal = SrcVal; 2230 if (Shift) { 2231 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2232 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2233 } 2234 2235 // Truncate down to an integer of the right size. 2236 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2237 2238 // Ignore zero sized fields like {}, they obviously contain no data. 2239 if (FieldSizeBits == 0) continue; 2240 2241 if (FieldSizeBits != AllocaSizeBits) 2242 EltVal = Builder.CreateTrunc(EltVal, 2243 IntegerType::get(SI->getContext(), FieldSizeBits)); 2244 Value *DestField = NewElts[i]; 2245 if (EltVal->getType() == FieldTy) { 2246 // Storing to an integer field of this size, just do it. 2247 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2248 // Bitcast to the right element type (for fp/vector values). 2249 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2250 } else { 2251 // Otherwise, bitcast the dest pointer (for aggregates). 2252 DestField = Builder.CreateBitCast(DestField, 2253 PointerType::getUnqual(EltVal->getType())); 2254 } 2255 new StoreInst(EltVal, DestField, SI); 2256 } 2257 2258 } else { 2259 ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2260 Type *ArrayEltTy = ATy->getElementType(); 2261 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2262 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2263 2264 uint64_t Shift; 2265 2266 if (TD->isBigEndian()) 2267 Shift = AllocaSizeBits-ElementOffset; 2268 else 2269 Shift = 0; 2270 2271 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2272 // Ignore zero sized fields like {}, they obviously contain no data. 2273 if (ElementSizeBits == 0) continue; 2274 2275 Value *EltVal = SrcVal; 2276 if (Shift) { 2277 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2278 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2279 } 2280 2281 // Truncate down to an integer of the right size. 2282 if (ElementSizeBits != AllocaSizeBits) 2283 EltVal = Builder.CreateTrunc(EltVal, 2284 IntegerType::get(SI->getContext(), 2285 ElementSizeBits)); 2286 Value *DestField = NewElts[i]; 2287 if (EltVal->getType() == ArrayEltTy) { 2288 // Storing to an integer field of this size, just do it. 2289 } else if (ArrayEltTy->isFloatingPointTy() || 2290 ArrayEltTy->isVectorTy()) { 2291 // Bitcast to the right element type (for fp/vector values). 2292 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2293 } else { 2294 // Otherwise, bitcast the dest pointer (for aggregates). 2295 DestField = Builder.CreateBitCast(DestField, 2296 PointerType::getUnqual(EltVal->getType())); 2297 } 2298 new StoreInst(EltVal, DestField, SI); 2299 2300 if (TD->isBigEndian()) 2301 Shift -= ElementOffset; 2302 else 2303 Shift += ElementOffset; 2304 } 2305 } 2306 2307 DeadInsts.push_back(SI); 2308} 2309 2310/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2311/// an integer. Load the individual pieces to form the aggregate value. 2312void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2313 SmallVector<AllocaInst*, 32> &NewElts) { 2314 // Extract each element out of the NewElts according to its structure offset 2315 // and form the result value. 2316 Type *AllocaEltTy = AI->getAllocatedType(); 2317 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2318 2319 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2320 << '\n'); 2321 2322 // There are two forms here: AI could be an array or struct. Both cases 2323 // have different ways to compute the element offset. 2324 const StructLayout *Layout = 0; 2325 uint64_t ArrayEltBitOffset = 0; 2326 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2327 Layout = TD->getStructLayout(EltSTy); 2328 } else { 2329 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2330 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2331 } 2332 2333 Value *ResultVal = 2334 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2335 2336 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2337 // Load the value from the alloca. If the NewElt is an aggregate, cast 2338 // the pointer to an integer of the same size before doing the load. 2339 Value *SrcField = NewElts[i]; 2340 Type *FieldTy = 2341 cast<PointerType>(SrcField->getType())->getElementType(); 2342 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2343 2344 // Ignore zero sized fields like {}, they obviously contain no data. 2345 if (FieldSizeBits == 0) continue; 2346 2347 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2348 FieldSizeBits); 2349 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2350 !FieldTy->isVectorTy()) 2351 SrcField = new BitCastInst(SrcField, 2352 PointerType::getUnqual(FieldIntTy), 2353 "", LI); 2354 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2355 2356 // If SrcField is a fp or vector of the right size but that isn't an 2357 // integer type, bitcast to an integer so we can shift it. 2358 if (SrcField->getType() != FieldIntTy) 2359 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2360 2361 // Zero extend the field to be the same size as the final alloca so that 2362 // we can shift and insert it. 2363 if (SrcField->getType() != ResultVal->getType()) 2364 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2365 2366 // Determine the number of bits to shift SrcField. 