1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This transformation implements the well known scalar replacement of 11 // aggregates transformation. This xform breaks up alloca instructions of 12 // aggregate type (structure or array) into individual alloca instructions for 13 // each member (if possible). Then, if possible, it transforms the individual 14 // alloca instructions into nice clean scalar SSA form. 15 // 16 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because they 17 // often interact, especially for C++ programs. As such, iterating between 18 // SRoA, then Mem2Reg until we run out of things to promote works well. 19 // 20 //===----------------------------------------------------------------------===// 21 22 #define DEBUG_TYPE "scalarrepl" 23 #include "llvm/Transforms/Scalar.h" 24 #include "llvm/Constants.h" 25 #include "llvm/DerivedTypes.h" 26 #include "llvm/Function.h" 27 #include "llvm/GlobalVariable.h" 28 #include "llvm/Instructions.h" 29 #include "llvm/IntrinsicInst.h" 30 #include "llvm/LLVMContext.h" 31 #include "llvm/Module.h" 32 #include "llvm/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" 52 using namespace llvm; 53 54 STATISTIC(NumReplaced, "Number of allocas broken up"); 55 STATISTIC(NumPromoted, "Number of allocas promoted"); 56 STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); 57 STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 58 STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 59 60 namespace { 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 } 195 196 char SROA_DT::ID = 0; 197 char SROA_SSAUp::ID = 0; 198 199 INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 200 "Scalar Replacement of Aggregates (DT)", false, false) 201 INITIALIZE_PASS_DEPENDENCY(DominatorTree) 202 INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 203 "Scalar Replacement of Aggregates (DT)", false, false) 204 205 INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 206 "Scalar Replacement of Aggregates (SSAUp)", false, false) 207 INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 208 "Scalar Replacement of Aggregates (SSAUp)", false, false) 209 210 // Public interface to the ScalarReplAggregates pass 211 FunctionPass *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 //===----------------------------------------------------------------------===// 222 223 namespace { 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. 227 class 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; 265 266 public: 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); 272 273 private: 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. 290 AllocaInst *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. 344 void 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. 388 bool 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. 414 bool 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. 525 void 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 int64_t SNumBytes = cast<ConstantInt>(MSI->getLength())->getSExtValue(); 578 if (SNumBytes > 0 && (SNumBytes >> 32) == 0) { 579 unsigned NumBytes = static_cast<unsigned>(SNumBytes); 580 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 581 582 // Compute the value replicated the right number of times. 583 APInt APVal(NumBytes*8, Val); 584 585 // Splat the value if non-zero. 586 if (Val) 587 for (unsigned i = 1; i != NumBytes; ++i) 588 APVal |= APVal << 8; 589 590 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 591 Value *New = ConvertScalar_InsertValue( 592 ConstantInt::get(User->getContext(), APVal), 593 Old, Offset, Builder); 594 Builder.CreateStore(New, NewAI); 595 596 // If the load we just inserted is now dead, then the memset overwrote 597 // the entire thing. 598 if (Old->use_empty()) 599 Old->eraseFromParent(); 600 } 601 MSI->eraseFromParent(); 602 continue; 603 } 604 605 // If this is a memcpy or memmove into or out of the whole allocation, we 606 // can handle it like a load or store of the scalar type. 607 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 608 assert(Offset == 0 && "must be store to start of alloca"); 609 610 // If the source and destination are both to the same alloca, then this is 611 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 612 // as appropriate. 613 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 614 615 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 616 // Dest must be OrigAI, change this to be a load from the original 617 // pointer (bitcasted), then a store to our new alloca. 618 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 619 Value *SrcPtr = MTI->getSource(); 620 PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 621 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 622 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 623 AIPTy = PointerType::get(AIPTy->getElementType(), 624 SPTy->getAddressSpace()); 625 } 626 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 627 628 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 629 SrcVal->setAlignment(MTI->getAlignment()); 630 Builder.CreateStore(SrcVal, NewAI); 631 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 632 // Src must be OrigAI, change this to be a load from NewAI then a store 633 // through the original dest pointer (bitcasted). 634 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 635 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 636 637 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 638 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 639 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 640 AIPTy = PointerType::get(AIPTy->getElementType(), 641 DPTy->getAddressSpace()); 642 } 643 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 644 645 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 646 NewStore->setAlignment(MTI->getAlignment()); 647 } else { 648 // Noop transfer. Src == Dst 649 } 650 651 MTI->eraseFromParent(); 652 continue; 653 } 654 655 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 656 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 657 II->getIntrinsicID() == Intrinsic::lifetime_end) { 658 // There's no need to preserve these, as the resulting alloca will be 659 // converted to a register anyways. 660 II->eraseFromParent(); 661 continue; 662 } 663 } 664 665 llvm_unreachable("Unsupported operation!"); 666 } 667 } 668 669 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 670 /// or vector value FromVal, extracting the bits from the offset specified by 671 /// Offset. This returns the value, which is of type ToType. 672 /// 673 /// This happens when we are converting an "integer union" to a single 674 /// integer scalar, or when we are converting a "vector union" to a vector with 675 /// insert/extractelement instructions. 676 /// 677 /// Offset is an offset from the original alloca, in bits that need to be 678 /// shifted to the right. 679 Value *ConvertToScalarInfo:: 680 ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, 681 uint64_t Offset, IRBuilder<> &Builder) { 682 // If the load is of the whole new alloca, no conversion is needed. 683 Type *FromType = FromVal->getType(); 684 if (FromType == ToType && Offset == 0) 685 return FromVal; 686 687 // If the result alloca is a vector type, this is either an element 688 // access or a bitcast to another vector type of the same size. 689 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) { 690 unsigned FromTypeSize = TD.getTypeAllocSize(FromType); 691 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 692 if (FromTypeSize == ToTypeSize) 693 return Builder.CreateBitCast(FromVal, ToType); 694 695 // Otherwise it must be an element access. 696 unsigned Elt = 0; 697 if (Offset) { 698 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 699 Elt = Offset/EltSize; 700 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 701 } 702 // Return the element extracted out of it. 703 Value *V = Builder.CreateExtractElement(FromVal, Builder.getInt32(Elt)); 704 if (V->getType() != ToType) 705 V = Builder.CreateBitCast(V, ToType); 706 return V; 707 } 708 709 // If ToType is a first class aggregate, extract out each of the pieces and 710 // use insertvalue's to form the FCA. 711 if (StructType *ST = dyn_cast<StructType>(ToType)) { 712 const StructLayout &Layout = *TD.getStructLayout(ST); 713 Value *Res = UndefValue::get(ST); 714 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 715 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 716 Offset+Layout.getElementOffsetInBits(i), 717 Builder); 718 Res = Builder.CreateInsertValue(Res, Elt, i); 719 } 720 return Res; 721 } 722 723 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 724 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 725 Value *Res = UndefValue::get(AT); 726 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 727 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 728 Offset+i*EltSize, Builder); 729 Res = Builder.CreateInsertValue(Res, Elt, i); 730 } 731 return Res; 732 } 733 734 // Otherwise, this must be a union that was converted to an integer value. 735 IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 736 737 // If this is a big-endian system and the load is narrower than the 738 // full alloca type, we need to do a shift to get the right bits. 739 int ShAmt = 0; 740 if (TD.isBigEndian()) { 741 // On big-endian machines, the lowest bit is stored at the bit offset 742 // from the pointer given by getTypeStoreSizeInBits. This matters for 743 // integers with a bitwidth that is not a multiple of 8. 744 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 745 TD.getTypeStoreSizeInBits(ToType) - Offset; 746 } else { 747 ShAmt = Offset; 748 } 749 750 // Note: we support negative bitwidths (with shl) which are not defined. 