1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This transformation implements the well known scalar replacement of 11 // aggregates transformation. This xform breaks up alloca instructions of 12 // aggregate type (structure or array) into individual alloca instructions for 13 // each member (if possible). Then, if possible, it transforms the individual 14 // alloca instructions into nice clean scalar SSA form. 15 // 16 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because 17 // often interact, especially for C++ programs. As such, iterating between 18 // SRoA, then Mem2Reg until we run out of things to promote works well. 19 // 20 //===----------------------------------------------------------------------===// 21 22 #define DEBUG_TYPE "scalarrepl" 23 #include "llvm/Transforms/Scalar.h" 24 #include "llvm/Constants.h" 25 #include "llvm/DerivedTypes.h" 26 #include "llvm/Function.h" 27 #include "llvm/GlobalVariable.h" 28 #include "llvm/Instructions.h" 29 #include "llvm/IntrinsicInst.h" 30 #include "llvm/LLVMContext.h" 31 #include "llvm/Module.h" 32 #include "llvm/Pass.h" 33 #include "llvm/Analysis/DebugInfo.h" 34 #include "llvm/Analysis/DIBuilder.h" 35 #include "llvm/Analysis/Dominators.h" 36 #include "llvm/Analysis/Loads.h" 37 #include "llvm/Analysis/ValueTracking.h" 38 #include "llvm/Target/TargetData.h" 39 #include "llvm/Transforms/Utils/PromoteMemToReg.h" 40 #include "llvm/Transforms/Utils/Local.h" 41 #include "llvm/Transforms/Utils/SSAUpdater.h" 42 #include "llvm/Support/CallSite.h" 43 #include "llvm/Support/Debug.h" 44 #include "llvm/Support/ErrorHandling.h" 45 #include "llvm/Support/GetElementPtrTypeIterator.h" 46 #include "llvm/Support/IRBuilder.h" 47 #include "llvm/Support/MathExtras.h" 48 #include "llvm/Support/raw_ostream.h" 49 #include "llvm/ADT/SetVector.h" 50 #include "llvm/ADT/SmallVector.h" 51 #include "llvm/ADT/Statistic.h" 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 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 457 Indices); 458 // See if all uses can be converted. 459 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 460 return false; 461 IsNotTrivial = true; // Can't be mem2reg'd. 462 HadNonMemTransferAccess = true; 463 continue; 464 } 465 466 // If this is a constant sized memset of a constant value (e.g. 0) we can 467 // handle it. 468 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 469 // Store of constant value. 470 if (!isa<ConstantInt>(MSI->getValue())) 471 return false; 472 473 // Store of constant size. 474 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength()); 475 if (!Len) 476 return false; 477 478 // If the size differs from the alloca, we can only convert the alloca to 479 // an integer bag-of-bits. 480 // FIXME: This should handle all of the cases that are currently accepted 481 // as vector element insertions. 482 if (Len->getZExtValue() != AllocaSize || Offset != 0) 483 ScalarKind = Integer; 484 485 IsNotTrivial = true; // Can't be mem2reg'd. 486 HadNonMemTransferAccess = true; 487 continue; 488 } 489 490 // If this is a memcpy or memmove into or out of the whole allocation, we 491 // can handle it like a load or store of the scalar type. 492 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 493 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 494 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 495 return false; 496 497 IsNotTrivial = true; // Can't be mem2reg'd. 498 continue; 499 } 500 501 // If this is a lifetime intrinsic, we can handle it. 502 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 503 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 504 II->getIntrinsicID() == Intrinsic::lifetime_end) { 505 continue; 506 } 507 } 508 509 // Otherwise, we cannot handle this! 510 return false; 511 } 512 513 return true; 514 } 515 516 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 517 /// directly. This happens when we are converting an "integer union" to a 518 /// single integer scalar, or when we are converting a "vector union" to a 519 /// vector with insert/extractelement instructions. 520 /// 521 /// Offset is an offset from the original alloca, in bits that need to be 522 /// shifted to the right. By the end of this, there should be no uses of Ptr. 523 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 524 uint64_t Offset) { 525 while (!Ptr->use_empty()) { 526 Instruction *User = cast<Instruction>(Ptr->use_back()); 527 528 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 529 ConvertUsesToScalar(CI, NewAI, Offset); 530 CI->eraseFromParent(); 531 continue; 532 } 533 534 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 535 // Compute the offset that this GEP adds to the pointer. 536 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 537 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 538 Indices); 539 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 540 GEP->eraseFromParent(); 541 continue; 542 } 543 544 IRBuilder<> Builder(User); 545 546 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 547 // The load is a bit extract from NewAI shifted right by Offset bits. 548 Value *LoadedVal = Builder.CreateLoad(NewAI); 549 Value *NewLoadVal 550 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 551 LI->replaceAllUsesWith(NewLoadVal); 552 LI->eraseFromParent(); 553 continue; 554 } 555 556 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 557 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 558 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 559 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 560 Builder); 561 Builder.CreateStore(New, NewAI); 562 SI->eraseFromParent(); 563 564 // If the load we just inserted is now dead, then the inserted store 565 // overwrote the entire thing. 566 if (Old->use_empty()) 567 Old->eraseFromParent(); 568 continue; 569 } 570 571 // If this is a constant sized memset of a constant value (e.g. 0) we can 572 // transform it into a store of the expanded constant value. 573 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 574 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 575 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 576 if (NumBytes != 0) { 577 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 578 579 // Compute the value replicated the right number of times. 580 APInt APVal(NumBytes*8, Val); 581 582 // Splat the value if non-zero. 583 if (Val) 584 for (unsigned i = 1; i != NumBytes; ++i) 585 APVal |= APVal << 8; 586 587 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 588 Value *New = ConvertScalar_InsertValue( 589 ConstantInt::get(User->getContext(), APVal), 590 Old, Offset, Builder); 591 Builder.CreateStore(New, NewAI); 592 593 // If the load we just inserted is now dead, then the memset overwrote 594 // the entire thing. 595 if (Old->use_empty()) 596 Old->eraseFromParent(); 597 } 598 MSI->eraseFromParent(); 599 continue; 600 } 601 602 // If this is a memcpy or memmove into or out of the whole allocation, we 603 // can handle it like a load or store of the scalar type. 604 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 605 assert(Offset == 0 && "must be store to start of alloca"); 606 607 // If the source and destination are both to the same alloca, then this is 608 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 609 // as appropriate. 610 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 611 612 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 613 // Dest must be OrigAI, change this to be a load from the original 614 // pointer (bitcasted), then a store to our new alloca. 615 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 616 Value *SrcPtr = MTI->getSource(); 617 PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 618 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 619 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 620 AIPTy = PointerType::get(AIPTy->getElementType(), 621 SPTy->getAddressSpace()); 622 } 623 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 624 625 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 626 SrcVal->setAlignment(MTI->getAlignment()); 627 Builder.CreateStore(SrcVal, NewAI); 628 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 629 // Src must be OrigAI, change this to be a load from NewAI then a store 630 // through the original dest pointer (bitcasted). 