2367 uint64_t Shift; 2368 if (Layout) // Struct case. 2369 Shift = Layout->getElementOffsetInBits(i); 2370 else // Array case. 2371 Shift = i*ArrayEltBitOffset; 2372 2373 if (TD->isBigEndian()) 2374 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2375 2376 if (Shift) { 2377 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2378 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2379 } 2380 2381 // Don't create an 'or x, 0' on the first iteration. 2382 if (!isa<Constant>(ResultVal) || 2383 !cast<Constant>(ResultVal)->isNullValue()) 2384 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2385 else 2386 ResultVal = SrcField; 2387 } 2388 2389 // Handle tail padding by truncating the result 2390 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2391 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2392 2393 LI->replaceAllUsesWith(ResultVal); 2394 DeadInsts.push_back(LI); 2395} 2396 2397/// HasPadding - Return true if the specified type has any structure or 2398/// alignment padding in between the elements that would be split apart 2399/// by SROA; return false otherwise. 2400static bool HasPadding(Type *Ty, const TargetData &TD) { 2401 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2402 Ty = ATy->getElementType(); 2403 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2404 } 2405 2406 // SROA currently handles only Arrays and Structs. 2407 StructType *STy = cast<StructType>(Ty); 2408 const StructLayout *SL = TD.getStructLayout(STy); 2409 unsigned PrevFieldBitOffset = 0; 2410 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2411 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2412 2413 // Check to see if there is any padding between this element and the 2414 // previous one. 2415 if (i) { 2416 unsigned PrevFieldEnd = 2417 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2418 if (PrevFieldEnd < FieldBitOffset) 2419 return true; 2420 } 2421 PrevFieldBitOffset = FieldBitOffset; 2422 } 2423 // Check for tail padding. 2424 if (unsigned EltCount = STy->getNumElements()) { 2425 unsigned PrevFieldEnd = PrevFieldBitOffset + 2426 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2427 if (PrevFieldEnd < SL->getSizeInBits()) 2428 return true; 2429 } 2430 return false; 2431} 2432 2433/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2434/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2435/// or 1 if safe after canonicalization has been performed. 2436bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2437 // Loop over the use list of the alloca. We can only transform it if all of 2438 // the users are safe to transform. 2439 AllocaInfo Info(AI); 2440 2441 isSafeForScalarRepl(AI, 0, Info); 2442 if (Info.isUnsafe) { 2443 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2444 return false; 2445 } 2446 2447 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2448 // source and destination, we have to be careful. In particular, the memcpy 2449 // could be moving around elements that live in structure padding of the LLVM 2450 // types, but may actually be used. In these cases, we refuse to promote the 2451 // struct. 2452 if (Info.isMemCpySrc && Info.isMemCpyDst && 2453 HasPadding(AI->getAllocatedType(), *TD)) 2454 return false; 2455 2456 // If the alloca never has an access to just *part* of it, but is accessed 2457 // via loads and stores, then we should use ConvertToScalarInfo to promote 2458 // the alloca instead of promoting each piece at a time and inserting fission 2459 // and fusion code. 2460 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2461 // If the struct/array just has one element, use basic SRoA. 2462 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2463 if (ST->getNumElements() > 1) return false; 2464 } else { 2465 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2466 return false; 2467 } 2468 } 2469 2470 return true; 2471} 2472 2473 2474 2475/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2476/// some part of a constant global variable. This intentionally only accepts 2477/// constant expressions because we don't can't rewrite arbitrary instructions. 