751 // We do this to support (f.e.) loads off the end of a structure where 752 // only some bits are used. 753 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 754 FromVal = Builder.CreateLShr(FromVal, 755 ConstantInt::get(FromVal->getType(), ShAmt)); 756 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 757 FromVal = Builder.CreateShl(FromVal, 758 ConstantInt::get(FromVal->getType(), -ShAmt)); 759 760 // Finally, unconditionally truncate the integer to the right width. 761 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 762 if (LIBitWidth < NTy->getBitWidth()) 763 FromVal = 764 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 765 LIBitWidth)); 766 else if (LIBitWidth > NTy->getBitWidth()) 767 FromVal = 768 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 769 LIBitWidth)); 770 771 // If the result is an integer, this is a trunc or bitcast. 772 if (ToType->isIntegerTy()) { 773 // Should be done. 774 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 775 // Just do a bitcast, we know the sizes match up. 776 FromVal = Builder.CreateBitCast(FromVal, ToType); 777 } else { 778 // Otherwise must be a pointer. 779 FromVal = Builder.CreateIntToPtr(FromVal, ToType); 780 } 781 assert(FromVal->getType() == ToType && "Didn't convert right?"); 782 return FromVal; 783 } 784 785 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 786 /// or vector value "Old" at the offset specified by Offset. 787 /// 788 /// This happens when we are converting an "integer union" to a 789 /// single integer scalar, or when we are converting a "vector union" to a 790 /// vector with insert/extractelement instructions. 791 /// 792 /// Offset is an offset from the original alloca, in bits that need to be 793 /// shifted to the right. 794 Value *ConvertToScalarInfo:: 795 ConvertScalar_InsertValue(Value *SV, Value *Old, 796 uint64_t Offset, IRBuilder<> &Builder) { 797 // Convert the stored type to the actual type, shift it left to insert 798 // then 'or' into place. 799 Type *AllocaType = Old->getType(); 800 LLVMContext &Context = Old->getContext(); 801 802 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 803 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 804 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 805 806 // Changing the whole vector with memset or with an access of a different 807 // vector type? 808 if (ValSize == VecSize) 809 return Builder.CreateBitCast(SV, AllocaType); 810 811 // Must be an element insertion. 812 Type *EltTy = VTy->getElementType(); 813 if (SV->getType() != EltTy) 814 SV = Builder.CreateBitCast(SV, EltTy); 815 uint64_t EltSize = TD.getTypeAllocSizeInBits(EltTy); 816 unsigned Elt = Offset/EltSize; 817 return Builder.CreateInsertElement(Old, SV, Builder.getInt32(Elt)); 818 } 819 820 // If SV is a first-class aggregate value, insert each value recursively. 821 if (StructType *ST = dyn_cast<StructType>(SV->getType())) { 822 const StructLayout &Layout = *TD.getStructLayout(ST); 823 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 824 Value *Elt = Builder.CreateExtractValue(SV, i); 825 Old = ConvertScalar_InsertValue(Elt, Old, 826 Offset+Layout.getElementOffsetInBits(i), 827 Builder); 828 } 829 return Old; 830 } 831 832 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 833 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 834 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 835 Value *Elt = Builder.CreateExtractValue(SV, i); 836 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 837 } 838 return Old; 839 } 840 841 // If SV is a float, convert it to the appropriate integer type. 842 // If it is a pointer, do the same. 843 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 844 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 845 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 846 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 847 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 848 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth)); 849 else if (SV->getType()->isPointerTy()) 850 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext())); 851 852 // Zero extend or truncate the value if needed. 853 if (SV->getType() != AllocaType) { 854 if (SV->getType()->getPrimitiveSizeInBits() < 855 AllocaType->getPrimitiveSizeInBits()) 856 SV = Builder.CreateZExt(SV, AllocaType); 857 else { 858 // Truncation may be needed if storing more than the alloca can hold 859 // (undefined behavior). 860 SV = Builder.CreateTrunc(SV, AllocaType); 861 SrcWidth = DestWidth; 862 SrcStoreWidth = DestStoreWidth; 863 } 864 } 865 866 // If this is a big-endian system and the store is narrower than the 867 // full alloca type, we need to do a shift to get the right bits. 868 int ShAmt = 0; 869 if (TD.isBigEndian()) { 870 // On big-endian machines, the lowest bit is stored at the bit offset 871 // from the pointer given by getTypeStoreSizeInBits. This matters for 872 // integers with a bitwidth that is not a multiple of 8. 873 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 874 } else { 875 ShAmt = Offset; 876 } 877 878 // Note: we support negative bitwidths (with shr) which are not defined. 879 // We do this to support (f.e.) stores off the end of a structure where 880 // only some bits in the structure are set. 881 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 882 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 883 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt)); 884 Mask <<= ShAmt; 885 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 886 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt)); 887 Mask = Mask.lshr(-ShAmt); 888 } 889 890 // Mask out the bits we are about to insert from the old value, and or 891 // in the new bits. 892 if (SrcWidth != DestWidth) { 893 assert(DestWidth > SrcWidth); 894 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 895 SV = Builder.CreateOr(Old, SV, "ins"); 896 } 897 return SV; 898 } 899 900 901 //===----------------------------------------------------------------------===// 902 // SRoA Driver 903 //===----------------------------------------------------------------------===// 904 905 906 bool SROA::runOnFunction(Function &F) { 907 TD = getAnalysisIfAvailable<TargetData>(); 908 909 bool Changed = performPromotion(F); 910 911 // FIXME: ScalarRepl currently depends on TargetData more than it 912 // theoretically needs to. It should be refactored in order to support 913 // target-independent IR. Until this is done, just skip the actual 914 // scalar-replacement portion of this pass. 915 if (!TD) return Changed; 916 917 while (1) { 918 bool LocalChange = performScalarRepl(F); 919 if (!LocalChange) break; // No need to repromote if no scalarrepl 920 Changed = true; 921 LocalChange = performPromotion(F); 922 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 923 } 924 925 return Changed; 926 } 927 928 namespace { 929 class AllocaPromoter : public LoadAndStorePromoter { 930 AllocaInst *AI; 931 DIBuilder *DIB; 932 SmallVector<DbgDeclareInst *, 4> DDIs; 933 SmallVector<DbgValueInst *, 4> DVIs; 934 public: 935 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 936 DIBuilder *DB) 937 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {} 938 939 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 940 // Remember which alloca we're promoting (for isInstInList). 941 this->AI = AI; 942 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) { 943 for (Value::use_iterator UI = DebugNode->use_begin(), 944 E = DebugNode->use_end(); UI != E; ++UI) 945 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI)) 946 DDIs.push_back(DDI); 947 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI)) 948 DVIs.push_back(DVI); 949 } 950 951 LoadAndStorePromoter::run(Insts); 952 AI->eraseFromParent(); 953 for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(), 954 E = DDIs.end(); I != E; ++I) { 955 DbgDeclareInst *DDI = *I; 956 DDI->eraseFromParent(); 957 } 958 for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(), 959 E = DVIs.end(); I != E; ++I) { 960 DbgValueInst *DVI = *I; 961 DVI->eraseFromParent(); 962 } 963 } 964 965 virtual bool isInstInList(Instruction *I, 966 const SmallVectorImpl<Instruction*> &Insts) const { 967 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 968 return LI->getOperand(0) == AI; 969 return cast<StoreInst>(I)->getPointerOperand() == AI; 970 } 971 972 virtual void updateDebugInfo(Instruction *Inst) const { 973 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(), 974 E = DDIs.end(); I != E; ++I) { 975 DbgDeclareInst *DDI = *I; 976 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 977 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB); 978 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 979 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB); 980 } 981 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(), 982 E = DVIs.end(); I != E; ++I) { 983 DbgValueInst *DVI = *I; 984 Value *Arg = NULL; 985 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 986 // If an argument is zero extended then use argument directly. The ZExt 987 // may be zapped by an optimization pass in future. 988 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 989 Arg = dyn_cast<Argument>(ZExt->getOperand(0)); 990 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 991 Arg = dyn_cast<Argument>(SExt->getOperand(0)); 992 if (!Arg) 993 Arg = SI->getOperand(0); 994 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 995 Arg = LI->getOperand(0); 996 } else { 997 continue; 998 } 999 Instruction *DbgVal = 1000 DIB->insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()), 1001 Inst); 1002 DbgVal->setDebugLoc(DVI->getDebugLoc()); 1003 } 1004 } 1005 }; 1006 } // end anon namespace 1007 1008 /// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1009 /// subsequently loaded can be rewritten to load both input pointers and then 1010 /// select between the result, allowing the load of the alloca to be promoted. 1011 /// From this: 1012 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1013 /// %V = load i32* %P2 1014 /// to: 1015 /// %V1 = load i32* %Alloca -> will be mem2reg'd 1016 /// %V2 = load i32* %Other 1017 /// %V = select i1 %cond, i32 %V1, i32 %V2 1018 /// 1019 /// We can do this to a select if its only uses are loads and if the operand to 1020 /// the select can be loaded unconditionally. 