631 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 632 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 633 634 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 635 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 636 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 637 AIPTy = PointerType::get(AIPTy->getElementType(), 638 DPTy->getAddressSpace()); 639 } 640 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 641 642 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 643 NewStore->setAlignment(MTI->getAlignment()); 644 } else { 645 // Noop transfer. Src == Dst 646 } 647 648 MTI->eraseFromParent(); 649 continue; 650 } 651 652 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 653 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 654 II->getIntrinsicID() == Intrinsic::lifetime_end) { 655 // There's no need to preserve these, as the resulting alloca will be 656 // converted to a register anyways. 657 II->eraseFromParent(); 658 continue; 659 } 660 } 661 662 llvm_unreachable("Unsupported operation!"); 663 } 664 } 665 666 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 667 /// or vector value FromVal, extracting the bits from the offset specified by 668 /// Offset. This returns the value, which is of type ToType. 669 /// 670 /// This happens when we are converting an "integer union" to a single 671 /// integer scalar, or when we are converting a "vector union" to a vector with 672 /// insert/extractelement instructions. 673 /// 674 /// Offset is an offset from the original alloca, in bits that need to be 675 /// shifted to the right. 676 Value *ConvertToScalarInfo:: 677 ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, 678 uint64_t Offset, IRBuilder<> &Builder) { 679 // If the load is of the whole new alloca, no conversion is needed. 680 Type *FromType = FromVal->getType(); 681 if (FromType == ToType && Offset == 0) 682 return FromVal; 683 684 // If the result alloca is a vector type, this is either an element 685 // access or a bitcast to another vector type of the same size. 686 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) { 687 unsigned FromTypeSize = TD.getTypeAllocSize(FromType); 688 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 689 if (FromTypeSize == ToTypeSize) 690 return Builder.CreateBitCast(FromVal, ToType); 691 692 // Otherwise it must be an element access. 693 unsigned Elt = 0; 694 if (Offset) { 695 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 696 Elt = Offset/EltSize; 697 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 698 } 699 // Return the element extracted out of it. 700 Value *V = Builder.CreateExtractElement(FromVal, Builder.getInt32(Elt)); 701 if (V->getType() != ToType) 702 V = Builder.CreateBitCast(V, ToType); 703 return V; 704 } 705 706 // If ToType is a first class aggregate, extract out each of the pieces and 707 // use insertvalue's to form the FCA. 708 if (StructType *ST = dyn_cast<StructType>(ToType)) { 709 const StructLayout &Layout = *TD.getStructLayout(ST); 710 Value *Res = UndefValue::get(ST); 711 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 712 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 713 Offset+Layout.getElementOffsetInBits(i), 714 Builder); 715 Res = Builder.CreateInsertValue(Res, Elt, i); 716 } 717 return Res; 718 } 719 720 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 721 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 722 Value *Res = UndefValue::get(AT); 723 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 724 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 725 Offset+i*EltSize, Builder); 726 Res = Builder.CreateInsertValue(Res, Elt, i); 727 } 728 return Res; 729 } 730 731 // Otherwise, this must be a union that was converted to an integer value. 732 IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 733 734 // If this is a big-endian system and the load is narrower than the 735 // full alloca type, we need to do a shift to get the right bits. 736 int ShAmt = 0; 737 if (TD.isBigEndian()) { 738 // On big-endian machines, the lowest bit is stored at the bit offset 739 // from the pointer given by getTypeStoreSizeInBits. This matters for 740 // integers with a bitwidth that is not a multiple of 8. 741 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 742 TD.getTypeStoreSizeInBits(ToType) - Offset; 743 } else { 744 ShAmt = Offset; 745 } 746 747 // Note: we support negative bitwidths (with shl) which are not defined. 748 // We do this to support (f.e.) loads off the end of a structure where 749 // only some bits are used. 750 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 751 FromVal = Builder.CreateLShr(FromVal, 752 ConstantInt::get(FromVal->getType(), ShAmt)); 753 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 754 FromVal = Builder.CreateShl(FromVal, 755 ConstantInt::get(FromVal->getType(), -ShAmt)); 756 757 // Finally, unconditionally truncate the integer to the right width. 758 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 759 if (LIBitWidth < NTy->getBitWidth()) 760 FromVal = 761 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 762 LIBitWidth)); 763 else if (LIBitWidth > NTy->getBitWidth()) 764 FromVal = 765 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 766 LIBitWidth)); 767 768 // If the result is an integer, this is a trunc or bitcast. 769 if (ToType->isIntegerTy()) { 770 // Should be done. 771 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 772 // Just do a bitcast, we know the sizes match up. 773 FromVal = Builder.CreateBitCast(FromVal, ToType); 774 } else { 775 // Otherwise must be a pointer. 776 FromVal = Builder.CreateIntToPtr(FromVal, ToType); 777 } 778 assert(FromVal->getType() == ToType && "Didn't convert right?"); 779 return FromVal; 780 } 781 782 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 783 /// or vector value "Old" at the offset specified by Offset. 784 /// 785 /// This happens when we are converting an "integer union" to a 786 /// single integer scalar, or when we are converting a "vector union" to a 787 /// vector with insert/extractelement instructions. 788 /// 789 /// Offset is an offset from the original alloca, in bits that need to be 790 /// shifted to the right. 791 Value *ConvertToScalarInfo:: 792 ConvertScalar_InsertValue(Value *SV, Value *Old, 793 uint64_t Offset, IRBuilder<> &Builder) { 794 // Convert the stored type to the actual type, shift it left to insert 795 // then 'or' into place. 796 Type *AllocaType = Old->getType(); 797 LLVMContext &Context = Old->getContext(); 798 799 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 800 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 801 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 802 803 // Changing the whole vector with memset or with an access of a different 804 // vector type? 805 if (ValSize == VecSize) 806 return Builder.CreateBitCast(SV, AllocaType); 807 808 // Must be an element insertion. 809 assert(SV->getType() == VTy->getElementType()); 810 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 811 unsigned Elt = Offset/EltSize; 812 return Builder.CreateInsertElement(Old, SV, Builder.getInt32(Elt)); 813 } 814 815 // If SV is a first-class aggregate value, insert each value recursively. 