2478static bool PointsToConstantGlobal(Value *V) { 2479 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2480 return GV->isConstant(); 2481 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2482 if (CE->getOpcode() == Instruction::BitCast || 2483 CE->getOpcode() == Instruction::GetElementPtr) 2484 return PointsToConstantGlobal(CE->getOperand(0)); 2485 return false; 2486} 2487 2488/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2489/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2490/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2491/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2492/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2493/// the alloca, and if the source pointer is a pointer to a constant global, we 2494/// can optimize this. 2495static bool 2496isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2497 bool isOffset, 2498 SmallVector<Instruction *, 4> &LifetimeMarkers) { 2499 // We track lifetime intrinsics as we encounter them. If we decide to go 2500 // ahead and replace the value with the global, this lets the caller quickly 2501 // eliminate the markers. 2502 2503 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2504 User *U = cast<Instruction>(*UI); 2505 2506 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2507 // Ignore non-volatile loads, they are always ok. 2508 if (!LI->isSimple()) return false; 2509 continue; 2510 } 2511 2512 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2513 // If uses of the bitcast are ok, we are ok. 2514 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset, 2515 LifetimeMarkers)) 2516 return false; 2517 continue; 2518 } 2519 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2520 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2521 // doesn't, it does. 2522 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2523 isOffset || !GEP->hasAllZeroIndices(), 2524 LifetimeMarkers)) 2525 return false; 2526 continue; 2527 } 2528 2529 if (CallSite CS = U) { 2530 // If this is the function being called then we treat it like a load and 2531 // ignore it. 2532 if (CS.isCallee(UI)) 2533 continue; 2534 2535 // If this is a readonly/readnone call site, then we know it is just a 2536 // load (but one that potentially returns the value itself), so we can 2537 // ignore it if we know that the value isn't captured. 2538 unsigned ArgNo = CS.getArgumentNo(UI); 2539 if (CS.onlyReadsMemory() && 2540 (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo))) 2541 continue; 2542 2543 // If this is being passed as a byval argument, the caller is making a 2544 // copy, so it is only a read of the alloca. 2545 if (CS.isByValArgument(ArgNo)) 2546 continue; 2547 } 2548 2549 // Lifetime intrinsics can be handled by the caller. 2550 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 2551 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 2552 II->getIntrinsicID() == Intrinsic::lifetime_end) { 2553 assert(II->use_empty() && "Lifetime markers have no result to use!"); 2554 LifetimeMarkers.push_back(II); 2555 continue; 2556 } 2557 } 2558 2559 // If this is isn't our memcpy/memmove, reject it as something we can't 2560 // handle. 2561 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2562 if (MI == 0) 2563 return false; 2564 2565 // If the transfer is using the alloca as a source of the transfer, then 2566 // ignore it since it is a load (unless the transfer is volatile). 2567 if (UI.getOperandNo() == 1) { 2568 if (MI->isVolatile()) return false; 2569 continue; 2570 } 2571 2572 // If we already have seen a copy, reject the second one. 2573 if (TheCopy) return false; 2574 2575 // If the pointer has been offset from the start of the alloca, we can't 2576 // safely handle this. 2577 if (isOffset) return false; 2578 2579 // If the memintrinsic isn't using the alloca as the dest, reject it. 2580 if (UI.getOperandNo() != 0) return false; 2581 2582 // If the source of the memcpy/move is not a constant global, reject it. 2583 if (!PointsToConstantGlobal(MI->getSource())) 2584 return false; 2585 2586 // Otherwise, the transform is safe. Remember the copy instruction. 2587 TheCopy = MI; 2588 } 2589 return true; 2590} 2591 2592/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2593/// modified by a copy from a constant global. If we can prove this, we can 2594/// replace any uses of the alloca with uses of the global directly. 2595MemTransferInst * 2596SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI, 2597 SmallVector<Instruction*, 4> &ToDelete) { 2598 MemTransferInst *TheCopy = 0; 2599 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete)) 2600 return TheCopy; 2601 return 0; 2602} 2603