1021 static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1022 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1023 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1024 1025 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1026 UI != UE; ++UI) { 1027 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1028 if (LI == 0 || !LI->isSimple()) return false; 1029 1030 // Both operands to the select need to be dereferencable, either absolutely 1031 // (e.g. allocas) or at this point because we can see other accesses to it. 1032 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1033 LI->getAlignment(), TD)) 1034 return false; 1035 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1036 LI->getAlignment(), TD)) 1037 return false; 1038 } 1039 1040 return true; 1041 } 1042 1043 /// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1044 /// subsequently loaded can be rewritten to load both input pointers in the pred 1045 /// blocks and then PHI the results, allowing the load of the alloca to be 1046 /// promoted. 1047 /// From this: 1048 /// %P2 = phi [i32* %Alloca, i32* %Other] 1049 /// %V = load i32* %P2 1050 /// to: 1051 /// %V1 = load i32* %Alloca -> will be mem2reg'd 1052 /// ... 1053 /// %V2 = load i32* %Other 1054 /// ... 1055 /// %V = phi [i32 %V1, i32 %V2] 1056 /// 1057 /// We can do this to a select if its only uses are loads and if the operand to 1058 /// the select can be loaded unconditionally. 1059 static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1060 // For now, we can only do this promotion if the load is in the same block as 1061 // the PHI, and if there are no stores between the phi and load. 1062 // TODO: Allow recursive phi users. 1063 // TODO: Allow stores. 1064 BasicBlock *BB = PN->getParent(); 1065 unsigned MaxAlign = 0; 1066 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1067 UI != UE; ++UI) { 1068 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1069 if (LI == 0 || !LI->isSimple()) return false; 1070 1071 // For now we only allow loads in the same block as the PHI. This is a 1072 // common case that happens when instcombine merges two loads through a PHI. 1073 if (LI->getParent() != BB) return false; 1074 1075 // Ensure that there are no instructions between the PHI and the load that 1076 // could store. 1077 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1078 if (BBI->mayWriteToMemory()) 1079 return false; 1080 1081 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1082 } 1083 1084 // Okay, we know that we have one or more loads in the same block as the PHI. 1085 // We can transform this if it is safe to push the loads into the predecessor 1086 // blocks. The only thing to watch out for is that we can't put a possibly 1087 // trapping load in the predecessor if it is a critical edge. 1088 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1089 BasicBlock *Pred = PN->getIncomingBlock(i); 1090 Value *InVal = PN->getIncomingValue(i); 1091 1092 // If the terminator of the predecessor has side-effects (an invoke), 1093 // there is no safe place to put a load in the predecessor. 1094 if (Pred->getTerminator()->mayHaveSideEffects()) 1095 return false; 1096 1097 // If the value is produced by the terminator of the predecessor 1098 // (an invoke), there is no valid place to put a load in the predecessor. 1099 if (Pred->getTerminator() == InVal) 1100 return false; 1101 1102 // If the predecessor has a single successor, then the edge isn't critical. 1103 if (Pred->getTerminator()->getNumSuccessors() == 1) 1104 continue; 1105 1106 // If this pointer is always safe to load, or if we can prove that there is 1107 // already a load in the block, then we can move the load to the pred block. 1108 if (InVal->isDereferenceablePointer() || 1109 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1110 continue; 1111 1112 return false; 1113 } 1114 1115 return true; 1116 } 1117 1118 1119 /// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1120 /// direct (non-volatile) loads and stores to it. If the alloca is close but 1121 /// not quite there, this will transform the code to allow promotion. As such, 1122 /// it is a non-pure predicate. 1123 static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1124 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1125 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1126 1127 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1128 UI != UE; ++UI) { 1129 User *U = *UI; 1130 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1131 if (!LI->isSimple()) 1132 return false; 1133 continue; 1134 } 1135 1136 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1137 if (SI->getOperand(0) == AI || !SI->isSimple()) 1138 return false; // Don't allow a store OF the AI, only INTO the AI. 1139 continue; 1140 } 1141 1142 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1143 // If the condition being selected on is a constant, fold the select, yes 1144 // this does (rarely) happen early on. 1145 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1146 Value *Result = SI->getOperand(1+CI->isZero()); 1147 SI->replaceAllUsesWith(Result); 1148 SI->eraseFromParent(); 1149 1150 // This is very rare and we just scrambled the use list of AI, start 1151 // over completely. 1152 return tryToMakeAllocaBePromotable(AI, TD); 1153 } 1154 1155 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1156 // loads, then we can transform this by rewriting the select. 1157 if (!isSafeSelectToSpeculate(SI, TD)) 1158 return false; 1159 1160 InstsToRewrite.insert(SI); 1161 continue; 1162 } 1163 1164 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1165 if (PN->use_empty()) { // Dead PHIs can be stripped. 1166 InstsToRewrite.insert(PN); 1167 continue; 1168 } 1169 1170 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1171 // in the pred blocks, then we can transform this by rewriting the PHI. 1172 if (!isSafePHIToSpeculate(PN, TD)) 1173 return false; 1174 1175 InstsToRewrite.insert(PN); 1176 continue; 1177 } 1178 1179 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1180 if (onlyUsedByLifetimeMarkers(BCI)) { 1181 InstsToRewrite.insert(BCI); 1182 continue; 1183 } 1184 } 1185 1186 return false; 1187 } 1188 1189 // If there are no instructions to rewrite, then all uses are load/stores and 1190 // we're done! 1191 if (InstsToRewrite.empty()) 1192 return true; 1193 1194 // If we have instructions that need to be rewritten for this to be promotable 1195 // take care of it now. 1196 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1197 if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) { 1198 // This could only be a bitcast used by nothing but lifetime intrinsics. 1199 for (BitCastInst::use_iterator I = BCI->use_begin(), E = BCI->use_end(); 1200 I != E;) { 1201 Use &U = I.getUse(); 1202 ++I; 1203 cast<Instruction>(U.getUser())->eraseFromParent(); 1204 } 1205 BCI->eraseFromParent(); 1206 continue; 1207 } 1208 1209 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1210 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1211 // loads with a new select. 1212 while (!SI->use_empty()) { 1213 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1214 1215 IRBuilder<> Builder(LI); 1216 LoadInst *TrueLoad = 1217 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1218 LoadInst *FalseLoad = 1219 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f"); 1220 1221 // Transfer alignment and TBAA info if present. 1222 TrueLoad->setAlignment(LI->getAlignment()); 1223 FalseLoad->setAlignment(LI->getAlignment()); 1224 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1225 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1226 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1227 } 1228 1229 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1230 V->takeName(LI); 1231 LI->replaceAllUsesWith(V); 1232 LI->eraseFromParent(); 1233 } 1234 1235 // Now that all the loads are gone, the select is gone too. 1236 SI->eraseFromParent(); 1237 continue; 1238 } 1239 1240 // Otherwise, we have a PHI node which allows us to push the loads into the 1241 // predecessors. 1242 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1243 if (PN->use_empty()) { 1244 PN->eraseFromParent(); 1245 continue; 1246 } 1247 1248 Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1249 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1250 PN->getName()+".ld", PN); 1251 1252 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1253 // matter which one we get and if any differ, it doesn't matter. 1254 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1255 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1256 unsigned Align = SomeLoad->getAlignment(); 1257 1258 // Rewrite all loads of the PN to use the new PHI. 1259 while (!PN->use_empty()) { 1260 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1261 LI->replaceAllUsesWith(NewPN); 1262 LI->eraseFromParent(); 1263 } 1264 1265 // Inject loads into all of the pred blocks. Keep track of which blocks we 1266 // insert them into in case we have multiple edges from the same block. 1267 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1268 1269 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1270 BasicBlock *Pred = PN->getIncomingBlock(i); 1271 LoadInst *&Load = InsertedLoads[Pred]; 1272 if (Load == 0) { 1273 Load = new LoadInst(PN->getIncomingValue(i), 1274 PN->getName() + "." + Pred->getName(), 1275 Pred->getTerminator()); 1276 Load->setAlignment(Align); 1277 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1278 } 1279 1280 NewPN->addIncoming(Load, Pred); 1281 } 1282 1283 PN->eraseFromParent(); 1284 } 1285 1286 ++NumAdjusted; 1287 return true; 1288 } 1289 1290 bool SROA::performPromotion(Function &F) { 1291 std::vector<AllocaInst*> Allocas; 1292 DominatorTree *DT = 0; 1293 if (HasDomTree) 1294 DT = &getAnalysis<DominatorTree>(); 1295 1296 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1297 DIBuilder DIB(*F.getParent()); 1298 bool Changed = false; 1299 SmallVector<Instruction*, 64> Insts; 1300 while (1) { 1301 Allocas.clear(); 1302 1303 // Find allocas that are safe to promote, by looking at all instructions in 1304 // the entry node 1305 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1306 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1307 if (tryToMakeAllocaBePromotable(AI, TD)) 1308 Allocas.push_back(AI); 1309 1310 if (Allocas.empty()) break; 1311 1312 if (HasDomTree) 1313 PromoteMemToReg(Allocas, *DT); 1314 else { 1315 SSAUpdater SSA; 1316 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1317 AllocaInst *AI = Allocas[i]; 1318 1319 // Build list of instructions to promote. 