816 if (StructType *ST = dyn_cast<StructType>(SV->getType())) { 817 const StructLayout &Layout = *TD.getStructLayout(ST); 818 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 819 Value *Elt = Builder.CreateExtractValue(SV, i); 820 Old = ConvertScalar_InsertValue(Elt, Old, 821 Offset+Layout.getElementOffsetInBits(i), 822 Builder); 823 } 824 return Old; 825 } 826 827 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 828 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 829 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 830 Value *Elt = Builder.CreateExtractValue(SV, i); 831 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 832 } 833 return Old; 834 } 835 836 // If SV is a float, convert it to the appropriate integer type. 837 // If it is a pointer, do the same. 838 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 839 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 840 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 841 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 842 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 843 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth)); 844 else if (SV->getType()->isPointerTy()) 845 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext())); 846 847 // Zero extend or truncate the value if needed. 848 if (SV->getType() != AllocaType) { 849 if (SV->getType()->getPrimitiveSizeInBits() < 850 AllocaType->getPrimitiveSizeInBits()) 851 SV = Builder.CreateZExt(SV, AllocaType); 852 else { 853 // Truncation may be needed if storing more than the alloca can hold 854 // (undefined behavior). 855 SV = Builder.CreateTrunc(SV, AllocaType); 856 SrcWidth = DestWidth; 857 SrcStoreWidth = DestStoreWidth; 858 } 859 } 860 861 // If this is a big-endian system and the store is narrower than the 862 // full alloca type, we need to do a shift to get the right bits. 863 int ShAmt = 0; 864 if (TD.isBigEndian()) { 865 // On big-endian machines, the lowest bit is stored at the bit offset 866 // from the pointer given by getTypeStoreSizeInBits. This matters for 867 // integers with a bitwidth that is not a multiple of 8. 868 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 869 } else { 870 ShAmt = Offset; 871 } 872 873 // Note: we support negative bitwidths (with shr) which are not defined. 874 // We do this to support (f.e.) stores off the end of a structure where 875 // only some bits in the structure are set. 876 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 877 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 878 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt)); 879 Mask <<= ShAmt; 880 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 881 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt)); 882 Mask = Mask.lshr(-ShAmt); 883 } 884 885 // Mask out the bits we are about to insert from the old value, and or 886 // in the new bits. 887 if (SrcWidth != DestWidth) { 888 assert(DestWidth > SrcWidth); 889 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 890 SV = Builder.CreateOr(Old, SV, "ins"); 891 } 892 return SV; 893 } 894 895 896 //===----------------------------------------------------------------------===// 897 // SRoA Driver 898 //===----------------------------------------------------------------------===// 899 900 901 bool SROA::runOnFunction(Function &F) { 902 TD = getAnalysisIfAvailable<TargetData>(); 903 904 bool Changed = performPromotion(F); 905 906 // FIXME: ScalarRepl currently depends on TargetData more than it 907 // theoretically needs to. It should be refactored in order to support 908 // target-independent IR. Until this is done, just skip the actual 909 // scalar-replacement portion of this pass. 910 if (!TD) return Changed; 911 912 while (1) { 913 bool LocalChange = performScalarRepl(F); 914 if (!LocalChange) break; // No need to repromote if no scalarrepl 915 Changed = true; 916 LocalChange = performPromotion(F); 917 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 918 } 919 920 return Changed; 921 } 922 923 namespace { 924 class AllocaPromoter : public LoadAndStorePromoter { 925 AllocaInst *AI; 926 DIBuilder *DIB; 927 SmallVector<DbgDeclareInst *, 4> DDIs; 928 SmallVector<DbgValueInst *, 4> DVIs; 929 public: 930 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 931 DIBuilder *DB) 932 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {} 933 934 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 935 // Remember which alloca we're promoting (for isInstInList). 936 this->AI = AI; 937 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) { 938 for (Value::use_iterator UI = DebugNode->use_begin(), 939 E = DebugNode->use_end(); UI != E; ++UI) 940 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI)) 941 DDIs.push_back(DDI); 942 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI)) 943 DVIs.push_back(DVI); 944 } 945 946 LoadAndStorePromoter::run(Insts); 947 AI->eraseFromParent(); 948 for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(), 949 E = DDIs.end(); I != E; ++I) { 950 DbgDeclareInst *DDI = *I; 951 DDI->eraseFromParent(); 952 } 953 for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(), 954 E = DVIs.end(); I != E; ++I) { 955 DbgValueInst *DVI = *I; 956 DVI->eraseFromParent(); 957 } 958 } 959 960 virtual bool isInstInList(Instruction *I, 961 const SmallVectorImpl<Instruction*> &Insts) const { 962 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 963 return LI->getOperand(0) == AI; 964 return cast<StoreInst>(I)->getPointerOperand() == AI; 965 } 966 967 virtual void updateDebugInfo(Instruction *Inst) const { 968 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(), 969 E = DDIs.end(); I != E; ++I) { 970 DbgDeclareInst *DDI = *I; 971 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 972 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB); 973 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 974 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB); 975 } 976 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(), 977 E = DVIs.end(); I != E; ++I) { 978 DbgValueInst *DVI = *I; 979 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 980 Instruction *DbgVal = NULL; 981 // If an argument is zero extended then use argument directly. The ZExt 982 // may be zapped by an optimization pass in future. 983 Argument *ExtendedArg = NULL; 984 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 985 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0)); 986 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 987 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0)); 988 if (ExtendedArg) 989 DbgVal = DIB->insertDbgValueIntrinsic(ExtendedArg, 0, 990 DIVariable(DVI->getVariable()), 991 SI); 992 else 993 DbgVal = DIB->insertDbgValueIntrinsic(SI->getOperand(0), 0, 994 DIVariable(DVI->getVariable()), 995 SI); 996 DbgVal->setDebugLoc(DVI->getDebugLoc()); 997 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 998 Instruction *DbgVal = 999 DIB->insertDbgValueIntrinsic(LI->getOperand(0), 0, 1000 DIVariable(DVI->getVariable()), LI); 1001 DbgVal->setDebugLoc(DVI->getDebugLoc()); 1002 } 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 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1522 UI.getOperandNo() == 0, Info, MI, 1523 true /*AllowWholeAccess*/); 1524 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1525 if (!LI->isSimple()) 1526 return MarkUnsafe(Info, User); 1527 Type *LIType = LI->getType(); 1528 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1529 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1530 Info.hasALoadOrStore = true; 1531 1532 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1533 // Store is ok if storing INTO the pointer, not storing the pointer 1534 if (!SI->isSimple() || SI->getOperand(0) == I) 1535 return MarkUnsafe(Info, User); 1536 1537 Type *SIType = SI->getOperand(0)->getType(); 1538 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1539 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1540 Info.hasALoadOrStore = true; 1541 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1542 if (II->getIntrinsicID() != Intrinsic::lifetime_start && 1543 II->getIntrinsicID() != Intrinsic::lifetime_end) 1544 return MarkUnsafe(Info, User); 1545 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1546 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1547 } else { 1548 return MarkUnsafe(Info, User); 1549 } 1550 if (Info.isUnsafe) return; 1551 } 1552 } 1553 1554 1555 /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1556 /// derived from the alloca, we can often still split the alloca into elements. 