1320 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1321 UI != E; ++UI) 1322 Insts.push_back(cast<Instruction>(*UI)); 1323 AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts); 1324 Insts.clear(); 1325 } 1326 } 1327 NumPromoted += Allocas.size(); 1328 Changed = true; 1329 } 1330 1331 return Changed; 1332 } 1333 1334 1335 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1336 /// SROA. It must be a struct or array type with a small number of elements. 1337 static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1338 Type *T = AI->getAllocatedType(); 1339 // Do not promote any struct into more than 32 separate vars. 1340 if (StructType *ST = dyn_cast<StructType>(T)) 1341 return ST->getNumElements() <= 32; 1342 // Arrays are much less likely to be safe for SROA; only consider 1343 // them if they are very small. 1344 if (ArrayType *AT = dyn_cast<ArrayType>(T)) 1345 return AT->getNumElements() <= 8; 1346 return false; 1347 } 1348 1349 1350 // performScalarRepl - This algorithm is a simple worklist driven algorithm, 1351 // which runs on all of the alloca instructions in the function, removing them 1352 // if they are only used by getelementptr instructions. 1353 // 1354 bool SROA::performScalarRepl(Function &F) { 1355 std::vector<AllocaInst*> WorkList; 1356 1357 // Scan the entry basic block, adding allocas to the worklist. 1358 BasicBlock &BB = F.getEntryBlock(); 1359 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1360 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1361 WorkList.push_back(A); 1362 1363 // Process the worklist 1364 bool Changed = false; 1365 while (!WorkList.empty()) { 1366 AllocaInst *AI = WorkList.back(); 1367 WorkList.pop_back(); 1368 1369 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1370 // with unused elements. 1371 if (AI->use_empty()) { 1372 AI->eraseFromParent(); 1373 Changed = true; 1374 continue; 1375 } 1376 1377 // If this alloca is impossible for us to promote, reject it early. 1378 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1379 continue; 1380 1381 // Check to see if this allocation is only modified by a memcpy/memmove from 1382 // a constant global. If this is the case, we can change all users to use 1383 // the constant global instead. This is commonly produced by the CFE by 1384 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1385 // is only subsequently read. 1386 SmallVector<Instruction *, 4> ToDelete; 1387 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) { 1388 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1389 DEBUG(dbgs() << " memcpy = " << *Copy << '\n'); 1390 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i) 1391 ToDelete[i]->eraseFromParent(); 1392 Constant *TheSrc = cast<Constant>(Copy->getSource()); 1393 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1394 Copy->eraseFromParent(); // Don't mutate the global. 1395 AI->eraseFromParent(); 1396 ++NumGlobals; 1397 Changed = true; 1398 continue; 1399 } 1400 1401 // Check to see if we can perform the core SROA transformation. We cannot 1402 // transform the allocation instruction if it is an array allocation 1403 // (allocations OF arrays are ok though), and an allocation of a scalar 1404 // value cannot be decomposed at all. 1405 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1406 1407 // Do not promote [0 x %struct]. 1408 if (AllocaSize == 0) continue; 1409 1410 // Do not promote any struct whose size is too big. 1411 if (AllocaSize > SRThreshold) continue; 1412 1413 // If the alloca looks like a good candidate for scalar replacement, and if 1414 // all its users can be transformed, then split up the aggregate into its 1415 // separate elements. 1416 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1417 DoScalarReplacement(AI, WorkList); 1418 Changed = true; 1419 continue; 1420 } 1421 1422 // If we can turn this aggregate value (potentially with casts) into a 1423 // simple scalar value that can be mem2reg'd into a register value. 1424 // IsNotTrivial tracks whether this is something that mem2reg could have 1425 // promoted itself. If so, we don't want to transform it needlessly. Note 1426 // that we can't just check based on the type: the alloca may be of an i32 1427 // but that has pointer arithmetic to set byte 3 of it or something. 1428 if (AllocaInst *NewAI = 1429 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1430 NewAI->takeName(AI); 1431 AI->eraseFromParent(); 1432 ++NumConverted; 1433 Changed = true; 1434 continue; 1435 } 1436 1437 // Otherwise, couldn't process this alloca. 1438 } 1439 1440 return Changed; 1441 } 1442 1443 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1444 /// predicate, do SROA now. 1445 void SROA::DoScalarReplacement(AllocaInst *AI, 1446 std::vector<AllocaInst*> &WorkList) { 1447 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1448 SmallVector<AllocaInst*, 32> ElementAllocas; 1449 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1450 ElementAllocas.reserve(ST->getNumContainedTypes()); 1451 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1452 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1453 AI->getAlignment(), 1454 AI->getName() + "." + Twine(i), AI); 1455 ElementAllocas.push_back(NA); 1456 WorkList.push_back(NA); // Add to worklist for recursive processing 1457 } 1458 } else { 1459 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1460 ElementAllocas.reserve(AT->getNumElements()); 1461 Type *ElTy = AT->getElementType(); 1462 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1463 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1464 AI->getName() + "." + Twine(i), AI); 1465 ElementAllocas.push_back(NA); 1466 WorkList.push_back(NA); // Add to worklist for recursive processing 1467 } 1468 } 1469 1470 // Now that we have created the new alloca instructions, rewrite all the 1471 // uses of the old alloca. 1472 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1473 1474 // Now erase any instructions that were made dead while rewriting the alloca. 1475 DeleteDeadInstructions(); 1476 AI->eraseFromParent(); 1477 1478 ++NumReplaced; 1479 } 1480 1481 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1482 /// recursively including all their operands that become trivially dead. 1483 void SROA::DeleteDeadInstructions() { 1484 while (!DeadInsts.empty()) { 1485 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1486 1487 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1488 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1489 // Zero out the operand and see if it becomes trivially dead. 1490 // (But, don't add allocas to the dead instruction list -- they are 1491 // already on the worklist and will be deleted separately.) 1492 *OI = 0; 1493 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1494 DeadInsts.push_back(U); 1495 } 1496 1497 I->eraseFromParent(); 1498 } 1499 } 1500 1501 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1502 /// performing scalar replacement of alloca AI. The results are flagged in 1503 /// the Info parameter. Offset indicates the position within AI that is 1504 /// referenced by this instruction. 1505 void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1506 AllocaInfo &Info) { 1507 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1508 Instruction *User = cast<Instruction>(*UI); 1509 1510 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1511 isSafeForScalarRepl(BC, Offset, Info); 1512 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1513 uint64_t GEPOffset = Offset; 1514 isSafeGEP(GEPI, GEPOffset, Info); 1515 if (!Info.isUnsafe) 1516 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1517 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1518 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1519 if (Length == 0) 1520 return MarkUnsafe(Info, User); 1521 if (Length->isNegative()) 1522 return MarkUnsafe(Info, User); 1523 1524 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1525 UI.getOperandNo() == 0, Info, MI, 1526 true /*AllowWholeAccess*/); 1527 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1528 if (!LI->isSimple()) 1529 return MarkUnsafe(Info, User); 1530 Type *LIType = LI->getType(); 1531 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1532 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1533 Info.hasALoadOrStore = true; 1534 1535 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1536 // Store is ok if storing INTO the pointer, not storing the pointer 1537 if (!SI->isSimple() || SI->getOperand(0) == I) 1538 return MarkUnsafe(Info, User); 1539 1540 Type *SIType = SI->getOperand(0)->getType(); 1541 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1542 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1543 Info.hasALoadOrStore = true; 1544 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1545 if (II->getIntrinsicID() != Intrinsic::lifetime_start && 1546 II->getIntrinsicID() != Intrinsic::lifetime_end) 1547 return MarkUnsafe(Info, User); 1548 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1549 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1550 } else { 1551 return MarkUnsafe(Info, User); 1552 } 1553 if (Info.isUnsafe) return; 1554 } 1555 } 1556 1557 1558 /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1559 /// derived from the alloca, we can often still split the alloca into elements. 1560 /// This is useful if we have a large alloca where one element is phi'd 1561 /// together somewhere: we can SRoA and promote all the other elements even if 1562 /// we end up not being able to promote this one. 1563 /// 1564 /// All we require is that the uses of the PHI do not index into other parts of 1565 /// the alloca. The most important use case for this is single load and stores 1566 /// that are PHI'd together, which can happen due to code sinking. 1567 void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1568 AllocaInfo &Info) { 1569 // If we've already checked this PHI, don't do it again. 1570 if (PHINode *PN = dyn_cast<PHINode>(I)) 1571 if (!Info.CheckedPHIs.insert(PN)) 1572 return; 1573 1574 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1575 Instruction *User = cast<Instruction>(*UI); 1576 1577 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1578 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1579 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1580 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1581 // but would have to prove that we're staying inside of an element being 1582 // promoted. 1583 if (!GEPI->hasAllZeroIndices()) 1584 return MarkUnsafe(Info, User); 1585 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1586 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1587 if (!LI->isSimple()) 1588 return MarkUnsafe(Info, User); 1589 Type *LIType = LI->getType(); 1590 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1591 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1592 Info.hasALoadOrStore = true; 1593 1594 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1595 // Store is ok if storing INTO the pointer, not storing the pointer 1596 if (!SI->isSimple() || SI->getOperand(0) == I) 1597 return MarkUnsafe(Info, User); 1598 1599 Type *SIType = SI->getOperand(0)->getType(); 1600 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1601 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1602 Info.hasALoadOrStore = true; 1603 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1604 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1605 } else { 1606 return MarkUnsafe(Info, User); 1607 } 1608 if (Info.isUnsafe) return; 1609 } 1610 } 1611 1612 /// isSafeGEP - Check if a GEP instruction can be handled for scalar 1613 /// replacement. It is safe when all the indices are constant, in-bounds 1614 /// references, and when the resulting offset corresponds to an element within 1615 /// the alloca type. The results are flagged in the Info parameter. Upon 1616 /// return, Offset is adjusted as specified by the GEP indices. 1617 void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1618 uint64_t &Offset, AllocaInfo &Info) { 1619 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1620 if (GEPIt == E) 1621 return; 1622 1623 // Walk through the GEP type indices, checking the types that this indexes 1624 // into. 1625 for (; GEPIt != E; ++GEPIt) { 1626 // Ignore struct elements, no extra checking needed for these. 1627 if ((*GEPIt)->isStructTy()) 1628 continue; 1629 1630 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1631 if (!IdxVal) 1632 return MarkUnsafe(Info, GEPI); 1633 } 1634 1635 // Compute the offset due to this GEP and check if the alloca has a 1636 // component element at that offset. 1637 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1638 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1639 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1640 MarkUnsafe(Info, GEPI); 1641 } 1642 1643 /// isHomogeneousAggregate - Check if type T is a struct or array containing 1644 /// elements of the same type (which is always true for arrays). If so, 1645 /// return true with NumElts and EltTy set to the number of elements and the 1646 /// element type, respectively. 1647 static bool isHomogeneousAggregate(Type *T, unsigned &NumElts, 1648 Type *&EltTy) { 1649 if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1650 NumElts = AT->getNumElements(); 1651 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1652 return true; 1653 } 1654 if (StructType *ST = dyn_cast<StructType>(T)) { 1655 NumElts = ST->getNumContainedTypes(); 1656 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1657 for (unsigned n = 1; n < NumElts; ++n) { 1658 if (ST->getContainedType(n) != EltTy) 1659 return false; 1660 } 1661 return true; 1662 } 1663 return false; 1664 } 1665 1666 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1667 /// "homogeneous" aggregates with the same element type and number of elements. 1668 static bool isCompatibleAggregate(Type *T1, Type *T2) { 1669 if (T1 == T2) 1670 return true; 1671 1672 unsigned NumElts1, NumElts2; 1673 Type *EltTy1, *EltTy2; 1674 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1675 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1676 NumElts1 == NumElts2 && 1677 EltTy1 == EltTy2) 1678 return true; 1679 1680 return false; 1681 } 1682 1683 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1684 /// alloca or has an offset and size that corresponds to a component element 1685 /// within it. The offset checked here may have been formed from a GEP with a 1686 /// pointer bitcasted to a different type. 1687 /// 1688 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1689 /// unit. If false, it only allows accesses known to be in a single element. 1690 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1691 Type *MemOpType, bool isStore, 1692 AllocaInfo &Info, Instruction *TheAccess, 1693 bool AllowWholeAccess) { 1694 // Check if this is a load/store of the entire alloca. 1695 if (Offset == 0 && AllowWholeAccess && 1696 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1697 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1698 // loads/stores (which are essentially the same as the MemIntrinsics with 1699 // regard to copying padding between elements). But, if an alloca is 1700 // flagged as both a source and destination of such operations, we'll need 1701 // to check later for padding between elements. 1702 if (!MemOpType || MemOpType->isIntegerTy()) { 1703 if (isStore) 1704 Info.isMemCpyDst = true; 1705 else 1706 Info.isMemCpySrc = true; 1707 return; 1708 } 1709 // This is also safe for references using a type that is compatible with 1710 // the type of the alloca, so that loads/stores can be rewritten using 1711 // insertvalue/extractvalue. 1712 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1713 Info.hasSubelementAccess = true; 1714 return; 1715 } 1716 } 1717 // Check if the offset/size correspond to a component within the alloca type. 1718 Type *T = Info.AI->getAllocatedType(); 1719 if (TypeHasComponent(T, Offset, MemSize)) { 1720 Info.hasSubelementAccess = true; 1721 return; 1722 } 1723 1724 return MarkUnsafe(Info, TheAccess); 1725 } 1726 1727 /// TypeHasComponent - Return true if T has a component type with the 1728 /// specified offset and size. If Size is zero, do not check the size. 1729 bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { 1730 Type *EltTy; 1731 uint64_t EltSize; 1732 if (StructType *ST = dyn_cast<StructType>(T)) { 1733 const StructLayout *Layout = TD->getStructLayout(ST); 1734 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1735 EltTy = ST->getContainedType(EltIdx); 1736 EltSize = TD->getTypeAllocSize(EltTy); 1737 Offset -= Layout->getElementOffset(EltIdx); 1738 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1739 EltTy = AT->getElementType(); 1740 EltSize = TD->getTypeAllocSize(EltTy); 1741 if (Offset >= AT->getNumElements() * EltSize) 1742 return false; 1743 Offset %= EltSize; 1744 } else { 1745 return false; 1746 } 1747 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1748 return true; 1749 // Check if the component spans multiple elements. 1750 if (Offset + Size > EltSize) 1751 return false; 1752 return TypeHasComponent(EltTy, Offset, Size); 1753 } 1754 1755 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1756 /// the instruction I, which references it, to use the separate elements. 1757 /// Offset indicates the position within AI that is referenced by this 1758 /// instruction. 1759 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1760 SmallVector<AllocaInst*, 32> &NewElts) { 1761 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1762 Use &TheUse = UI.getUse(); 1763 Instruction *User = cast<Instruction>(*UI++); 1764 1765 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1766 RewriteBitCast(BC, AI, Offset, NewElts); 1767 continue; 1768 } 1769 1770 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1771 RewriteGEP(GEPI, AI, Offset, NewElts); 1772 continue; 1773 } 1774 1775 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1776 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1777 uint64_t MemSize = Length->getZExtValue(); 1778 if (Offset == 0 && 1779 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1780 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1781 // Otherwise the intrinsic can only touch a single element and the 1782 // address operand will be updated, so nothing else needs to be done. 1783 continue; 1784 } 1785 1786 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1787 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 1788 II->getIntrinsicID() == Intrinsic::lifetime_end) { 1789 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts); 1790 } 1791 continue; 1792 } 1793 1794 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1795 Type *LIType = LI->getType(); 1796 1797 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1798 // Replace: 1799 // %res = load { i32, i32 }* %alloc 1800 // with: 1801 // %load.0 = load i32* %alloc.0 1802 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1803 // %load.1 = load i32* %alloc.1 1804 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1805 // (Also works for arrays instead of structs) 1806 Value *Insert = UndefValue::get(LIType); 1807 