1557 /// This is useful if we have a large alloca where one element is phi'd 1558 /// together somewhere: we can SRoA and promote all the other elements even if 1559 /// we end up not being able to promote this one. 1560 /// 1561 /// All we require is that the uses of the PHI do not index into other parts of 1562 /// the alloca. The most important use case for this is single load and stores 1563 /// that are PHI'd together, which can happen due to code sinking. 1564 void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1565 AllocaInfo &Info) { 1566 // If we've already checked this PHI, don't do it again. 1567 if (PHINode *PN = dyn_cast<PHINode>(I)) 1568 if (!Info.CheckedPHIs.insert(PN)) 1569 return; 1570 1571 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1572 Instruction *User = cast<Instruction>(*UI); 1573 1574 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1575 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1576 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1577 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1578 // but would have to prove that we're staying inside of an element being 1579 // promoted. 1580 if (!GEPI->hasAllZeroIndices()) 1581 return MarkUnsafe(Info, User); 1582 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1583 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1584 if (!LI->isSimple()) 1585 return MarkUnsafe(Info, User); 1586 Type *LIType = LI->getType(); 1587 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1588 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1589 Info.hasALoadOrStore = true; 1590 1591 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1592 // Store is ok if storing INTO the pointer, not storing the pointer 1593 if (!SI->isSimple() || SI->getOperand(0) == I) 1594 return MarkUnsafe(Info, User); 1595 1596 Type *SIType = SI->getOperand(0)->getType(); 1597 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1598 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1599 Info.hasALoadOrStore = true; 1600 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1601 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1602 } else { 1603 return MarkUnsafe(Info, User); 1604 } 1605 if (Info.isUnsafe) return; 1606 } 1607 } 1608 1609 /// isSafeGEP - Check if a GEP instruction can be handled for scalar 1610 /// replacement. It is safe when all the indices are constant, in-bounds 1611 /// references, and when the resulting offset corresponds to an element within 1612 /// the alloca type. The results are flagged in the Info parameter. Upon 1613 /// return, Offset is adjusted as specified by the GEP indices. 1614 void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1615 uint64_t &Offset, AllocaInfo &Info) { 1616 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1617 if (GEPIt == E) 1618 return; 1619 1620 // Walk through the GEP type indices, checking the types that this indexes 1621 // into. 1622 for (; GEPIt != E; ++GEPIt) { 1623 // Ignore struct elements, no extra checking needed for these. 1624 if ((*GEPIt)->isStructTy()) 1625 continue; 1626 1627 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1628 if (!IdxVal) 1629 return MarkUnsafe(Info, GEPI); 1630 } 1631 1632 // Compute the offset due to this GEP and check if the alloca has a 1633 // component element at that offset. 1634 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1635 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1636 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1637 MarkUnsafe(Info, GEPI); 1638 } 1639 1640 /// isHomogeneousAggregate - Check if type T is a struct or array containing 1641 /// elements of the same type (which is always true for arrays). If so, 1642 /// return true with NumElts and EltTy set to the number of elements and the 1643 /// element type, respectively. 1644 static bool isHomogeneousAggregate(Type *T, unsigned &NumElts, 1645 Type *&EltTy) { 1646 if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1647 NumElts = AT->getNumElements(); 1648 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1649 return true; 1650 } 1651 if (StructType *ST = dyn_cast<StructType>(T)) { 1652 NumElts = ST->getNumContainedTypes(); 1653 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1654 for (unsigned n = 1; n < NumElts; ++n) { 1655 if (ST->getContainedType(n) != EltTy) 1656 return false; 1657 } 1658 return true; 1659 } 1660 return false; 1661 } 1662 1663 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1664 /// "homogeneous" aggregates with the same element type and number of elements. 1665 static bool isCompatibleAggregate(Type *T1, Type *T2) { 1666 if (T1 == T2) 1667 return true; 1668 1669 unsigned NumElts1, NumElts2; 1670 Type *EltTy1, *EltTy2; 1671 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1672 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1673 NumElts1 == NumElts2 && 1674 EltTy1 == EltTy2) 1675 return true; 1676 1677 return false; 1678 } 1679 1680 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1681 /// alloca or has an offset and size that corresponds to a component element 1682 /// within it. The offset checked here may have been formed from a GEP with a 1683 /// pointer bitcasted to a different type. 1684 /// 1685 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1686 /// unit. If false, it only allows accesses known to be in a single element. 1687 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1688 Type *MemOpType, bool isStore, 1689 AllocaInfo &Info, Instruction *TheAccess, 1690 bool AllowWholeAccess) { 1691 // Check if this is a load/store of the entire alloca. 1692 if (Offset == 0 && AllowWholeAccess && 1693 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1694 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1695 // loads/stores (which are essentially the same as the MemIntrinsics with 1696 // regard to copying padding between elements). But, if an alloca is 1697 // flagged as both a source and destination of such operations, we'll need 1698 // to check later for padding between elements. 1699 if (!MemOpType || MemOpType->isIntegerTy()) { 1700 if (isStore) 1701 Info.isMemCpyDst = true; 1702 else 1703 Info.isMemCpySrc = true; 1704 return; 1705 } 1706 // This is also safe for references using a type that is compatible with 1707 // the type of the alloca, so that loads/stores can be rewritten using 1708 // insertvalue/extractvalue. 1709 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1710 Info.hasSubelementAccess = true; 1711 return; 1712 } 1713 } 1714 // Check if the offset/size correspond to a component within the alloca type. 1715 Type *T = Info.AI->getAllocatedType(); 1716 if (TypeHasComponent(T, Offset, MemSize)) { 1717 Info.hasSubelementAccess = true; 1718 return; 1719 } 1720 1721 return MarkUnsafe(Info, TheAccess); 1722 } 1723 1724 /// TypeHasComponent - Return true if T has a component type with the 1725 /// specified offset and size. If Size is zero, do not check the size. 1726 bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { 1727 Type *EltTy; 1728 uint64_t EltSize; 1729 if (StructType *ST = dyn_cast<StructType>(T)) { 1730 const StructLayout *Layout = TD->getStructLayout(ST); 1731 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1732 EltTy = ST->getContainedType(EltIdx); 1733 EltSize = TD->getTypeAllocSize(EltTy); 1734 Offset -= Layout->getElementOffset(EltIdx); 1735 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1736 EltTy = AT->getElementType(); 1737 EltSize = TD->getTypeAllocSize(EltTy); 1738 if (Offset >= AT->getNumElements() * EltSize) 1739 return false; 1740 Offset %= EltSize; 1741 } else { 1742 return false; 1743 } 1744 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1745 return true; 1746 // Check if the component spans multiple elements. 1747 if (Offset + Size > EltSize) 1748 return false; 1749 return TypeHasComponent(EltTy, Offset, Size); 1750 } 1751 1752 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1753 /// the instruction I, which references it, to use the separate elements. 1754 /// Offset indicates the position within AI that is referenced by this 1755 /// instruction. 1756 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1757 SmallVector<AllocaInst*, 32> &NewElts) { 1758 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1759 Use &TheUse = UI.getUse(); 1760 Instruction *User = cast<Instruction>(*UI++); 1761 1762 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1763 RewriteBitCast(BC, AI, Offset, NewElts); 1764 continue; 1765 } 1766 1767 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1768 RewriteGEP(GEPI, AI, Offset, NewElts); 1769 continue; 1770 } 1771 1772 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1773 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1774 uint64_t MemSize = Length->getZExtValue(); 1775 if (Offset == 0 && 1776 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1777 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1778 // Otherwise the intrinsic can only touch a single element and the 1779 // address operand will be updated, so nothing else needs to be done. 1780 continue; 1781 } 1782 1783 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1784 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 1785 II->getIntrinsicID() == Intrinsic::lifetime_end) { 1786 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts); 1787 } 1788 continue; 1789 } 1790 1791 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1792 Type *LIType = LI->getType(); 1793 1794 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1795 // Replace: 1796 // %res = load { i32, i32 }* %alloc 1797 // with: 1798 // %load.0 = load i32* %alloc.0 1799 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1800 // %load.1 = load i32* %alloc.1 1801 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1802 // (Also works for arrays instead of structs) 1803 Value *Insert = UndefValue::get(LIType); 1804 IRBuilder<> Builder(LI); 1805 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1806 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1807 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1808 } 1809 LI->replaceAllUsesWith(Insert); 1810 DeadInsts.push_back(LI); 1811 } else if (LIType->isIntegerTy() && 1812 TD->getTypeAllocSize(LIType) == 1813 TD->getTypeAllocSize(AI->getAllocatedType())) { 1814 // If this is a load of the entire alloca to an integer, rewrite it. 1815 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1816 } 1817 continue; 1818 } 1819 1820 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1821 Value *Val = SI->getOperand(0); 1822 Type *SIType = Val->getType(); 1823 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1824 // Replace: 1825 // store { i32, i32 } %val, { i32, i32 }* %alloc 1826 // with: 1827 // %val.0 = extractvalue { i32, i32 } %val, 0 1828 // store i32 %val.0, i32* %alloc.0 1829 // %val.1 = extractvalue { i32, i32 } %val, 1 1830 // store i32 %val.1, i32* %alloc.1 1831 // (Also works for arrays instead of structs) 1832 IRBuilder<> Builder(SI); 1833 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1834 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1835 Builder.CreateStore(Extract, NewElts[i]); 1836 } 1837 DeadInsts.push_back(SI); 1838 } else if (SIType->isIntegerTy() && 1839 TD->getTypeAllocSize(SIType) == 1840 TD->getTypeAllocSize(AI->getAllocatedType())) { 1841 // If this is a store of the entire alloca from an integer, rewrite it. 1842 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1843 } 1844 continue; 1845 } 1846 1847 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1848 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1849 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1850 // the new pointer. 1851 if (!isa<AllocaInst>(I)) continue; 1852 1853 assert(Offset == 0 && NewElts[0] && 1854 "Direct alloca use should have a zero offset"); 1855 1856 // If we have a use of the alloca, we know the derived uses will be 1857 // utilizing just the first element of the scalarized result. Insert a 1858 // bitcast of the first alloca before the user as required. 1859 AllocaInst *NewAI = NewElts[0]; 1860 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1861 NewAI->moveBefore(BCI); 1862 TheUse = BCI; 1863 continue; 1864 } 1865 } 1866 } 1867 1868 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1869 /// and recursively continue updating all of its uses. 1870 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1871 SmallVector<AllocaInst*, 32> &NewElts) { 1872 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1873 if (BC->getOperand(0) != AI) 1874 return; 1875 1876 // The bitcast references the original alloca. Replace its uses with 1877 // references to the first new element alloca. 1878 Instruction *Val = NewElts[0]; 1879 if (Val->getType() != BC->getDestTy()) { 1880 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1881 Val->takeName(BC); 1882 } 1883 BC->replaceAllUsesWith(Val); 1884 DeadInsts.push_back(BC); 1885 } 1886 1887 /// FindElementAndOffset - Return the index of the element containing Offset 1888 /// within the specified type, which must be either a struct or an array. 1889 /// Sets T to the type of the element and Offset to the offset within that 1890 /// element. IdxTy is set to the type of the index result to be used in a 1891 /// GEP instruction. 1892 uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, 1893 Type *&IdxTy) { 1894 uint64_t Idx = 0; 1895 if (StructType *ST = dyn_cast<StructType>(T)) { 1896 const StructLayout *Layout = TD->getStructLayout(ST); 1897 Idx = Layout->getElementContainingOffset(Offset); 1898 T = ST->getContainedType(Idx); 1899 Offset -= Layout->getElementOffset(Idx); 1900 IdxTy = Type::getInt32Ty(T->getContext()); 1901 return Idx; 1902 } 1903 ArrayType *AT = cast<ArrayType>(T); 1904 T = AT->getElementType(); 1905 uint64_t EltSize = TD->getTypeAllocSize(T); 1906 Idx = Offset / EltSize; 1907 Offset -= Idx * EltSize; 1908 IdxTy = Type::getInt64Ty(T->getContext()); 1909 return Idx; 1910 } 1911 1912 /// RewriteGEP - Check if this GEP instruction moves the pointer across 1913 /// elements of the alloca that are being split apart, and if so, rewrite 1914 /// the GEP to be relative to the new element. 1915 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1916 SmallVector<AllocaInst*, 32> &NewElts) { 1917 uint64_t OldOffset = Offset; 1918 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1919 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1920 1921 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1922 1923 Type *T = AI->getAllocatedType(); 1924 Type *IdxTy; 1925 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1926 if (GEPI->getOperand(0) == AI) 1927 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1928 1929 T = AI->getAllocatedType(); 1930 uint64_t EltOffset = Offset; 1931 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1932 1933 // If this GEP does not move the pointer across elements of the alloca 1934 // being split, then it does not needs to be rewritten. 1935 if (Idx == OldIdx) 1936 return; 1937 1938 Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1939 SmallVector<Value*, 8> NewArgs; 1940 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1941 while (EltOffset != 0) { 1942 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1943 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1944 } 1945 Instruction *Val = NewElts[Idx]; 1946 if (NewArgs.size() > 1) { 1947 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI); 1948 Val->takeName(GEPI); 1949 } 1950 if (Val->getType() != GEPI->getType()) 1951 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1952 GEPI->replaceAllUsesWith(Val); 1953 DeadInsts.push_back(GEPI); 1954 } 1955 1956 /// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it 1957 /// to mark the lifetime of the scalarized memory. 