IRBuilder<> Builder(LI); 1808 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1809 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1810 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1811 } 1812 LI->replaceAllUsesWith(Insert); 1813 DeadInsts.push_back(LI); 1814 } else if (LIType->isIntegerTy() && 1815 TD->getTypeAllocSize(LIType) == 1816 TD->getTypeAllocSize(AI->getAllocatedType())) { 1817 // If this is a load of the entire alloca to an integer, rewrite it. 1818 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1819 } 1820 continue; 1821 } 1822 1823 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1824 Value *Val = SI->getOperand(0); 1825 Type *SIType = Val->getType(); 1826 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1827 // Replace: 1828 // store { i32, i32 } %val, { i32, i32 }* %alloc 1829 // with: 1830 // %val.0 = extractvalue { i32, i32 } %val, 0 1831 // store i32 %val.0, i32* %alloc.0 1832 // %val.1 = extractvalue { i32, i32 } %val, 1 1833 // store i32 %val.1, i32* %alloc.1 1834 // (Also works for arrays instead of structs) 1835 IRBuilder<> Builder(SI); 1836 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1837 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1838 Builder.CreateStore(Extract, NewElts[i]); 1839 } 1840 DeadInsts.push_back(SI); 1841 } else if (SIType->isIntegerTy() && 1842 TD->getTypeAllocSize(SIType) == 1843 TD->getTypeAllocSize(AI->getAllocatedType())) { 1844 // If this is a store of the entire alloca from an integer, rewrite it. 1845 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1846 } 1847 continue; 1848 } 1849 1850 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1851 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1852 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1853 // the new pointer. 1854 if (!isa<AllocaInst>(I)) continue; 1855 1856 assert(Offset == 0 && NewElts[0] && 1857 "Direct alloca use should have a zero offset"); 1858 1859 // If we have a use of the alloca, we know the derived uses will be 1860 // utilizing just the first element of the scalarized result. Insert a 1861 // bitcast of the first alloca before the user as required. 1862 AllocaInst *NewAI = NewElts[0]; 1863 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1864 NewAI->moveBefore(BCI); 1865 TheUse = BCI; 1866 continue; 1867 } 1868 } 1869 } 1870 1871 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1872 /// and recursively continue updating all of its uses. 1873 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1874 SmallVector<AllocaInst*, 32> &NewElts) { 1875 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1876 if (BC->getOperand(0) != AI) 1877 return; 1878 1879 // The bitcast references the original alloca. Replace its uses with 1880 // references to the alloca containing offset zero (which is normally at 1881 // index zero, but might not be in cases involving structs with elements 1882 // of size zero). 1883 Type *T = AI->getAllocatedType(); 1884 uint64_t EltOffset = 0; 1885 Type *IdxTy; 1886 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1887 Instruction *Val = NewElts[Idx]; 1888 if (Val->getType() != BC->getDestTy()) { 1889 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1890 Val->takeName(BC); 1891 } 1892 BC->replaceAllUsesWith(Val); 1893 DeadInsts.push_back(BC); 1894 } 1895 1896 /// FindElementAndOffset - Return the index of the element containing Offset 1897 /// within the specified type, which must be either a struct or an array. 1898 /// Sets T to the type of the element and Offset to the offset within that 1899 /// element. IdxTy is set to the type of the index result to be used in a 1900 /// GEP instruction. 1901 uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, 1902 Type *&IdxTy) { 1903 uint64_t Idx = 0; 1904 if (StructType *ST = dyn_cast<StructType>(T)) { 1905 const StructLayout *Layout = TD->getStructLayout(ST); 1906 Idx = Layout->getElementContainingOffset(Offset); 1907 T = ST->getContainedType(Idx); 1908 Offset -= Layout->getElementOffset(Idx); 1909 IdxTy = Type::getInt32Ty(T->getContext()); 1910 return Idx; 1911 } 1912 ArrayType *AT = cast<ArrayType>(T); 1913 T = AT->getElementType(); 1914 uint64_t EltSize = TD->getTypeAllocSize(T); 1915 Idx = Offset / EltSize; 1916 Offset -= Idx * EltSize; 1917 IdxTy = Type::getInt64Ty(T->getContext()); 1918 return Idx; 1919 } 1920 1921 /// RewriteGEP - Check if this GEP instruction moves the pointer across 1922 /// elements of the alloca that are being split apart, and if so, rewrite 1923 /// the GEP to be relative to the new element. 1924 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1925 SmallVector<AllocaInst*, 32> &NewElts) { 1926 uint64_t OldOffset = Offset; 1927 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1928 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1929 1930 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1931 1932 Type *T = AI->getAllocatedType(); 1933 Type *IdxTy; 1934 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1935 if (GEPI->getOperand(0) == AI) 1936 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1937 1938 T = AI->getAllocatedType(); 1939 uint64_t EltOffset = Offset; 1940 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1941 1942 // If this GEP does not move the pointer across elements of the alloca 1943 // being split, then it does not needs to be rewritten. 1944 if (Idx == OldIdx) 1945 return; 1946 1947 Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1948 SmallVector<Value*, 8> NewArgs; 1949 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1950 while (EltOffset != 0) { 1951 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1952 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1953 } 1954 Instruction *Val = NewElts[Idx]; 1955 if (NewArgs.size() > 1) { 1956 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI); 1957 Val->takeName(GEPI); 1958 } 1959 if (Val->getType() != GEPI->getType()) 1960 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1961 GEPI->replaceAllUsesWith(Val); 1962 DeadInsts.push_back(GEPI); 1963 } 1964 1965 /// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it 1966 /// to mark the lifetime of the scalarized memory. 1967 void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 1968 uint64_t Offset, 1969 SmallVector<AllocaInst*, 32> &NewElts) { 1970 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0)); 1971 // Put matching lifetime markers on everything from Offset up to 1972 // Offset+OldSize. 1973 Type *AIType = AI->getAllocatedType(); 1974 uint64_t NewOffset = Offset; 1975 Type *IdxTy; 1976 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy); 1977 1978 IRBuilder<> Builder(II); 1979 uint64_t Size = OldSize->getLimitedValue(); 1980 1981 if (NewOffset) { 1982 // Splice the first element and index 'NewOffset' bytes in. SROA will 1983 // split the alloca again later. 1984 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy()); 1985 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset)); 1986 1987 IdxTy = NewElts[Idx]->getAllocatedType(); 1988 uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset; 1989 if (EltSize > Size) { 1990 EltSize = Size; 1991 Size = 0; 1992 } else { 1993 Size -= EltSize; 1994 } 1995 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 1996 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize)); 1997 else 1998 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize)); 1999 ++Idx; 2000 } 2001 2002 for (; Idx != NewElts.size() && Size; ++Idx) { 2003 IdxTy = NewElts[Idx]->getAllocatedType(); 2004 uint64_t EltSize = TD->getTypeAllocSize(IdxTy); 2005 if (EltSize > Size) { 2006 EltSize = Size; 2007 Size = 0; 2008 } else { 2009 Size -= EltSize; 2010 } 2011 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2012 Builder.CreateLifetimeStart(NewElts[Idx], 2013 Builder.getInt64(EltSize)); 2014 else 2015 Builder.CreateLifetimeEnd(NewElts[Idx], 2016 Builder.getInt64(EltSize)); 2017 } 2018 DeadInsts.push_back(II); 2019 } 2020 2021 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2022 /// Rewrite it to copy or set the elements of the scalarized memory. 2023 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2024 AllocaInst *AI, 2025 SmallVector<AllocaInst*, 32> &NewElts) { 2026 // If this is a memcpy/memmove, construct the other pointer as the 2027 // appropriate type. The "Other" pointer is the pointer that goes to memory 2028 // that doesn't have anything to do with the alloca that we are promoting. For 2029 // memset, this Value* stays null. 2030 Value *OtherPtr = 0; 2031 unsigned MemAlignment = MI->getAlignment(); 2032 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2033 if (Inst == MTI->getRawDest()) 2034 OtherPtr = MTI->getRawSource(); 2035 else { 2036 assert(Inst == MTI->getRawSource()); 2037 OtherPtr = MTI->getRawDest(); 2038 } 2039 } 2040 2041 // If there is an other pointer, we want to convert it to the same pointer 2042 // type as AI has, so we can GEP through it safely. 2043 if (OtherPtr) { 2044 unsigned AddrSpace = 2045 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2046 2047 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2048 // optimization, but it's also required to detect the corner case where 2049 // both pointer operands are referencing the same memory, and where 2050 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2051 // function is only called for mem intrinsics that access the whole 2052 // aggregate, so non-zero GEPs are not an issue here.) 2053 OtherPtr = OtherPtr->stripPointerCasts(); 2054 2055 // Copying the alloca to itself is a no-op: just delete it. 2056 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2057 // This code will run twice for a no-op memcpy -- once for each operand. 2058 // Put only one reference to MI on the DeadInsts list. 2059 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2060 E = DeadInsts.end(); I != E; ++I) 2061 if (*I == MI) return; 2062 DeadInsts.push_back(MI); 2063 return; 2064 } 2065 2066 // If the pointer is not the right type, insert a bitcast to the right 2067 // type. 2068 Type *NewTy = 2069 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2070 2071 if (OtherPtr->getType() != NewTy) 2072 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2073 } 2074 2075 // Process each element of the aggregate. 