1958 void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 1959 uint64_t Offset, 1960 SmallVector<AllocaInst*, 32> &NewElts) { 1961 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0)); 1962 // Put matching lifetime markers on everything from Offset up to 1963 // Offset+OldSize. 1964 Type *AIType = AI->getAllocatedType(); 1965 uint64_t NewOffset = Offset; 1966 Type *IdxTy; 1967 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy); 1968 1969 IRBuilder<> Builder(II); 1970 uint64_t Size = OldSize->getLimitedValue(); 1971 1972 if (NewOffset) { 1973 // Splice the first element and index 'NewOffset' bytes in. SROA will 1974 // split the alloca again later. 1975 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy()); 1976 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset)); 1977 1978 IdxTy = NewElts[Idx]->getAllocatedType(); 1979 uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset; 1980 if (EltSize > Size) { 1981 EltSize = Size; 1982 Size = 0; 1983 } else { 1984 Size -= EltSize; 1985 } 1986 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 1987 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize)); 1988 else 1989 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize)); 1990 ++Idx; 1991 } 1992 1993 for (; Idx != NewElts.size() && Size; ++Idx) { 1994 IdxTy = NewElts[Idx]->getAllocatedType(); 1995 uint64_t EltSize = TD->getTypeAllocSize(IdxTy); 1996 if (EltSize > Size) { 1997 EltSize = Size; 1998 Size = 0; 1999 } else { 2000 Size -= EltSize; 2001 } 2002 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2003 Builder.CreateLifetimeStart(NewElts[Idx], 2004 Builder.getInt64(EltSize)); 2005 else 2006 Builder.CreateLifetimeEnd(NewElts[Idx], 2007 Builder.getInt64(EltSize)); 2008 } 2009 DeadInsts.push_back(II); 2010 } 2011 2012 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2013 /// Rewrite it to copy or set the elements of the scalarized memory. 2014 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2015 AllocaInst *AI, 2016 SmallVector<AllocaInst*, 32> &NewElts) { 2017 // If this is a memcpy/memmove, construct the other pointer as the 2018 // appropriate type. The "Other" pointer is the pointer that goes to memory 2019 // that doesn't have anything to do with the alloca that we are promoting. For 2020 // memset, this Value* stays null. 2021 Value *OtherPtr = 0; 2022 unsigned MemAlignment = MI->getAlignment(); 2023 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2024 if (Inst == MTI->getRawDest()) 2025 OtherPtr = MTI->getRawSource(); 2026 else { 2027 assert(Inst == MTI->getRawSource()); 2028 OtherPtr = MTI->getRawDest(); 2029 } 2030 } 2031 2032 // If there is an other pointer, we want to convert it to the same pointer 2033 // type as AI has, so we can GEP through it safely. 2034 if (OtherPtr) { 2035 unsigned AddrSpace = 2036 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2037 2038 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2039 // optimization, but it's also required to detect the corner case where 2040 // both pointer operands are referencing the same memory, and where 2041 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2042 // function is only called for mem intrinsics that access the whole 2043 // aggregate, so non-zero GEPs are not an issue here.) 2044 OtherPtr = OtherPtr->stripPointerCasts(); 2045 2046 // Copying the alloca to itself is a no-op: just delete it. 2047 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2048 // This code will run twice for a no-op memcpy -- once for each operand. 2049 // Put only one reference to MI on the DeadInsts list. 2050 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2051 E = DeadInsts.end(); I != E; ++I) 2052 if (*I == MI) return; 2053 DeadInsts.push_back(MI); 2054 return; 2055 } 2056 2057 // If the pointer is not the right type, insert a bitcast to the right 2058 // type. 2059 Type *NewTy = 2060 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2061 2062 if (OtherPtr->getType() != NewTy) 2063 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2064 } 2065 2066 // Process each element of the aggregate. 2067 bool SROADest = MI->getRawDest() == Inst; 2068 2069 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2070 2071 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2072 // If this is a memcpy/memmove, emit a GEP of the other element address. 2073 Value *OtherElt = 0; 2074 unsigned OtherEltAlign = MemAlignment; 2075 2076 if (OtherPtr) { 2077 Value *Idx[2] = { Zero, 2078 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2079 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, 2080 OtherPtr->getName()+"."+Twine(i), 2081 MI); 2082 uint64_t EltOffset; 2083 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2084 Type *OtherTy = OtherPtrTy->getElementType(); 2085 if (StructType *ST = dyn_cast<StructType>(OtherTy)) { 2086 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2087 } else { 2088 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2089 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2090 } 2091 2092 // The alignment of the other pointer is the guaranteed alignment of the 2093 // element, which is affected by both the known alignment of the whole 2094 // mem intrinsic and the alignment of the element. If the alignment of 2095 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2096 // known alignment is just 4 bytes. 2097 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2098 } 2099 2100 Value *EltPtr = NewElts[i]; 2101 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2102 2103 // If we got down to a scalar, insert a load or store as appropriate. 2104 if (EltTy->isSingleValueType()) { 2105 if (isa<MemTransferInst>(MI)) { 2106 if (SROADest) { 2107 // From Other to Alloca. 2108 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2109 new StoreInst(Elt, EltPtr, MI); 2110 } else { 2111 // From Alloca to Other. 2112 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2113 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2114 } 2115 continue; 2116 } 2117 assert(isa<MemSetInst>(MI)); 2118 2119 // If the stored element is zero (common case), just store a null 2120 // constant. 2121 Constant *StoreVal; 2122 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2123 if (CI->isZero()) { 2124 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2125 } else { 2126 // If EltTy is a vector type, get the element type. 2127 Type *ValTy = EltTy->getScalarType(); 2128 2129 // Construct an integer with the right value. 2130 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2131 APInt OneVal(EltSize, CI->getZExtValue()); 2132 APInt TotalVal(OneVal); 2133 // Set each byte. 2134 for (unsigned i = 0; 8*i < EltSize; ++i) { 2135 TotalVal = TotalVal.shl(8); 2136 TotalVal |= OneVal; 2137 } 2138 2139 // Convert the integer value to the appropriate type. 2140 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2141 if (ValTy->isPointerTy()) 2142 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2143 else if (ValTy->isFloatingPointTy()) 2144 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2145 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2146 2147 // If the requested value was a vector constant, create it. 2148 if (EltTy->isVectorTy()) { 2149 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements(); 2150 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2151 StoreVal = ConstantVector::get(Elts); 2152 } 2153 } 2154 new StoreInst(StoreVal, EltPtr, MI); 2155 continue; 2156 } 2157 // Otherwise, if we're storing a byte variable, use a memset call for 2158 // this element. 2159 } 2160 2161 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2162 2163 IRBuilder<> Builder(MI); 2164 2165 // Finally, insert the meminst for this element. 