2076 bool SROADest = MI->getRawDest() == Inst; 2077 2078 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2079 2080 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2081 // If this is a memcpy/memmove, emit a GEP of the other element address. 2082 Value *OtherElt = 0; 2083 unsigned OtherEltAlign = MemAlignment; 2084 2085 if (OtherPtr) { 2086 Value *Idx[2] = { Zero, 2087 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2088 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, 2089 OtherPtr->getName()+"."+Twine(i), 2090 MI); 2091 uint64_t EltOffset; 2092 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2093 Type *OtherTy = OtherPtrTy->getElementType(); 2094 if (StructType *ST = dyn_cast<StructType>(OtherTy)) { 2095 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2096 } else { 2097 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2098 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2099 } 2100 2101 // The alignment of the other pointer is the guaranteed alignment of the 2102 // element, which is affected by both the known alignment of the whole 2103 // mem intrinsic and the alignment of the element. If the alignment of 2104 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2105 // known alignment is just 4 bytes. 2106 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2107 } 2108 2109 Value *EltPtr = NewElts[i]; 2110 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2111 2112 // If we got down to a scalar, insert a load or store as appropriate. 2113 if (EltTy->isSingleValueType()) { 2114 if (isa<MemTransferInst>(MI)) { 2115 if (SROADest) { 2116 // From Other to Alloca. 2117 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2118 new StoreInst(Elt, EltPtr, MI); 2119 } else { 2120 // From Alloca to Other. 2121 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2122 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2123 } 2124 continue; 2125 } 2126 assert(isa<MemSetInst>(MI)); 2127 2128 // If the stored element is zero (common case), just store a null 2129 // constant. 2130 Constant *StoreVal; 2131 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2132 if (CI->isZero()) { 2133 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2134 } else { 2135 // If EltTy is a vector type, get the element type. 2136 Type *ValTy = EltTy->getScalarType(); 2137 2138 // Construct an integer with the right value. 2139 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2140 APInt OneVal(EltSize, CI->getZExtValue()); 2141 APInt TotalVal(OneVal); 2142 // Set each byte. 2143 for (unsigned i = 0; 8*i < EltSize; ++i) { 2144 TotalVal = TotalVal.shl(8); 2145 TotalVal |= OneVal; 2146 } 2147 2148 // Convert the integer value to the appropriate type. 2149 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2150 if (ValTy->isPointerTy()) 2151 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2152 else if (ValTy->isFloatingPointTy()) 2153 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2154 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2155 2156 // If the requested value was a vector constant, create it. 2157 if (EltTy->isVectorTy()) { 2158 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements(); 2159 StoreVal = ConstantVector::getSplat(NumElts, StoreVal); 2160 } 2161 } 2162 new StoreInst(StoreVal, EltPtr, MI); 2163 continue; 2164 } 2165 // Otherwise, if we're storing a byte variable, use a memset call for 2166 // this element. 2167 } 2168 2169 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2170 if (!EltSize) 2171 continue; 2172 2173 IRBuilder<> Builder(MI); 2174 2175 // Finally, insert the meminst for this element. 2176 if (isa<MemSetInst>(MI)) { 2177 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2178 MI->isVolatile()); 2179 } else { 2180 assert(isa<MemTransferInst>(MI)); 2181 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2182 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2183 2184 if (isa<MemCpyInst>(MI)) 2185 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2186 else 2187 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2188 } 2189 } 2190 DeadInsts.push_back(MI); 2191 } 2192 2193 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2194 /// overwrites the entire allocation. Extract out the pieces of the stored 2195 /// integer and store them individually. 2196 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2197 SmallVector<AllocaInst*, 32> &NewElts){ 2198 // Extract each element out of the integer according to its structure offset 2199 // and store the element value to the individual alloca. 2200 Value *SrcVal = SI->getOperand(0); 2201 Type *AllocaEltTy = AI->getAllocatedType(); 2202 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2203 2204 IRBuilder<> Builder(SI); 2205 2206 // Handle tail padding by extending the operand 2207 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2208 SrcVal = Builder.CreateZExt(SrcVal, 2209 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2210 2211 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2212 << '\n'); 2213 2214 // There are two forms here: AI could be an array or struct. Both cases 2215 // have different ways to compute the element offset. 2216 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2217 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2218 2219 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2220 // Get the number of bits to shift SrcVal to get the value. 2221 Type *FieldTy = EltSTy->getElementType(i); 2222 uint64_t Shift = Layout->getElementOffsetInBits(i); 2223 2224 if (TD->isBigEndian()) 2225 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2226 2227 Value *EltVal = SrcVal; 2228 if (Shift) { 2229 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2230 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2231 } 2232 2233 // Truncate down to an integer of the right size. 2234 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2235 2236 // Ignore zero sized fields like {}, they obviously contain no data. 2237 if (FieldSizeBits == 0) continue; 2238 2239 if (FieldSizeBits != AllocaSizeBits) 2240 EltVal = Builder.CreateTrunc(EltVal, 2241 IntegerType::get(SI->getContext(), FieldSizeBits)); 2242 Value *DestField = NewElts[i]; 2243 if (EltVal->getType() == FieldTy) { 2244 // Storing to an integer field of this size, just do it. 2245 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2246 // Bitcast to the right element type (for fp/vector values). 2247 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2248 } else { 2249 // Otherwise, bitcast the dest pointer (for aggregates). 2250 DestField = Builder.CreateBitCast(DestField, 2251 PointerType::getUnqual(EltVal->getType())); 2252 } 2253 new StoreInst(EltVal, DestField, SI); 2254 } 2255 2256 } else { 2257 ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2258 Type *ArrayEltTy = ATy->getElementType(); 2259 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2260 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2261 2262 uint64_t Shift; 2263 2264 if (TD->isBigEndian()) 2265 Shift = AllocaSizeBits-ElementOffset; 2266 else 2267 Shift = 0; 2268 2269 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2270 // Ignore zero sized fields like {}, they obviously contain no data. 2271 if (ElementSizeBits == 0) continue; 2272 2273 Value *EltVal = SrcVal; 2274 if (Shift) { 2275 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2276 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2277 } 2278 2279 // Truncate down to an integer of the right size. 2280 if (ElementSizeBits != AllocaSizeBits) 2281 EltVal = Builder.CreateTrunc(EltVal, 2282 IntegerType::get(SI->getContext(), 2283 ElementSizeBits)); 2284 Value *DestField = NewElts[i]; 2285 if (EltVal->getType() == ArrayEltTy) { 2286 // Storing to an integer field of this size, just do it. 2287 } else if (ArrayEltTy->isFloatingPointTy() || 2288 ArrayEltTy->isVectorTy()) { 2289 // Bitcast to the right element type (for fp/vector values). 2290 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2291 } else { 2292 // Otherwise, bitcast the dest pointer (for aggregates). 2293 DestField = Builder.CreateBitCast(DestField, 2294 PointerType::getUnqual(EltVal->getType())); 2295 } 2296 new StoreInst(EltVal, DestField, SI); 2297 2298 if (TD->isBigEndian()) 2299 Shift -= ElementOffset; 2300 else 2301 Shift += ElementOffset; 2302 } 2303 } 2304 2305 DeadInsts.push_back(SI); 2306 } 2307 2308 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2309 /// an integer. Load the individual pieces to form the aggregate value. 2310 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2311 SmallVector<AllocaInst*, 32> &NewElts) { 2312 // Extract each element out of the NewElts according to its structure offset 2313 // and form the result value. 2314 Type *AllocaEltTy = AI->getAllocatedType(); 2315 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2316 2317 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2318 << '\n'); 2319 2320 // There are two forms here: AI could be an array or struct. Both cases 2321 // have different ways to compute the element offset. 