2166 if (isa<MemSetInst>(MI)) { 2167 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2168 MI->isVolatile()); 2169 } else { 2170 assert(isa<MemTransferInst>(MI)); 2171 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2172 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2173 2174 if (isa<MemCpyInst>(MI)) 2175 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2176 else 2177 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2178 } 2179 } 2180 DeadInsts.push_back(MI); 2181 } 2182 2183 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2184 /// overwrites the entire allocation. Extract out the pieces of the stored 2185 /// integer and store them individually. 2186 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2187 SmallVector<AllocaInst*, 32> &NewElts){ 2188 // Extract each element out of the integer according to its structure offset 2189 // and store the element value to the individual alloca. 2190 Value *SrcVal = SI->getOperand(0); 2191 Type *AllocaEltTy = AI->getAllocatedType(); 2192 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2193 2194 IRBuilder<> Builder(SI); 2195 2196 // Handle tail padding by extending the operand 2197 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2198 SrcVal = Builder.CreateZExt(SrcVal, 2199 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2200 2201 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2202 << '\n'); 2203 2204 // There are two forms here: AI could be an array or struct. Both cases 2205 // have different ways to compute the element offset. 2206 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2207 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2208 2209 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2210 // Get the number of bits to shift SrcVal to get the value. 2211 Type *FieldTy = EltSTy->getElementType(i); 2212 uint64_t Shift = Layout->getElementOffsetInBits(i); 2213 2214 if (TD->isBigEndian()) 2215 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2216 2217 Value *EltVal = SrcVal; 2218 if (Shift) { 2219 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2220 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2221 } 2222 2223 // Truncate down to an integer of the right size. 2224 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2225 2226 // Ignore zero sized fields like {}, they obviously contain no data. 2227 if (FieldSizeBits == 0) continue; 2228 2229 if (FieldSizeBits != AllocaSizeBits) 2230 EltVal = Builder.CreateTrunc(EltVal, 2231 IntegerType::get(SI->getContext(), FieldSizeBits)); 2232 Value *DestField = NewElts[i]; 2233 if (EltVal->getType() == FieldTy) { 2234 // Storing to an integer field of this size, just do it. 2235 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2236 // Bitcast to the right element type (for fp/vector values). 2237 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2238 } else { 2239 // Otherwise, bitcast the dest pointer (for aggregates). 2240 DestField = Builder.CreateBitCast(DestField, 2241 PointerType::getUnqual(EltVal->getType())); 2242 } 2243 new StoreInst(EltVal, DestField, SI); 2244 } 2245 2246 } else { 2247 ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2248 Type *ArrayEltTy = ATy->getElementType(); 2249 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2250 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2251 2252 uint64_t Shift; 2253 2254 if (TD->isBigEndian()) 2255 Shift = AllocaSizeBits-ElementOffset; 2256 else 2257 Shift = 0; 2258 2259 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2260 // Ignore zero sized fields like {}, they obviously contain no data. 2261 if (ElementSizeBits == 0) continue; 2262 2263 Value *EltVal = SrcVal; 2264 if (Shift) { 2265 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2266 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2267 } 2268 2269 // Truncate down to an integer of the right size. 2270 if (ElementSizeBits != AllocaSizeBits) 2271 EltVal = Builder.CreateTrunc(EltVal, 2272 IntegerType::get(SI->getContext(), 2273 ElementSizeBits)); 2274 Value *DestField = NewElts[i]; 2275 if (EltVal->getType() == ArrayEltTy) { 2276 // Storing to an integer field of this size, just do it. 2277 } else if (ArrayEltTy->isFloatingPointTy() || 2278 ArrayEltTy->isVectorTy()) { 2279 // Bitcast to the right element type (for fp/vector values). 2280 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2281 } else { 2282 // Otherwise, bitcast the dest pointer (for aggregates). 2283 DestField = Builder.CreateBitCast(DestField, 2284 PointerType::getUnqual(EltVal->getType())); 2285 } 2286 new StoreInst(EltVal, DestField, SI); 2287 2288 if (TD->isBigEndian()) 2289 Shift -= ElementOffset; 2290 else 2291 Shift += ElementOffset; 2292 } 2293 } 2294 2295 DeadInsts.push_back(SI); 2296 } 2297 2298 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2299 /// an integer. Load the individual pieces to form the aggregate value. 2300 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2301 SmallVector<AllocaInst*, 32> &NewElts) { 2302 // Extract each element out of the NewElts according to its structure offset 2303 // and form the result value. 2304 Type *AllocaEltTy = AI->getAllocatedType(); 2305 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2306 2307 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2308 << '\n'); 2309 2310 // There are two forms here: AI could be an array or struct. Both cases 2311 // have different ways to compute the element offset. 2312 const StructLayout *Layout = 0; 2313 uint64_t ArrayEltBitOffset = 0; 2314 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2315 Layout = TD->getStructLayout(EltSTy); 2316 } else { 2317 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2318 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2319 } 2320 2321 Value *ResultVal = 2322 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2323 2324 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2325 // Load the value from the alloca. If the NewElt is an aggregate, cast 2326 // the pointer to an integer of the same size before doing the load. 2327 Value *SrcField = NewElts[i]; 2328 Type *FieldTy = 2329 cast<PointerType>(SrcField->getType())->getElementType(); 2330 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2331 2332 // Ignore zero sized fields like {}, they obviously contain no data. 2333 if (FieldSizeBits == 0) continue; 2334 2335 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2336 FieldSizeBits); 2337 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2338 !FieldTy->isVectorTy()) 2339 SrcField = new BitCastInst(SrcField, 2340 PointerType::getUnqual(FieldIntTy), 2341 "", LI); 2342 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2343 2344 // If SrcField is a fp or vector of the right size but that isn't an 2345 // integer type, bitcast to an integer so we can shift it. 2346 if (SrcField->getType() != FieldIntTy) 2347 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2348 2349 // Zero extend the field to be the same size as the final alloca so that 2350 // we can shift and insert it. 2351 if (SrcField->getType() != ResultVal->getType()) 2352 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2353 2354 // Determine the number of bits to shift SrcField. 2355 uint64_t Shift; 2356 if (Layout) // Struct case. 2357 Shift = Layout->getElementOffsetInBits(i); 2358 else // Array case. 2359 Shift = i*ArrayEltBitOffset; 2360 2361 if (TD->isBigEndian()) 2362 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2363 2364 if (Shift) { 2365 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2366 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2367 } 2368 2369 // Don't create an 'or x, 0' on the first iteration. 2370 if (!isa<Constant>(ResultVal) || 2371 !cast<Constant>(ResultVal)->isNullValue()) 2372 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2373 else 2374 ResultVal = SrcField; 2375 } 2376 2377 // Handle tail padding by truncating the result 2378 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2379 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2380 2381 LI->replaceAllUsesWith(ResultVal); 2382 DeadInsts.push_back(LI); 2383 } 2384 2385 /// HasPadding - Return true if the specified type has any structure or 2386 /// alignment padding in between the elements that would be split apart 2387 /// by SROA; return false otherwise. 