2322 const StructLayout *Layout = 0; 2323 uint64_t ArrayEltBitOffset = 0; 2324 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2325 Layout = TD->getStructLayout(EltSTy); 2326 } else { 2327 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2328 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2329 } 2330 2331 Value *ResultVal = 2332 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2333 2334 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2335 // Load the value from the alloca. If the NewElt is an aggregate, cast 2336 // the pointer to an integer of the same size before doing the load. 2337 Value *SrcField = NewElts[i]; 2338 Type *FieldTy = 2339 cast<PointerType>(SrcField->getType())->getElementType(); 2340 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2341 2342 // Ignore zero sized fields like {}, they obviously contain no data. 2343 if (FieldSizeBits == 0) continue; 2344 2345 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2346 FieldSizeBits); 2347 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2348 !FieldTy->isVectorTy()) 2349 SrcField = new BitCastInst(SrcField, 2350 PointerType::getUnqual(FieldIntTy), 2351 "", LI); 2352 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2353 2354 // If SrcField is a fp or vector of the right size but that isn't an 2355 // integer type, bitcast to an integer so we can shift it. 2356 if (SrcField->getType() != FieldIntTy) 2357 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2358 2359 // Zero extend the field to be the same size as the final alloca so that 2360 // we can shift and insert it. 2361 if (SrcField->getType() != ResultVal->getType()) 2362 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2363 2364 // Determine the number of bits to shift SrcField. 2365 uint64_t Shift; 2366 if (Layout) // Struct case. 2367 Shift = Layout->getElementOffsetInBits(i); 2368 else // Array case. 2369 Shift = i*ArrayEltBitOffset; 2370 2371 if (TD->isBigEndian()) 2372 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2373 2374 if (Shift) { 2375 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2376 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2377 } 2378 2379 // Don't create an 'or x, 0' on the first iteration. 2380 if (!isa<Constant>(ResultVal) || 2381 !cast<Constant>(ResultVal)->isNullValue()) 2382 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2383 else 2384 ResultVal = SrcField; 2385 } 2386 2387 // Handle tail padding by truncating the result 2388 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2389 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2390 2391 LI->replaceAllUsesWith(ResultVal); 2392 DeadInsts.push_back(LI); 2393 } 2394 2395 /// HasPadding - Return true if the specified type has any structure or 2396 /// alignment padding in between the elements that would be split apart 2397 /// by SROA; return false otherwise. 2398 static bool HasPadding(Type *Ty, const TargetData &TD) { 2399 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2400 Ty = ATy->getElementType(); 2401 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2402 } 2403 2404 // SROA currently handles only Arrays and Structs. 2405 StructType *STy = cast<StructType>(Ty); 2406 const StructLayout *SL = TD.getStructLayout(STy); 2407 unsigned PrevFieldBitOffset = 0; 2408 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2409 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2410 2411 // Check to see if there is any padding between this element and the 2412 // previous one. 2413 if (i) { 2414 unsigned PrevFieldEnd = 2415 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2416 if (PrevFieldEnd < FieldBitOffset) 2417 return true; 2418 } 2419 PrevFieldBitOffset = FieldBitOffset; 2420 } 2421 // Check for tail padding. 2422 if (unsigned EltCount = STy->getNumElements()) { 2423 unsigned PrevFieldEnd = PrevFieldBitOffset + 2424 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2425 if (PrevFieldEnd < SL->getSizeInBits()) 2426 return true; 2427 } 2428 return false; 2429 } 2430 2431 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2432 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2433 /// or 1 if safe after canonicalization has been performed. 2434 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2435 // Loop over the use list of the alloca. We can only transform it if all of 2436 // the users are safe to transform. 2437 AllocaInfo Info(AI); 2438 2439 isSafeForScalarRepl(AI, 0, Info); 2440 if (Info.isUnsafe) { 2441 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2442 return false; 2443 } 2444 2445 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2446 // source and destination, we have to be careful. In particular, the memcpy 2447 // could be moving around elements that live in structure padding of the LLVM 2448 // types, but may actually be used. In these cases, we refuse to promote the 2449 // struct. 2450 if (Info.isMemCpySrc && Info.isMemCpyDst && 2451 HasPadding(AI->getAllocatedType(), *TD)) 2452 return false; 2453 2454 // If the alloca never has an access to just *part* of it, but is accessed 2455 // via loads and stores, then we should use ConvertToScalarInfo to promote 2456 // the alloca instead of promoting each piece at a time and inserting fission 2457 // and fusion code. 2458 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2459 // If the struct/array just has one element, use basic SRoA. 2460 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2461 if (ST->getNumElements() > 1) return false; 2462 } else { 2463 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2464 return false; 2465 } 2466 } 2467 2468 return true; 2469 } 2470 2471 2472 2473 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2474 /// some part of a constant global variable. This intentionally only accepts 2475 /// constant expressions because we don't can't rewrite arbitrary instructions. 2476 static bool PointsToConstantGlobal(Value *V) { 2477 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2478 return GV->isConstant(); 2479 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2480 if (CE->getOpcode() == Instruction::BitCast || 2481 CE->getOpcode() == Instruction::GetElementPtr) 2482 return PointsToConstantGlobal(CE->getOperand(0)); 2483 return false; 2484 } 2485 2486 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2487 /// pointer to an alloca. Ignore any reads of the pointer, return false if we 2488 /// see any stores or other unknown uses. If we see pointer arithmetic, keep 2489 /// track of whether it moves the pointer (with isOffset) but otherwise traverse 2490 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2491 /// the alloca, and if the source pointer is a pointer to a constant global, we 2492 /// can optimize this. 2493 static bool 2494 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2495 bool isOffset, 2496 SmallVector<Instruction *, 4> &LifetimeMarkers) { 2497 // We track lifetime intrinsics as we encounter them. If we decide to go 2498 // ahead and replace the value with the global, this lets the caller quickly 2499 // eliminate the markers. 2500 2501 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2502 User *U = cast<Instruction>(*UI); 2503 2504 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2505 // Ignore non-volatile loads, they are always ok. 2506 if (!LI->isSimple()) return false; 2507 continue; 2508 } 2509 2510 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2511 // If uses of the bitcast are ok, we are ok. 2512 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset, 2513 LifetimeMarkers)) 2514 return false; 2515 continue; 2516 } 2517 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2518 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2519 // doesn't, it does. 2520 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2521 isOffset || !GEP->hasAllZeroIndices(), 2522 LifetimeMarkers)) 2523 return false; 2524 continue; 2525 } 2526 2527 if (CallSite CS = U) { 2528 // If this is the function being called then we treat it like a load and 2529 // ignore it. 2530 if (CS.isCallee(UI)) 2531 continue; 2532 2533 // If this is a readonly/readnone call site, then we know it is just a 2534 // load (but one that potentially returns the value itself), so we can 2535 // ignore it if we know that the value isn't captured. 2536 unsigned ArgNo = CS.getArgumentNo(UI); 2537 if (CS.onlyReadsMemory() && 2538 (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo))) 2539 continue; 2540 2541 // If this is being passed as a byval argument, the caller is making a 2542 // copy, so it is only a read of the alloca. 2543 if (CS.isByValArgument(ArgNo)) 2544 continue; 2545 } 2546 2547 // Lifetime intrinsics can be handled by the caller. 2548 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 2549 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 2550 II->getIntrinsicID() == Intrinsic::lifetime_end) { 2551 assert(II->use_empty() && "Lifetime markers have no result to use!"); 2552 LifetimeMarkers.push_back(II); 2553 continue; 2554 } 2555 } 2556 2557 // If this is isn't our memcpy/memmove, reject it as something we can't 2558 // handle. 2559 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2560 if (MI == 0) 2561 return false; 2562 2563 // If the transfer is using the alloca as a source of the transfer, then 2564 // ignore it since it is a load (unless the transfer is volatile). 2565 if (UI.getOperandNo() == 1) { 2566 if (MI->isVolatile()) return false; 2567 continue; 2568 } 2569 2570 // If we already have seen a copy, reject the second one. 2571 if (TheCopy) return false; 2572 2573 // If the pointer has been offset from the start of the alloca, we can't 2574 // safely handle this. 2575 if (isOffset) return false; 2576 2577 // If the memintrinsic isn't using the alloca as the dest, reject it. 2578 if (UI.getOperandNo() != 0) return false; 2579 2580 // If the source of the memcpy/move is not a constant global, reject it. 2581 if (!PointsToConstantGlobal(MI->getSource())) 2582 return false; 2583 2584 // Otherwise, the transform is safe. Remember the copy instruction. 2585 TheCopy = MI; 2586 } 2587 return true; 2588 } 2589 2590 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2591 /// modified by a copy from a constant global. If we can prove this, we can 2592 /// replace any uses of the alloca with uses of the global directly. 2593 MemTransferInst * 2594 SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI, 2595 SmallVector<Instruction*, 4> &ToDelete) { 2596 MemTransferInst *TheCopy = 0; 2597 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete)) 2598 return TheCopy; 2599 return 0; 2600 } 2601