2388 static bool HasPadding(Type *Ty, const TargetData &TD) { 2389 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2390 Ty = ATy->getElementType(); 2391 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2392 } 2393 2394 // SROA currently handles only Arrays and Structs. 2395 StructType *STy = cast<StructType>(Ty); 2396 const StructLayout *SL = TD.getStructLayout(STy); 2397 unsigned PrevFieldBitOffset = 0; 2398 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2399 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2400 2401 // Check to see if there is any padding between this element and the 2402 // previous one. 2403 if (i) { 2404 unsigned PrevFieldEnd = 2405 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2406 if (PrevFieldEnd < FieldBitOffset) 2407 return true; 2408 } 2409 PrevFieldBitOffset = FieldBitOffset; 2410 } 2411 // Check for tail padding. 2412 if (unsigned EltCount = STy->getNumElements()) { 2413 unsigned PrevFieldEnd = PrevFieldBitOffset + 2414 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2415 if (PrevFieldEnd < SL->getSizeInBits()) 2416 return true; 2417 } 2418 return false; 2419 } 2420 2421 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2422 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2423 /// or 1 if safe after canonicalization has been performed. 2424 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2425 // Loop over the use list of the alloca. We can only transform it if all of 2426 // the users are safe to transform. 2427 AllocaInfo Info(AI); 2428 2429 isSafeForScalarRepl(AI, 0, Info); 2430 if (Info.isUnsafe) { 2431 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2432 return false; 2433 } 2434 2435 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2436 // source and destination, we have to be careful. In particular, the memcpy 2437 // could be moving around elements that live in structure padding of the LLVM 2438 // types, but may actually be used. In these cases, we refuse to promote the 2439 // struct. 2440 if (Info.isMemCpySrc && Info.isMemCpyDst && 2441 HasPadding(AI->getAllocatedType(), *TD)) 2442 return false; 2443 2444 // If the alloca never has an access to just *part* of it, but is accessed 2445 // via loads and stores, then we should use ConvertToScalarInfo to promote 2446 // the alloca instead of promoting each piece at a time and inserting fission 2447 // and fusion code. 2448 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2449 // If the struct/array just has one element, use basic SRoA. 2450 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2451 if (ST->getNumElements() > 1) return false; 2452 } else { 2453 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2454 return false; 2455 } 2456 } 2457 2458 return true; 2459 } 2460 2461 2462 2463 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2464 /// some part of a constant global variable. This intentionally only accepts 2465 /// constant expressions because we don't can't rewrite arbitrary instructions. 2466 static bool PointsToConstantGlobal(Value *V) { 2467 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2468 return GV->isConstant(); 2469 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2470 if (CE->getOpcode() == Instruction::BitCast || 2471 CE->getOpcode() == Instruction::GetElementPtr) 2472 return PointsToConstantGlobal(CE->getOperand(0)); 2473 return false; 2474 } 2475 2476 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2477 /// pointer to an alloca. Ignore any reads of the pointer, return false if we 2478 /// see any stores or other unknown uses. If we see pointer arithmetic, keep 2479 /// track of whether it moves the pointer (with isOffset) but otherwise traverse 2480 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2481 /// the alloca, and if the source pointer is a pointer to a constant global, we 2482 /// can optimize this. 2483 static bool 2484 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2485 bool isOffset, 2486 SmallVector<Instruction *, 4> &LifetimeMarkers) { 2487 // We track lifetime intrinsics as we encounter them. If we decide to go 2488 // ahead and replace the value with the global, this lets the caller quickly 2489 // eliminate the markers. 2490 2491 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2492 User *U = cast<Instruction>(*UI); 2493 2494 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2495 // Ignore non-volatile loads, they are always ok. 2496 if (!LI->isSimple()) return false; 2497 continue; 2498 } 2499 2500 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2501 // If uses of the bitcast are ok, we are ok. 2502 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset, 2503 LifetimeMarkers)) 2504 return false; 2505 continue; 2506 } 2507 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2508 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2509 // doesn't, it does. 2510 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2511 isOffset || !GEP->hasAllZeroIndices(), 2512 LifetimeMarkers)) 2513 return false; 2514 continue; 2515 } 2516 2517 if (CallSite CS = U) { 2518 // If this is the function being called then we treat it like a load and 2519 // ignore it. 2520 if (CS.isCallee(UI)) 2521 continue; 2522 2523 // If this is a readonly/readnone call site, then we know it is just a 2524 // load (but one that potentially returns the value itself), so we can 2525 // ignore it if we know that the value isn't captured. 2526 unsigned ArgNo = CS.getArgumentNo(UI); 2527 if (CS.onlyReadsMemory() && 2528 (CS.getInstruction()->use_empty() || 2529 CS.paramHasAttr(ArgNo+1, Attribute::NoCapture))) 2530 continue; 2531 2532 // If this is being passed as a byval argument, the caller is making a 2533 // copy, so it is only a read of the alloca. 2534 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2535 continue; 2536 } 2537 2538 // Lifetime intrinsics can be handled by the caller. 2539 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 2540 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 2541 II->getIntrinsicID() == Intrinsic::lifetime_end) { 2542 assert(II->use_empty() && "Lifetime markers have no result to use!"); 2543 LifetimeMarkers.push_back(II); 2544 continue; 2545 } 2546 } 2547 2548 // If this is isn't our memcpy/memmove, reject it as something we can't 2549 // handle. 2550 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2551 if (MI == 0) 2552 return false; 2553 2554 // If the transfer is using the alloca as a source of the transfer, then 2555 // ignore it since it is a load (unless the transfer is volatile). 2556 if (UI.getOperandNo() == 1) { 2557 if (MI->isVolatile()) return false; 2558 continue; 2559 } 2560 2561 // If we already have seen a copy, reject the second one. 2562 if (TheCopy) return false; 2563 2564 // If the pointer has been offset from the start of the alloca, we can't 2565 // safely handle this. 2566 if (isOffset) return false; 2567 2568 // If the memintrinsic isn't using the alloca as the dest, reject it. 2569 if (UI.getOperandNo() != 0) return false; 2570 2571 // If the source of the memcpy/move is not a constant global, reject it. 2572 if (!PointsToConstantGlobal(MI->getSource())) 2573 return false; 2574 2575 // Otherwise, the transform is safe. Remember the copy instruction. 2576 TheCopy = MI; 2577 } 2578 return true; 2579 } 2580 2581 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2582 /// modified by a copy from a constant global. If we can prove this, we can 2583 /// replace any uses of the alloca with uses of the global directly. 2584 MemTransferInst * 2585 SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI, 2586 SmallVector<Instruction*, 4> &ToDelete) { 2587 MemTransferInst *TheCopy = 0; 2588 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete)) 2589 return TheCopy; 2590 return 0; 2591 } 2592